I cringe at the camera. I hate the sound of my own voice. I’m too lazy to set up lighting, do my hair, take it again because I stumbled on a word.

So this week I tried to skip it entirely.

I gave HeyGen 15 seconds of grainy café footage and one headshot. 5 minutes later I had an AI version of me that could record videos while I did literally anything else.

It worked better than I wanted it to. With one big caveat I’ll get to.

What I was actually testing

Before I go further: I wasn’t testing the voice this week.

This week was about one question: is the video good enough?

Because that’s the harder problem. Voice I can solve. I can record audio without a camera, without lighting, without my hair done. I can do that in two minutes. The video is the part I dread.

So I went in looking at the footage. Not the audio.

What HeyGen actually is

HeyGen turns a script into a video of an AI avatar speaking it. You can pick from their library of pre-made avatars (300+, every demographic) or train one on yourself.

I chose Avatar V, their newest model. The previous version, Avatar IV, has a reputation for not quite getting there. Stiff movement, slightly off lip sync, the uncanny valley problem. Avatar V is supposed to fix that.

The free plan gives you one Avatar V test video. That’s it. One shot. After that you’re paying. Plans start at $24/month annually, $29 monthly.

The 5-minute experiment

I sat down in a café. Sweaty. No makeup. Bad lighting. Phone camera. Recorded 15 seconds of myself reading the HeyGen script to the lens.

The footage looked like shit. That was the point. Most of you don’t have a studio either.

Here’s what happened next.

Step 1: Upload the clip.

HeyGen analyzed it and generated 3 voice variations and a handful of avatar options. Office setting. Casual. Podcast. Different outfits, different backgrounds.

Step 2: Listen to the voices.

None of them sounded like me. They sounded AI-generated. One was less bad than the others. I picked it and moved on, because today the voice wasn’t the point.

Step 3: Look at the avatars.

Some came close to my face. None nailed it. The hair was right, the general shape was right, but the eyes were off. Slightly dead, slightly someone-else’s.

I almost stopped here.

Step 4: Upload one headshot.

This was the unlock.

HeyGen used the headshot for the face and pulled the motion from the café selfie video. See the yellow marked headshot and video above.

Suddenly the avatar moved like me. Sat like me. Tilted her head like me.

The official result

Step 5: Notice what’s wrong.

Two things. The teeth were slightly off (fixable, I wasn’t smiling in my 15 sec selfie video). And the avatar was wearing a wedding ring.

I don’t wear rings. I’m not married. My hands weren’t even in the footage I uploaded.

The AI just... decided. Maybe it was trying to tell me something. Maybe HeyGen thinks being in your thirties should come with a husband by now.

I’ll get back to the AI on that one.

Total time: 5 minutes. Total cost: zero.

The brand kit nobody talks about

This is the part most reviews skip.

I pasted my website URL (brisk.vision) into HeyGen. It went and read the site. Pulled my colors, my fonts, my visual style. Built me a brand template I didn’t ask for.

Then it offered me caption styles, intro animations, outros, transitions. All matched to the brand. All editable.

The captions were the bit that surprised me. HeyGen automatically generated them in my brand colors. Highlighted keywords in my accent shade. Used my font. The kind of polish that takes ages to do manually in CapCut or Descript.

Note: In the draft preview, you can only see the captions. You get the Avatar talking only via Video generation. I have used the free one I had. Hence I am showing below recording only for the captions, but imagine the Avatar image actually talking.

Fun fact: I’m Croatian.

So the next thing I tested was the Croatian translation.

Croatian is my benchmark for any AI tool. The grammar is brutal. Cases, genders, verb forms that change based on who’s doing the action. Most AI models butcher it.

HeyGen did okay-ish. Two issues: the translation pretended I was male (had to manually fix every verb form), and it was too literal. Not how I’d actually talk. But functional. A starting point.

If your audience speaks anything other than English, this matters more than the avatar quality.

What I couldn’t test (and won’t, until I pay)

The free plan gives you one Avatar V export. Including a watermark. I used it.

So I couldn’t test the office setting. The walking avatar. Different transitions. Different intro animations. I saw the previews. I didn’t see the finished output.

This is HeyGen’s freemium trap. You won’t know if you’ll actually use the product until you commit a month at $24-29.

I did however manage to get another free low quality video that actually looked a lot like a tired version of me (if we ignore the lip sync & voice):

My rule: never pay for SaaS until I know I’ll use it weekly.

So I’m not subscribing. Yet.

If you’re considering it, give yourself 30 days on the free plan first. Make a list of what you’d actually publish with it. If the list has fewer than 4 things, save your money.

Who should pay for HeyGen

Pay for it if

→ You publish video content weekly and hate being on camera

→ You sell a course or training and need 10+ explainer videos

→ You do client outreach and want personalized video at scale

→ You translate content into 3+ languages regularly

→ You do mind having the HeyGen watermark in your video

Skip it if

→ You post videos occasionally and don’t mind your face

→ You only need one avatar for one project

→ You’re testing whether you’ll commit to video at all (free plan first)

One pricing thing nobody warns you about: Avatar V (the good one) burns through “premium credits” fast. The Creator plan gives you 200 credits a month, which is roughly 10 minutes of premium video.

If you’re publishing 3-4 videos a week, you’ll hit the ceiling in week one. The Pro plan ($79/month annually) gets you 2,000 credits, which is the realistic tier for anyone serious about this.

The part I want my mum to read

Here’s what made me stop and think after 5 minutes of fucking around with this.

If I can build a believable avatar of myself in 5 minutes with one bad selfie, so can anyone else. They don’t even need 5 minutes. Voice cloning tools work with 3 seconds of audio. Three seconds.

I have a podcast clip somewhere. Voice notes I’ve sent. Substack videos. Instagram stories from years ago. My voice is everywhere.

So is yours.

I thought about my parents. They’re old. They trust voices. If someone called my mum tomorrow crying in my voice, asking for money because I was in trouble, I don’t know if she’d ask the right questions.

The fix is one thing: a family code word. Pick a random word. Tell your parents, your kids, your siblings, your partner. If anyone calls in a panic asking for money, ask for the word. No word, hang up and call them back on a number you already have.

Voice cloning attacks jumped over 400% in 2025. This isn’t theoretical. It’s happening to families right now. Florida. Arizona. Hong Kong. Mostly to people whose kids never thought to have this conversation.

Have the conversation. Tonight. It takes 30 seconds.

Where this gets really interesting (next week)

Last week I showed how Claude generates Instagram carousels for you. That one brought a wave of new subscribers. Thank you for being here.

This week was the manual HeyGen experiment. Slow. One-off. Click by click. Good enough for someone testing the waters.

Next week is the part I’m actually excited about.

Claude Code wired to HeyGen.

One prompt → research → script → finished avatar video.

Most popular use case: automated Reels.

No clicking. Plus the voice fix: I’ll show you how to plug a proper voice cloning tool into HeyGen so the avatar finally sounds like a human, not a corporate AI assistant.

It’s more advanced than this week. If you’ve never opened Claude Code, I’ll link back to my older walkthroughs so you can catch up. I’ll go step by step. And I’ll tell you honestly whether the setup time is worth it.

For most of you, this week’s manual workflow is enough. For the rest, next Thursday is going to be fun.

I started this experiment because I hate recording videos. I’m finishing it knowing the video technology is good enough to replace me. The voice isn’t there yet. But that’s a fixable problem, and next week I’ll fix it.

Tonight, do two things. Pick a family code word. Then try HeyGen’s free plan with one 15-second video of yourself. See what your AI twin looks like.

Tell me what type of jewelry it gives you.

PS - Let me know how it goes. I encourage you to share your good & bad AI Avatar version.

Leave a comment

Subscribers get the full HeyGen + Claude Code setup guide incl. terminal commands directly into their Inbox next week.

You are not a subscriber? Change it now!

Subscribe now

This article is part of my ongoing series on making AI simple and useful for non-tech people. Subscribe to get future articles on AI tools that actually matter for everyday life. And if you found something that helped you - sharing is caring 😉

Share

I had three chefs in my kitchen. A research agent scanning for content ideas. A writer agent drafting LinkedIn posts. A carousel agent turning posts into slides.

All built in Claude Code. All cooking unsupervised. All doing their jobs.

On Tuesday morning I checked my drafts folder. 126 files.

The research agent had been running in a loop all night, generating “content angles” nobody ordered. No budget cap. No service hours. No one watching the pass.

That’s not an agent problem. That’s a kitchen management problem (metaphorically speaking).

And that’s exactly what Paperclip was built to solve.

When one chef isn’t enough

In last week’s article, I walked you through the 5 levels of the AI Kitchen. Levels 1 through 4 are all about one chef: one agent, one task, one kitchen.

Read here: https://lucijatomsic.substack.com/p/agents

Level 5 is different. One chef became a team.

Picture a restaurant. You have a head chef, line cooks at each station, a pastry chef, a prep crew. Tickets come in, tickets go out. There’s a pass, a schedule, a budget per station, and an audit log for every plate.

Now picture the same restaurant without a manager. Without a ticket system. Without a budget. Every cook just... does whatever they want.

That’s what running multiple AI agents without Paperclip looks like.

I lived it. Ten terminal tabs. No way to track which agent was doing what. No cost controls. Total context loss every time my machine rebooted.

Here’s what nobody tells you about multi-agent setups:

→ No budget tracking. One agent burned through 20€ in tokens while I slept.

→ No audit trail. What did the agent cook at 3 AM? Nobody knows.

→ Session loss on reboot. Your entire brigade wakes up with amnesia every morning.

→ No org chart. Who reports to whom? Who has decision rights over the final plate?

I didn’t have an AI problem. I had a management problem.

And management is something I actually know how to do.

Share

What Paperclip actually is

Paperclip is NOT a chef. It doesn’t cook. It doesn’t write. It doesn’t code. It doesn’t research.

Paperclip is the restaurant. The front-of-house manager, the shift schedule, the POS system, the budget ledger, and the ticket rail.

The best one-liner I’ve seen about it: “If OpenClaw is an employee, Paperclip is the company.”

In kitchen terms: if Claude Code is the chef, Paperclip is the restaurant that chef works in.

Here’s what it gives you:

→ Org chart. Head chef agent, sous-chef agents, line cook agents. Real hierarchy. Real reporting lines. Not just a pile of terminals.

→ Ticket board. Every task is a ticket with a clear owner, status, and thread. Like Asana, but for AI agents.

→ Budget controls. Set a monthly token limit per agent. Agent hits budget? It pauses. Not warns. Pauses. This alone would have prevented my 126-file disaster.

→ Heartbeats. Agents don’t run 24/7 burning tokens. They wake up on a schedule (every 4, 8, or 12 hours), check their tickets, do their work, go back to sleep. Service hours, not an all-night bender.

→ Audit log. Every decision, every tool call, every dollar spent. Append-only. Nothing can be deleted or edited. When a plate goes wrong (and it will), you have the full ticket history.

→ You’re the restaurateur. Agents propose plans. You approve. They want to “hire” a new line cook? You say yes or no. They want to change the menu? They ask you first.

Subscribe now

How the restaurant runs

This is where the metaphor clicks. Because Paperclip is built on the one mental model every consultant and manager already has: an organization.

Restaurateur (You)
└── Head Chef Agent (CEO)
    ├── Marketing Agent
    │   └── Copywriter Agent
    ├── Tech Lead Agent
    │   ├── Engineer Agent
    │   └── QA Agent
    └── Research Agent

You give the head chef a mission: “Build a landing page for our new AI consulting service.”

The head chef breaks it into tickets. It proposes to “hire” a tech lead and a marketing agent. You approve from the dashboard.

The tech lead hires an engineer. Marketing hires a copywriter. Each agent gets a job description, a budget, and a heartbeat schedule.

You watch from the pass. Like running any team, except this one actually does what it said it would do.

How heartbeats work (service hours for your brigade)

This is the part that changed how I think about agents.

You set a schedule. Every 8 hours.

At 8:00 AM, the head chef agent wakes up. It reads its identity file (who am I?), checks the plan (what are we building?), looks at the ticket board (what’s assigned to me?), breaks work into steps, executes, logs what it learned, and goes back to sleep.

At 8:05, the engineer wakes. Reads its assignment from the head chef. Writes code. Saves progress. Goes back to sleep.

At 4:00 PM, they all wake again. Head chef reviews the engineer’s output. Copywriter submits a draft. QA checks everything.

You get a summary in your inbox.

They don’t run 24/7 burning your API budget. They work in shifts. Service hours, not an open kitchen.

Every agent gets 4 files that define how it behaves:

AGENTS.md tells the agent who it is and what its role is.

HEARTBEAT.md is its opening routine. The checklist it runs every time it wakes up.

SOUL.md is its personality. The head chef’s soul file says things like “You own the P&L. You default to action.”

TOOLS.md defines what the agent can and can’t touch.

You can edit all of these. That’s how you shape behavior without writing code.

Paperclip vs. everything else

Here’s how all the tools from last week’s article fit together, mapped to the AI Kitchen levels.

Level 1. Claude Chat. You’re at the counter. You order, Claude cooks, you eat. No kitchen of your own.

Level 2. Claude Cowork. You’re in your kitchen. Claude is your sous-chef. You’re still holding the knife.

Level 3. Claude Code / OpenClaw. You wrote the recipe and left. The chef cooks alone in your kitchen. Your gas bill. Your fire alarm.

Level 4. Claude Managed Agents. Same chef, same recipe. But now the cooking happens in Anthropic’s kitchen. You don’t own the stove. Launched April 8, 2026.

Level 5. Paperclip. One chef became a brigade. You’re the restaurateur. Org chart, budgets, ticket system, heartbeats, audit log. You run the house.

They’re not competitors. They’re levels.

Paperclip uses Claude Code, OpenClaw, Codex, Cursor, or any other agent as its chefs. It doesn’t replace them. It gives them a restaurant to work in.

How to set up Paperclip (the full walkthrough)

Every step. Real commands. If you followed my Claude Code setup guide, you can do this.

A. What you need before you start

→ A computer (Mac, Windows, or Linux).

→ Node.js version 20 or higher. Node.js is a program that lets you run certain tools on your computer. Free. Download from nodejs.org. Check with node --version in your terminal.

→ pnpm version 9.15 or higher. A package manager. Install it by typing npm install -g pnpm in your terminal.

→ An API key from Anthropic. Your access pass to use Claude as the brain behind your agents. Get one at console.anthropic.com. It costs money per use.

If “terminal” still feels foreign: I wrote a full Claude Code setup guide that walks you through VS Code so you never touch a raw terminal. Start there.

B. Install Paperclip (one command)

Open your terminal. Type this. Press Enter.

npx paperclipai onboard --yes

Paperclip installs itself, creates its own database, and walks you through setup. One command. Done.

If you are still scared of the terminal, I put together an installation guide for dummies. Subscribers can access it in my Resource section.

Subscribe now

C. Open the dashboard

Open your browser. Go to:

http://localhost:3100

This is your restaurant headquarters. Everything happens here.

On first load, an onboarding wizard appears and walks you through 4 steps. Follow it.

D. Create your restaurant (company)

The wizard asks for two things:

1. Company name. Call it whatever you want. “Content Team” or “brisk. Agents” (in my case hehe) or your project name. Each company in Paperclip is fully isolated. Its own agents, budgets, and audit trail. You can run multiple companies from one install.

2. Company mission. This matters. Every ticket your agents work on traces back to this mission. Be specific.

Good: “Produce one LinkedIn post and one newsletter article per week about AI tools for non-technical professionals.”

Bad: “Do content stuff.”

E. Hire your head chef (first agent)

Click “Create Agent.” Fill in:

→ Name: CEO (or Head Chef, whatever you prefer)

→ Model: Claude Sonnet 4.6 (cheaper than Opus, smart enough for coordination)

→ Adapter: Claude

Click “Test now” to verify the connection. You’ll see “hello probe succeeded.”

Paperclip creates the 4 identity files automatically: AGENTS.md, HEARTBEAT.md, SOUL.md, TOOLS.md. Edit them later to customize.

F. Give it a ticket

Type your first task:

“Research 5 trending AI tools this week and write a one-paragraph summary of each. Focus on tools relevant to non-technical professionals.”

Click publish. The head chef wakes up, reads the ticket, starts working.

Check the status of each ticket in the issues board - it is a kanban board, similar like in Jira or other project management tools.

G. Set the budget and service hours

→ Monthly token budget: Start with 20€. You can always increase.

→ Heartbeat schedule: Every 8 hours.

→ At the limit: The agent pauses. It doesn’t warn you and keep spending. It stops.

H. Watch the pass

Check your inbox in the dashboard. The head chef will:

1. Read the ticket.

2. Propose a plan.

3. Ask to “hire” additional agents if the task needs it.

4. Wait for your approval before any big decision.

When the head chef proposes hiring an engineer, you’ll see: “I recommend hiring a Copywriter agent to handle the content drafts. Approve?”

Click approve. The new agent gets created. The head chef delegates. The org chart updates automatically.

You’re the restaurateur. You run the pass.

Share

What people are actually building

Not theoretical. Real projects running right now.

Content operation

Head chef coordinates. Researcher finds trending topics. Writer drafts articles and posts. Editor reviews for quality. Social media agent formats for different platforms.

Each agent wakes every 8 hours. Produces content autonomously. Human approves before anything gets published. Monthly budget: 100-200€ across all agents.

Lead generation

Head chef plans outreach strategy. Research agent finds B2B prospects. Copywriter crafts personalized messages. QA reviews everything before it goes out.

Monthly budget: ~265€ total. Head chef on Opus 4.7 (50€/month cap). Specialists on Sonnet 4.6 (40-100€/month each).

SaaS development team

Head chef defines the product vision. Tech lead breaks it into tasks. Engineer writes code. QA tests it. Designer creates the spec.

One team built an MVP in 2 weeks that would have taken 2 months with a traditional approach. Agents collaborate through the ticket system. Human reviews every major decision.

Pre-built restaurant templates

Paperclip is building “Clipmart.” A marketplace where you download a ready-made company template with one click. A content agency. A lead gen operation. A dev shop.

Full org structures, agent configs, and skills pre-configured. Still in development, but you can already import community templates:

npx paperclipai company import --from ./trail-of-bits-security

One command. Entire brigade. Ready to cook.

Should you actually do this?

Yes, if:

→ You’re comfortable with Claude and want to see what’s next.

→ You manage projects or teams already. The mental model will click.

→ You’re willing to type a few commands in a terminal.

→ You want to understand where business is heading in the next 2-3 years.

Not yet, if:

→ You’re still learning how to use Claude for basic tasks. Start at Level 1.

→ You don’t want to pay for API tokens. Agents cost real money.

→ “Self-hosting” sounds like a foreign language.

→ You need production-ready today. Paperclip launched 7 weeks ago. Powerful, but early. One honest reviewer called it “a proof of concept wearing a product’s clothing.”

Your kitchen, your rules

A year ago, “using AI” meant writing a better prompt. You were at the counter, ordering one dish at a time.

Six months ago, it meant working alongside a sous-chef. You were in your kitchen, cooking together.

Today, it means running a restaurant. Setting the menu. Hiring the brigade. Watching the pass. Approving the big decisions. Checking the books.

The skills that make a good consultant (clear goals, structured delegation, knowing when to intervene and when to step back) are exactly the skills that make someone good at running a brigade of AI agents.

That’s not a coincidence. That’s the whole point of this metaphor.

You don’t need to open a restaurant tonight. But the managers and consultants who understand this shift will have a 2-year head start on everyone still ordering at the counter.

Open a terminal. Type the command. Meet your brigade.

PS - Let me know how it goes. I encourage you to share the companies you built in the comments. We can help each other out by exporting our success stories!

Leave a comment

Subscribers get the full Paperclip Installation Guide incl. Terminal commands including my explanations for dummies in the Resource section.

You are not a subscriber? Change it now!

Subscribe now

This article is part of my ongoing series on making AI simple and useful for non-tech people. Subscribe to get future articles on AI tools that actually matter for everyday life. And if you found something that helped you - sharing is caring 😉

Share

Neither did I, three months ago. I nodded along on LinkedIn, liked a few posts about “agentic workflows,” and quietly googled it when nobody was looking. The answers were worse than the confusion.

Here’s what I found.

The simplest explanation that actually works

A chatbot is a short-order cook at a counter. You order. It cooks the plate in front of you. It hands it over. Ask for the next dish, it cooks that one too. Walk away, the kitchen goes quiet.

An agent is a chef with a kitchen. You say “make me dinner with what’s in the fridge.” The chef opens the fridge, checks what’s there, picks a dish, cooks it, tastes it, adjusts the salt, plates it - while you’re in another room.

That’s the difference. Counter cooks take orders. Chefs run kitchens.

Three things turn a counter cook into a chef:

→ It plans. It figures out the steps on its own.

→ It acts. It uses real tools: your browser, your files, your inbox.

→ It loops. It tastes its own work, catches mistakes, tries again.

Miss one of those three and it’s not an agent. It’s a chatbot in a chef’s hat.

And inside every real chef-agent, there are 6 things working together in the kitchen. I’ll show you all 6 in a minute. First, the levels.

Where you are right now (the 5 levels)

For now: skip ahead only when Levels 1 to 3 feel boring.

Share

The 5 tools everyone keeps talking about

A. Claude Chat (claude.ai). The counter cook. You walk up, you order, you get a dish. Free or €17/month for Pro. Start here. Period.

B. Claude Cowork. Your sous-chef in your own kitchen. Cowork went from research preview to general availability on April 9, 2026, on Pro, Team, and Enterprise. Mac and Windows.

The difference between Claude Chat and Cowork is the difference between ordering at the counter and bringing Claude into your kitchen. At the counter, Claude cooks one dish at a time with nothing but what you describe. In your kitchen, Claude opens the fridge, reads your files, and cooks alongside you.

Real example: You point Cowork at a folder with 6 client briefs and say “consolidate these into one summary deck.” It reads all 6, finds the patterns, builds the deck. You watch. You course-correct.

Cowork works WITH you. An agent works FOR you. Cowork is the bridge between the two.

C. OpenClaw. Free, open source, runs on your computer. Powerful. Also a security nightmare. Cisco flagged it. Bitdefender found nearly 900 malicious packages in its plugin registry. One user’s agent built itself a dating profile. Skip it unless you’re a developer who understands sandboxing.

D. Claude Code. Anthropic’s command-line tool for building with Claude. Not an agent itself — it’s where you hire your chef and write the recipes. You point it at a folder on your computer, describe what you want, and it reads your files, writes code, runs tests, fixes errors, and loops until the job is done.

Built with developers in mind, but I use it weekly to build my own agents and I don’t code. The terminal is the only barrier, and it’s smaller than it looks. More on how I actually use it in Section 4.

E. Claude Managed Agents. Brand new. Public beta launched April 8, 2026. Anthropic now hosts your agents for you. You define the job. They own the kitchen. Your chef still cooks in it. Notion, Asana, and Sentry are already using it.

Yes, all 5 are Anthropic-adjacent. That’s because they’re what I actually use. Build agents on any model you want. The concepts don’t change.

What’s actually inside an agent (the 6 components)

This is the part nobody on social media is writing about. And it’s the part that decides whether your agent works or produces garbage.

A great framework on a basic model beats a great model with no framework. Every time.

Here are the 6 things every working agent needs. Think of it as the kitchen setup before the chef walks in.

Subscribe now

Most non-techies build component 02 (a prompt) and call it an agent. Then they wonder why dinner is burnt.

I built a full guide that walks through each of these 6 components in detail. With examples, templates, and the exact files I use for my own agents. It’s free for subscribers. I’ll drop the link at the end.

How I’m building mine right now

Three real agents. All work in progress. All for brisk.

• Agent 1 reads my “things that inspired me / pissed me off this week” Notion page and finds the strongest content angles.

• Agent 2 drafts LinkedIn posts using my voice file and post templates.

• Agent 3 turns approved posts into branded carousel slides.

All three require my approval before anything goes live. Non-negotiable.

Here’s the build process:

A. Plan it in conversation

Before you build anything, you think. I open Claude (regular chat or Cowork) and write something like this:

I want to build an agent that helps me with my weekly LinkedIn content.

Here's my current workflow:
- Sunday: I pick a topic from my ideas list
- Monday: I draft a LinkedIn post
- Tuesday: I edit and schedule it

I want an agent that can:
1. Research trending AI topics from my list of sources
2. Draft a LinkedIn post in my voice (I have a voice file)
3. Create a carousel from the post content
4. Prepare everything for me to review before posting

What's the best way to structure this?

Claude helps me think through the logic. What inputs, what outputs, where the agent should pause for my approval. Don’t skip this phase.

Claude Code has a /plan mode that does the same thing. Type the command, describe the agent, and it maps the architecture before writing a single line of code.

B. Build it in Claude Code

Once the plan is clear, Claude Code reads your files, writes the code, runs it, fixes errors, and loops until it works. You watch and course-correct.

You don’t need to code. You need to know what you want, step by step, and describe it clearly. Claude Code writes the code. You provide the thinking.

I tell it something like:

Build me an agent that does the following:

1. Reads my content ideas from /content/ideas.md
2. Picks the top-priority topic
3. Researches it using web search
4. Drafts a LinkedIn post following the rules in /content/voice-file.md
5. Saves the draft to /content/drafts/ for my review
6. When I approve, generates a carousel using my brand guidelines
7. Prepares the final post and carousel for scheduling

A note on the terminal: Claude Code runs in one, and that scares a lot of non-technical people off. It shouldn’t. I wrote a full guide on how to set up Claude Code inside VS Code so you never have to touch a terminal directly. If that’s the wall you’re hitting, start there:

C. Run it, break it, fix it

The first version will be messy. That’s the point.

Run it on a real task. See what it gets wrong. Tell Claude Code to fix it. Iterate. Every cycle gets better.

What I learned building mine:

→ Start with ONE agent doing ONE thing. Not five.

→ Give it your voice file. Without it, output sounds like generic AI slop.

→ Always add an approval step. Always.

→ Keep the scope small. “Draft one LinkedIn post” is good. “Run my content strategy” is too much.

Share brisk.ai

The honest part: what can go wrong

→ Cost blowups. Agents running overnight can burn hundreds of euros in API tokens. Set a spending limit before you hit run. Always.

→ Security. OpenClaw is the cautionary tale. Broad permissions plus no boundaries equals chaos. Don’t give an agent access to anything you’d panic about losing.

→ Memory loss. Some agents forget everything on reboot. Your “employee” shows up with amnesia. Check what kind of memory the tool has before you trust it with anything serious.

→ The “it worked too well” problem. Agents do exactly what you tell them. The dating profile chef wasn’t broken. It was doing its job. Nobody told it where to stop. Governance (component 05) is not optional.

→ Complexity creep. One agent becomes three becomes ten. Suddenly nobody knows which agent did what or why the research agent is producing 61 files at 4 AM. That’s exactly where next week’s article comes in.

Where to start tonight

Open claude.ai. Free account if you don’t have one.

Give it one real task from your actual job. Not a joke, not a test. A real task. The proposal you’ve been putting off. The email you don’t want to write. The report you’d rather not start.

See what happens. Tell me in the comments.

That’s the whole assignment.

Next Thursday:

Level 5. When one chef becomes a restaurant.

I’ll walk you through Paperclip - the tool I wish existed when I was building my first messy brigade of agents at 3 AM.

Until then: which level are you on? I built a free cheat sheet that tells you exactly when a good prompt is enough, when you need a single agent, when Managed Agents make sense, and when you need the full Paperclip setup.

Subscribers get the full 6 Components Framework on how to set up AI Agents that actually work including my prompts in the Resource section.

You are not a subscriber? Change it now!

Subscribe now

This article is part of my ongoing series on making AI simple and useful for non-tech people. Subscribe to get future articles on AI tools that actually matter for everyday life.

Leave a comment

I know 17k doesn’t sound like much anymore. A few years ago it was a big deal.

I haven’t posted on it in over 4 years.

The algorithm kept changing. One month your content reached your audience. The next month Instagram decided who sees what. Then ads flooded the platform. Suddenly the game rewarded mass production over quality. Post more. Post faster. Post every day or disappear.

I refused to play that game. So I stopped.

I burned out. Walked away completely.

But I kept watching. And lately I noticed something. Creators are using AI to generate carousels and images at scale. Entire visual brands built in minutes.

The problem: a lot of it is slop. Generic. Soulless. The kind of content that makes you scroll faster, not slower.

But it doesn’t have to be.

I spent the last few weeks figuring out how to do this properly. Not mass-produced garbage. Carousels that actually look like someone with taste made them. Original images, not stock photos. Brand-consistent visuals, not random AI output.

I’m not going back to Instagram (at least not yet). But I’m not letting you miss out on this either.

Here’s the step-by-step system using Claude and Nano Banana 2.

The Problem With “Make Me a Carousel”

Open Claude. Paste an article. Type “turn this into an Instagram carousel.”

You’ll get text on slides. Technically correct. Visually dead. It looks like every other AI-generated carousel on the platform. Generic hook. Generic layout. Zero personality.

That’s not Claude’s fault. That’s a prompt problem.

Claude isn’t a design tool. Claude is a content strategist that happens to output files. When you treat it like a template machine, you get template results.

The fix is a system. Three pieces.

How I Actually Build Carousels Now

A. The brand system.

You need a document that tells Claude exactly how your brand looks and sounds. Brand name, target audience, color palette with hex codes, typography, voice description.

Upload it to a Claude Project as a Skill. Now Claude remembers your visual identity for every future conversation. No more “make this teal” or “use this font.” It already knows.

I built a brand system template and a prompt that generates one for any brand. I’ll share both with subscribers.

Subscribe now

B. The content prompt.

This is where the magic happens. You give Claude your long-form content and a specific prompt that tells it how to think about carousels.

Not “make me slides.” Something much more precise.

The prompt instructs Claude to:

→ Read your article and identify the strongest standalone ideas

→ Structure them into a narrative arc (hook slide, tension, payoff, CTA)

→ Decide how many slides the content actually needs (not always 10)

→ Write slide copy that works for the carousel format, not just chopped-up paragraphs

→ Match each slide to the visual energy it needs

Claude understands that slide 1 must stop the scroll. That the middle slides need to build tension. That the last slide requires a clear call to action. It’s not copying text onto rectangles. It’s reconstructing your content for a completely different medium.

I’ll share the exact prompt with subscribers. It’s the backbone of the whole system.

C. The visuals.

This is where Nano Banana 2 comes in. Instead of stock photos or solid-color backgrounds, you generate original images.

Nano Banana 2 is Google’s image generation model. It runs on Gemini 3.1 Flash. Costs about $0.04 per image. 4K resolution. Text rendering that actually works.

The workflow:

→ Find a visual aesthetic you like (Pinterest is perfect for this)

→ Screenshot the reference image

→ Upload it to Claude

→ Ask Claude to analyze the style and write a detailed image generation prompt

→ Copy that prompt into Nano Banana 2

Do this in batches. Spend one session generating 20-30 images. Over a few weeks, you’ll have a library of 60-100 reusable assets.

Then upload your batch of images to the same Claude Project alongside your article and brand system. Ask Claude to evaluate each image by mood, visual energy, and space for text overlay, then match them to slides.

I picked 3 random images just for demonstration purposes. You can see them repeating in my final carousel. Therefore, the more images you give Claude to pick from, the better.

Claude picks a high-energy image for the hook. A calm image for context. An image with negative space for a headline. Design decisions a human art director would make. Except it takes minutes.

Not bad for a first draft, huh?

But What If You Don’t Have an Instagram Yet?

Maybe you’ve been thinking about starting one. Or you have a brand but no visual identity to work with. You don’t know what cohesive even looks like for your content.

There’s a system for that too. And it starts with one thing: a URL.

Step 1: Brand research.

Open Claude Code (the terminal version of Claude). Paste this prompt:

You are a world-class brand strategist and creative director. I'm going to give you a brand's URL. Visit the website and extract everything needed to fully understand this brand.

Extract:
- The brand name and what it sells
- Its target audience in detail
- Its brand tone and voice
- Its visual style, colors, and aesthetic
- Its key messages and value propositions
- Its competitive positioning
- The emotions it wants to evoke in its customers

URL: [PASTE URL HERE]

Give me a complete brand intelligence report before we create anything.

Claude Code visits the website and builds the full brand profile. About 3 minutes.

Step 2: Generate 40 image prompts.

Once Claude Code has the brand report, paste this:

You are now a world-class advertising creative director specializing in creating scroll-stopping visual concepts.

Using everything you just learned about this brand, generate 40 complete image generation prompts for Nano Banana 2.

For each prompt:
- Make it specific to this brand's visual identity and audience.
- Vary the format across product, lifestyle, emotional, minimalist, and bold graphic styles.
- Make each prompt production-ready with fully described lighting, composition, mood, style, and subject.

Format: Ad 1: [FULL PROMPT] Ad 2: [FULL PROMPT] ... through Ad 40.

Each prompt should feel like it was written by a senior art director who has studied this brand for months.

About 5 minutes.

Step 3: Generate the images.

Take the 40 prompts in batches of 5. Paste each into Nano Banana 2 with this add-on at the end:

Square format, hyperrealistic, 8k, consistent brand colors, no text, no words, production-ready advertising visual.

Nano Banana 2 generates each image in seconds. 40 images in about 20 minutes.

Step 4: Quality review.

Go back to Claude Code. Paste this:

Here are the 40 ad prompts we generated for this brand: [PASTE THE 40 PROMPTS]

Now act as a creative director reviewing this campaign before sending it to the client.

For each ad tell me:
- Whether the concept is strong or weak for this brand.
- Which are the 10 strongest ads and why.
- Which 5 should be regenerated with better direction.
- What the improved prompt should be for each of the weak ones.

Give me a final shortlist of the 15 best ads, ranked by expected performance.

Claude edits its own work and leaves you with only the best. About 5 minutes.

Total time from URL to 15 polished visuals: under 35 minutes.

Here’s what makes this powerful. This isn’t just for your own brand. Run the same system for a client. New URL, same prompts, same workflow. You could charge good money per brand depending on the scope. One system. Unlimited clients. No design skills required.

The Part That Matters

Both systems share one principle: your AI is only as good as the system you give it.

A brand system document. A library of custom images. A content prompt that tells Claude how to think, not just what to do. That’s the difference between AI slop and content you’d actually post.

You don’t need to learn code. You don’t need Canva. You don’t need a design background.

You need structure. Claude runs the rest.

Tonight, open Claude. Create a new Project. Write a one-page brand system. Generate 5 images in Nano Banana 2. Upload an article you’ve already written. Ask Claude to build a carousel.

The first one takes 30 minutes. The second one takes 10. By the third, you’ll wonder why you ever did this manually.

The full prompt stack, the brand system template, and the content-to-carousel prompt are all packaged for subscribers. You’ll get them in the welcome email.

Btw - this approach works for any social media platform if you adjust the sizes.

Happy building.

Lu

Leave a comment

It wasn’t perfect. The design was a little plain, the layout needed adjusting, and I made every rookie mistake possible — starting without a clear brief, jumping straight into building before thinking, asking Claude to “make it look better” without telling it what better meant.

The second website took half the time and looked twice as good.

The difference wasn’t Claude. It was having a process.

This is article 4 in a series on Claude Code. So far we’ve covered setting up VS Code and GitHub, and publishing your first site with Vercel. If you missed those, start there — this guide assumes you have a working local environment and know how to deploy.

This one is about building websites that actually look good. Not just functional. Good enough to show a client, charge for, or publish under your own name.

By the end, you’ll have a repeatable process for building clean websites with Claude Code — the exact prompts I use at every stage, a trick for stealing any website’s design legally using your browser’s inspect tool, and the skills that make Claude dramatically better at design from the first line of code. The next two articles will cover how to present work to clients, handle feedback, and price what you build. But first, you need to be able to build it well.

This is that process. Start to finish, in plain English, with the exact prompts I use at every stage.

Before you build: the brief

The most important step happens before you open VS Code.

Non-technical people usually skip this. They open Claude, type “build me a website about my consulting services,” and then spend three hours correcting things that a five-minute conversation would have prevented.

Claude is literal. It builds exactly what you describe. If your description is vague, the result will be vague.

To solve this, I created a template I force myself to fill in before every project. Copy it, paste it into a document, answer every line. Then paste the whole thing into Claude as your first message.

PROJECT BRIEF

What is this website for?
[One sentence. Example: A portfolio site for a management consultant specialising in change management.]

Who will visit it?
[Example: Potential clients — HR directors and COOs at mid-market companies.]

What do you want visitors to do?
[Example: Read about my services and send me an enquiry via the contact form.]

How many pages?
[Example: 4 pages — Home, About, Services, Contact.]

What should the design feel like?
[Example: Clean, minimal, professional. Think McKinsey website, not a creative agency.]

Any websites you like the look of?
[Example: domain.com, domain.com — I like the layout and colour scheme.]

What content do I have ready?
[Example: My bio, a list of services, one case study, and a professional photo.]

What must it include?
[Example: Contact form, mobile-friendly design, links to my LinkedIn.]

What must it NOT include?
[Example: No stock photos, no pricing, no blog section.]

Read it back before you send it. If anything is vague, make it specific. Claude will thank you by building the right thing the first time.

Subscribe now

The right way to start a new project

Claude Code now works in two ways. Pick the one that fits how you work.

Option A: The VS Code extension (easier start)

Install the Claude Code extension directly from the VS Code Marketplace. It gives you a dedicated Claude panel inside VS Code — no terminal juggling, inline diffs showing exactly what changed, and keyboard shortcuts for common actions. If you’re not a developer by trade, start here. I wrote a step-by-step guide on how to do it:

Option B: The terminal (more control)

Open VS Code, create a new empty folder on your desktop, name it clearly — “ClientName-Website” or “MyPortfolio-2026” works fine. Go to File → Open Folder and select it. Then open the terminal inside VS Code.

Either way, the next step is the same.

Before you paste the brief, create a CLAUDE.md file for this specific project. Ask Claude:

Create a CLAUDE.md file for this project with these preferences:
- Always use clean, minimal design unless told otherwise
- Mobile-first — design for phone screens before desktop
- No stock photo placeholders — use solid colours instead
- Keep code organised and well-commented so changes are easy later
- Ask me before making any changes that affect the overall layout

Now paste your brief. Claude will read both the CLAUDE.md and the brief before writing a single line of code.

That combination — project preferences plus clear brief — is what separates a clean first draft from an hour of corrections.

The build process

Don’t ask Claude to build the whole website in one go.

I know that’s tempting. It feels efficient. It isn’t. Claude will make decisions about structure, design, and content all at once, and when you want to change one thing it often breaks another.

Build in three phases instead.

Phase 1: Structure

Ask Claude to build the skeleton first — pages, navigation, layout — with no real design and placeholder text.

Build the structure of this website based on the brief I gave you.
Create all the pages with navigation between them. Use placeholder text for content. No colours or styling yet — just clean structure. I want to review before we move to design.

Review it. Click through every page. Check the navigation works. Make sure the structure makes sense before moving on.

If something looks wrong, use /rewind in the chat to undo changes and try a different direction. This is your safety net — use it freely.

Phase 2: Design

Once the structure is approved, add the visual layer.

Now apply the design. Use the style references from the brief.
Keep it clean and minimal. Use this colour palette: [describe or paste hex codes]. Apply consistent typography throughout.
Show me the homepage design first before styling the other pages.

Review the homepage. Approve it. Then ask Claude to apply the same design to the remaining pages.

Phase 3: Content

Replace all placeholder text with real content.

Replace all placeholder text with the following content:
[paste your actual text here]

Keep the formatting consistent with the design. Don't change any layout or styling — only replace the text and images.

Three phases. Three review points. Far fewer surprises at the end.

Claude Code commands worth knowing

Claude Code has a set of slash commands you’ll use constantly while building. These five are the ones that actually matter for a website project.

• /rewind — Undoes your last change and restores the previous version. Use it the moment something breaks instead of trying to talk Claude back to where you were. It’s faster and cleaner.

• /compact — Summarises the conversation when sessions get long. Claude has a context limit. On a multi-hour build, hitting /compact keeps it from losing track of what you asked earlier. Run it between phases.

• /model — Switches between Claude models mid-session. If you’re doing a quick content change, you don’t need Opus. Drop down to Sonnet and save your usage for the complex build work.

• /cost — Shows what the session has used so far. Useful if you’re on API billing and want to keep an eye on spend as the project grows.

• /clear — Starts a fresh context without closing the session entirely. Use it when you’re switching from one page to another and want Claude focused only on the new task.

You don’t need to memorise all of these before you start. But /rewind and /compact will save you time on your first project.

What Claude builds well (and what it doesn’t)

Not every website is the same. Here’s what Claude handles easily versus where it will frustrate you.

Easy:

✦ Portfolio sites — one to a few pages, mostly text and images, clean layout

✦ Landing pages — single goal, clear structure, no complexity

✦ Brochure sites — 3-8 pages, about/services/contact format, standard business website

Medium (works with careful prompting):

✦ Simple contact or booking forms — Claude can build these but test them thoroughly before going live

✦ Basic image galleries — works well for small collections, gets messy at scale

Hard (not recommended without developer support):

✦ E-commerce — payment processing, inventory, checkout flows are beyond this scope

✦ Membership or login systems — requires backend infrastructure Claude can sketch but not safely build alone

✦ Anything connected to a database — beyond what this guide covers

If you want to build for clients and they ask for something in the Hard category, be honest. Refer them to a developer. It’s better than delivering something that breaks.

Making it look good without a designer

Claude’s default aesthetic is functional but forgettable.

The fix is giving it a clear visual direction before it starts designing. Vague instructions produce vague results. These three prompts work reliably:

For clean and minimal:

Design this with a clean, minimal aesthetic. White background, one accent colour (use #1a1a2e — dark navy), generous white space, simple sans-serif typography. Think less is more throughout.

For bold and modern:

Design this with a bold, confident aesthetic. Dark background (#0d0d0d), white text, one strong accent colour (use #ff6b35 — burnt orange). Large typography, strong contrast, modern feel.

For warm and professional:

Design this with a warm, approachable professional aesthetic.
Off-white background (#faf8f5), warm grey text, soft terracotta accent (#c17f5a). Feels human and trustworthy, not corporate.

Share

The result - 3 different designs received in seconds. Pick the one you like most or adapt the prompt any way you want.

If you have a website you like the look of, tell Claude directly:

I like the visual style of [website URL]. Use it as inspiration for the design — similar colour palette, similar typography weight, similar use of white space. Don't copy it, just use it as a reference.

Steal the design. Legally.

There’s a faster way to get the look you want — one almost nobody talks about.

If you find a website with a design you love, even a complicated one, you can copy its actual code and hand it directly to Claude. Claude reads code the same way it reads text. Give it the HTML and CSS of a site you admire, and it can rebuild that same aesthetic for your project.

Here’s how to grab it.

Step 1: Open the website you like in Chrome or Firefox. Right-click anywhere on the page and select “Inspect” (or press F12 on your keyboard). A panel opens on the right side of your screen — this is the code behind what you’re seeing.

Step 2: Click the “Elements” tab at the top of that panel. You’ll see the HTML structure of the page. To grab the CSS (the part that controls colour, fonts, spacing, layout), click the “Styles” tab that appears in the lower half of the inspector.

Step 3: To copy the full HTML, right-click the very first line in the Elements tab — it usually starts with <html — and select “Copy → Copy outerHTML.” For the CSS, you can select and copy the styles shown in the Styles panel, or go to Sources → find the .css file → select all and copy.

Step 4: Paste what you copied into Claude Code with this prompt:

Here is the HTML and CSS from a website whose design I like. I don't want to copy this website — I want you to use it as a design reference. Analyse the colours, typography, spacing, and layout approach, then apply a similar aesthetic to my project. Build in my content, not theirs.

Claude will extract the design logic — the colour palette, the font choices, the spacing system — and apply it to your site. You get the look without copying anyone’s content.

One thing to know: very complex sites (think large agency websites with lots of animations and interactive elements) will give Claude a lot to read at once. If the code is very long, focus on copying just the section you like rather than the entire page. A homepage hero section is usually enough to give Claude a strong design direction.

The skills that make Claude dramatically better at design

Claude Code is a capable designer by default. But skills turn it into something closer to having a senior designer reviewing every choice before the code gets written.

Skills are small instruction files you install once and they run in the background. Every time Claude touches design work, it reads those instructions first. The results are noticeably different.

Here are the ones worth installing for website builds.

The Frontend Design skill (start here)

This is Anthropic’s official skill for design quality, and the most installed design skill in the Claude Code ecosystem. It gives Claude a design system and philosophy before it touches any code — producing distinctive typography, purposeful colour choices, and layouts that don’t look like every other AI-generated site.

Without it, Claude defaults to what developers now call “AI slop” — the same Inter font, the same purple gradient, the same card layout on every project.

Install it with one command in your terminal:

claude plugin add anthropic/frontend-design

Once installed, invoke it at the start of any design phase:

/frontend-design

If the terminal feels unfamiliar, don’t let it stop you. Reply to this email or drop a comment below and I’ll send you the skill files directly. You can upload them straight into Claude Code without touching the terminal at all.

The Web Design Guidelines skill (quality check)

Where the Frontend Design skill focuses on creative direction, this Vercel skill focuses on correctness. It reviews your UI code against 100+ rules covering accessibility, performance, and UX best practices. Think of it as a checklist running in the background — catching things like buttons without proper labels, text that’s too small on mobile, or images missing descriptions.

Install it:

npx skills add vercel-labs/agent-skills --skill web-design-guidelines

The Accessibility skill (don’t skip this)

Accessible websites rank better on Google and work for more people. This skill catches the errors that are easy to miss: keyboard navigation that breaks, form inputs without labels, colour contrast that fails on bright screens.

Install it:

npx skills add supercent-io/skills-template@web-accessibility -g -y

The full install block (if you want all three at once)

claude plugin add anthropic/frontend-design
npx skills add vercel-labs/agent-skills --skill web-design-guidelines
npx skills add supercent-io/skills-template@web-accessibility -g -y

Run this in your terminal once. Skills install globally — you don’t repeat it for every project.

After installing, ask Claude: “What design skills do you have installed?” It will confirm which ones are active. Then start your build the same way as before, and watch the difference in what it produces.

A note on Claude Cowork

Anthropic launched Cowork in early 2026 as a graphical, no-terminal tool for non-technical users. If someone you know is too uncomfortable with VS Code to follow this guide, Cowork is worth showing them.

For building actual websites, though, Claude Code gives you more control — you can see the files, version them properly, and host them cleanly. This series stays focused on Claude Code for that reason.

Testing before you call it done

Never hand over a website — or publish your own — without running this checklist.

Better yet, ask Claude to run it first:

Before I review the final version, go through this checklist and fix anything that fails:

1. Does every navigation link work?
2. Does the contact form send correctly?
3. Is every page fully responsive on mobile screens?
4. Are there any broken images or missing assets?
5. Is the page load time reasonable (no unnecessary large files)?
6. Does it look correct in both light and dark browser modes?

Tell me what you checked and what you fixed.

Then do your own manual check on top of Claude’s.

Open the site on your phone. Click every button. Fill in the contact form and see if you receive the email. Open it in a different browser. Ask one other person to click through it cold and tell you if anything confused them.

Ship nothing you haven’t tested on a real phone with real fingers.

Your build checklist

✦ Brief completed before opening VS Code

✦ New project folder created and opened (or VS Code extension installed)

✦ CLAUDE.md created with project-specific preferences

✦ Brief pasted as first Claude message

✦ Design skills installed (/frontend-design + web-design-guidelines + accessibility)

✦ Phase 1 (structure) reviewed and approved

✦ Phase 2 (design) reviewed and approved

✦ Phase 3 (content) added and reviewed

✦ /compact run between long phases

✦ Claude checklist run before final review

✦ Manual test on mobile completed

✦ One other person reviewed it cold

That’s a website built properly. Not just built.

Let me know in the comments how your build went. I’m genuinely curious.

Lu.

Leave a comment

P.S. - 3 ways to go deeper:

1. Start from the beginning — the full Claude Code beginner’s guide →

1. Next up: You built it for a client. How do you present it, handle feedback, and hand it over without the chaos? Article 5. → coming soon.

2. Coming soon: What do you charge for a website you built in an afternoon? How to price your work when AI does most of the building. Article 6. → coming soon.

This article is part of my ongoing series on making AI simple and useful for non-tech people. Subscribe to get future articles on AI tools that actually matter for everyday life.

Subscribe now

Every other day, I talk to someone who is genuinely anxious about AI. They’ve heard of Claude. They’ve heard of Claude Code. The FOMO is real and it is loud, and nobody in that conversation seems to know where to actually start.

And when I ask them if they’ve heard of NotebookLM - a tool that is mostly free, requires zero technical knowledge, and is one of the most genuinely useful things Google has ever built - they look at me like I made it up.

That’s the gap I want to close with this guide.

Because here’s the irony

NotebookLM is the exact tool that helps you bridge the FOMO. You don’t need to understand how AI works to use it. You don’t need to write clever prompts. You just need documents or links. And it turns those into slide decks, audio overviews, cinematic videos, infographics, and structured research - mostly for free - so you can actually learn and work with all the other tools you’ve been afraid of missing out on.

I’ve been using it weekly for over a year. I’ve tested it against every other research tool available. Nothing else stays open as consistently.

And to show you exactly how to fight that FOMO - not just talk about it - I'm going to use a real topic as the working example throughout this entire guide: how to start learning to use Claude Cowork. Every workflow, every output, every prompt you'll see here is built around that topic. So by the end, you'll have a complete NotebookLM playbook and a head start on the tool that everyone around you is panicking about.

This is the guide I wish existed when I started.

What NotebookLM Actually Is

NotebookLM is not a chatbot.

Every AI tool you’ve used before - ChatGPT, Claude, Gemini - works from its own knowledge base. It was trained on the internet. When you ask it a question, it draws from everything it knows. That’s useful, but it also means you have no control over what it pulls.

NotebookLM is different. It only knows what you put in.

You create a notebook. You upload your sources. Every answer NotebookLM gives you comes exclusively from those sources, with citations that link directly back to the original document. If your sources don’t contain the answer, it tells you that instead of making something up.

The interface has three panels:

→ Sources - where you upload and manage your documents

→ Chat - where you ask questions and get cited, grounded answers

→ Studio - where you generate outputs: slides, audio, video, infographics, reports, and more

Everything lives inside a notebook. One notebook per project, per client, per research question. That structure is the whole point.

Why I Use NotebookLM Before Any Other AI Tool

Here’s what happens when you open ChatGPT or Claude with a research question.

You ask one thing. The model gives you a solid answer. You ask a follow-up. Suddenly you’re exploring a tangent you didn’t plan for. Twenty minutes later, you’re deep in a rabbit hole that started with “explain this concept” and ended somewhere near “here are six adjacent frameworks you might find useful.” The model is intelligent and generous. That’s exactly the problem.

NotebookLM solves this. By design.

Because it can only answer from what you give it, there are no detours. No creative tangents. No training data bleeding in. You stay focused on your topic because that’s all the tool can see.

Here’s my actual workflow - and I’ll use the exact example from this guide to show you: learning how to use Claude Cowork.

Subscribe now

→ Step 1: I found the four best YouTube tutorials on Claude Cowork from AI creators I trust - people who actually use the tools, not just talk about them - plus the official Anthropic Cowork tutorial pages. I uploaded all of them as sources - you can either name them clearly or let NotebookLM do the job. If I weren’t sure I had enough sources, I could ask NotebookLM right there: “What perspectives or sources am I missing to fully understand how to get started with Claude Cowork?” It will tell you.

→ Step 2 (most IMPORTANT to me): I asked it: “Summarize all sources into one overview - what Claude Cowork is, how it works, and the three things a complete beginner should do first.” The output was clean, cited, and covered exactly what I needed. No tangents about Claude Code, no detour into AI architecture. Just Cowork, from the sources I chose. I saved that summary as a note called “Cowork Beginner Summary” and converted it into a source (watch a video of how exactly I did it below) .

→ Step 3: I ran Studio tools on that single summarized source - not on the five raw originals. The slide deck it built from the tight summary was sharper and more focused than anything I could have prompted from scratch. The infographic was cleaner. Every output reflected exactly the angle I’d chosen, not the full noise of five unfiltered documents (I am showing details below).

→ Step 4: When I wanted to go further, I exported the summary and opened Claude. I pasted it as the entire context and said: “Based only on this, build me a 30-day learning plan for Claude Cowork.” No rabbit holes. No training data bleeding in. The model worked exclusively on what I gave it.

NotebookLM isn’t competing with Claude or ChatGPT. It’s the step before them. It’s how you stay in control.

Free, Pro, or Ultra? Here’s What You Actually Need

This is not a “you need to pay to get value” situation. The free tier is genuinely powerful.

Here’s what each plan gives you:

Free (no subscription)

A word of honesty: slide decks, infographics, and video overviews technically exist on the free tier but you will hit the cap fast - likely after one or two generations per notebook. If those outputs are central to your workflow, free will frustrate you quickly. Everything else - chat, mind maps, Deep Research, flashcards, audio overviews - works well within the free limits for most people.

Pro (~$20/month via Google AI Pro or Google One AI Premium)

Pro makes sense if you’re maintaining large research bases, running client work, or already using Google Workspace and Gemini across your apps. The storage and Gemini Advanced often justify the price independently of NotebookLM.

Ultra (~$250/month via Google AI Ultra)

Ultra is not for individuals with one notebook. It’s for creators, agencies, and teams producing large volumes of content who will actually use the cinematic video feature and bundled creative tools heavily. If you only want NotebookLM, the free or Pro tier covers almost everything.

Your Sources: The Foundation of Everything

Every output NotebookLM creates is only as good as the sources you give it.

This is the most important sentence in this guide. Not the prompts. Not the tier. The sources.

What you can upload

• PDFs

• Google Docs

• URLs (web pages, articles)

• YouTube videos (it reads the transcript)

• Images

• Text files and other machine-readable documents

How to structure your notebooks

One notebook per coherent topic or project. Don’t mix a client brief with a market research collection - they belong in separate notebooks.

Name your sources descriptively. Or let NotebookLM name them for you. For the Claude Cowork notebook, my sources looked like this:

Each one named. Each one purposeful. When I toggle only the YouTube sources, I get practitioner perspectives - real people who tested the tool. When I toggle only the Anthropic docs, I get official guidance. Comparing them reveals the gap between what’s documented and what people are actually doing in practice.

Share

The note-to-source loop

This is NotebookLM’s most underused feature. When a chat answer is good - when it cleanly synthesizes what you need - save it as a note. Then convert that note into a source.

For the Cowork example: after asking NotebookLM to summarize all nine sources, I saved the output as “Cowork Beginner Summary” and made it a source. Every slide deck, infographic, and quiz I generated after that was built from that one clean synthesis - not from raw youtube videos and tutorials. The outputs were tighter because the input was tighter. That’s the whole point.

Watch the video to see the whole workflow.

The Studio: Every Output, Explained

The Studio is where NotebookLM transforms your sources into deliverables. Each format has its own card. Click it, optionally add a prompt, and let it run.

Here’s every tool, and how to actually use each one.

1 - Mind Maps

A mind map visualizes the main topics and connections across your sources as a branching diagram. Click the Mind Map chip in chat, and NotebookLM builds it from your loaded sources.

Once generated, you can zoom, pan, expand and collapse branches, and click any node to ask a targeted follow-up question in chat. Use this at the start of a new project to see how your sources cluster before you dive deep.

For the Cowork notebook, the mind map surfaced three clusters I hadn’t consciously organized: core features, workflow patterns, and how Cowork differs from Claude Code. I clicked the “workflow patterns” node and asked: “Which of my sources covers this most thoroughly?” One YouTube video turned out to cover it in far more depth than the others. I wouldn’t have known that without the map.

2 - Audio Overviews + Interactive Mode

This is the feature that made NotebookLM famous. Two AI hosts discuss the key ideas in your sources in a podcast-style conversation. You choose the format - deep dive, debate, overview - and NotebookLM scripts and records it from your materials only.

The feature most people miss is Interactive Mode. After generating an overview, press Join. When the hosts call on you, ask a question in real time. The hosts answer using your sources, then continue the conversation. It turns a passive listening experience into a study session you can run while commuting.

3 - Standard Video Overviews

Standard Video Overviews extend the audio experience into visual form. The output is a narrated video combining images, charts, and text pulled from your sources. You can direct them for different audiences - “explain this for a non-technical executive” or “focus on the technical implementation for a developer.”

Available across all tiers, with higher daily limits on paid plans.

4 - Cinematic Video Overviews [Ultra only]

Launched in March 2026, Cinematic Video Overviews are a different category of output entirely. Instead of narrated slides, you get a fully animated explainer video - coherent visuals, narration, and a story arc - generated by a three-model AI stack including Gemini 3 and Veo 3.

Here’s the generation flow:

→ Build a notebook with well-structured sources → Open Studio, choose Video, then select Cinematic → Add a prompt to direct it: “3-minute explainer for a non-technical audience” or “marketing overview for a product launch” → Wait a few minutes for processing → Review and download your 2-5 minute cinematic video

Limit: 20 cinematic videos per day on Ultra. Available in English to users over 18. This is the flagship feature that justifies the Ultra price - for creators and agencies producing high-quality explainer content at scale.

5 - Slide Decks + Prompt-Based Revision

Click the Slide Deck card and NotebookLM builds a 12-15 slide presentation from your sources - with a narrative arc, key data points, speaker notes, and visual design. No prompt required for the first draft.

The real power is in revision. After the initial deck generates, use the Revise button to iterate with natural language: “Make slide 3 more visual.” “Rewrite the speaker notes for a technical audience.” “Add a comparison table on slide xyz.”

For the Cowork notebook, I ran the deck card on my “Cowork Beginner Summary” source with this prompt: “Create a beginner-friendly presentation on how to start using Claude Cowork. Audience: non-technical professionals. Tone: practical and direct. 10 slides max.” The first draft was 90% there. One revision - “add a slide comparing Cowork to Claude Code for someone who’s heard of both but used neither” - and it was done.

Pro tip: create a reusable specification note that describes your target audience, preferred tone, slide count, and visual style. Reference it in every deck prompt for consistent output across projects (I used my brand colors).

6 - Infographics

NotebookLM generates infographics that visualize your notebook content. You can incorporate brand assets - upload a logo or mascot as a source, create a note with the copy you want featured, convert that note to a source, then trigger the infographic card with a prompt specifying style and audience.

For the Cowork notebook, I used my summarized source and prompted: “Create an infographic showing the 5 first steps to getting started with Claude Cowork. Audience: non-technical professionals. Keep it visual and scannable.” The result covered exactly those five steps - nothing more, nothing outside my sources. Like Magic.

That’s the whole point of building from the summary first.

Ultra accounts can remove NotebookLM watermarks from exported visuals.

7 - Reports, Study Guides, Flashcards, and Quizzes

The Studio includes cards for structured reports, study guides, flashcards, and practice quizzes - all generated from your uploaded sources and nothing else.

For educators: upload your lecture notes, readings, and textbook chapters. Generate a complete study package in a few clicks - study guide, flashcard set, practice quiz - grounded in your actual course materials.

For professionals: the report card turns a research collection into a structured document with sections and citations faster than writing it yourself.

8 - Data Tables → Google Sheets

Ask NotebookLM to compare your sources across any dimension - competitor pricing, feature sets, research findings - and the Data Table card builds a structured table. Export it directly to Google Sheets.

This bridges qualitative research and quantitative comparison. It’s particularly useful for competitive analysis, literature matrices, and product feature work.

9 - Deep Research

Deep Research is the Studio’s most powerful mode. Give it a question, and it runs a multi-step research process across your sources - and in some configurations, across the web - and outputs a structured report with full citations.

Import that report back into your notebook as a source. Now you’ve bootstrapped a curated synthesis that feeds all your other outputs. Free users get approximately 10 Deep Research runs per month. Pro users get around 20 per day. This is one of the most compelling reasons to consider upgrading.

The Full Workflow: Research → Learn → Analyze → Ship

Every NotebookLM project follows the same four-phase pipeline. Here’s how it played out with the Claude Cowork notebook.

→ Research: I uploaded the five YouTube tutorials and the Anthropic docs. I asked NotebookLM: “What are the five most important things a non-technical professional needs to understand before using Claude Cowork for the first time?” The answer was grounded, cited, and took about 30 seconds. I ran the same question again with only the Anthropic docs toggled on, then again with only the YouTube sources. The gap between official documentation and what creators were actually teaching was immediately visible. I saved the synthesized answer as a note, converted it to a source.

→ Learn: I generated a mind map to see the main clusters - core features, workflows, and the Cowork vs. Claude Code comparison. Then I created an audio overview and listened to it on a walk. By the time I sat back down, I had a working mental model of the tool without having opened it once. I ran a short quiz from the Studio to check what had stuck.

→ Analyze: Back in chat, I asked: “Based on all sources, what is the single biggest mistake someone new to Claude Cowork is likely to make in their first week?” One source had a very specific answer. I wouldn’t have found it without asking. I added the answer to my notes and flagged it for the infographic.

→ Ship: From the summarized source, I generated the slide deck, the infographic, and an audio overview formatted as a beginner onboarding guide. Three different outputs. One source. Fifteen minutes total. Each one would have taken me hours to build manually - and none of them would have been as grounded.

Prompting patterns that work well:

✦ “Summarize all sources in 500 words - what [topic] is, how it works, and what a beginner should do first.”

✦ “What perspectives or sources am I missing to fully understand [topic]?”

✦ “What is the single biggest mistake someone new to [tool] is likely to make?”

✦ “Using only the [specific source], what does it say about [subtopic]?”

→ The most powerful prompt is always the simplest: “What am I missing?”

NotebookLM + Gemini: The Combination That Changes Everything

NotebookLM and Gemini are not competing products. They are designed to work together.

The concept: NotebookLM is your second brain - curated, grounded, cited knowledge from sources you control. Gemini is the builder - the model that takes that knowledge and turns it into apps, websites, campaigns, and multimedia projects.

When you connect them, your research base becomes the foundation for everything else you build.

How to connect them:

→ Create or open a notebook in NotebookLM and upload your sources

→ Open Gemini in your browser or mobile app and start a new chat

→ Click the + attachment icon in the chat input and select NotebookLM from the source list

→ Choose one or more notebooks to attach

→ Ask questions or give instructions - Gemini answers using your notebooks as live context

Note: the NotebookLM attachment option is still rolling out. Some accounts may not see it yet, and availability differs between personal, Workspace, and education accounts.

Four workflows worth building

1. Gem as persistent advisor

Create a Gemini Gem - a custom AI agent - that always references a specific NotebookLM notebook. As you add sources to the notebook, the Gem updates automatically. Build a “Content Strategy Advisor” Gem for your research notebook. Build a “Course TA” Gem for your learning notebook.

2. Prototype builder

Attach a research notebook to Gemini and ask it to build a working tool - an audit app, a recommendation engine, a segmentation dashboard - in Gemini Canvas. The entire build is grounded in your actual research, not generic training data.

3. Cross-notebook synthesis

Attach multiple notebooks to a single Gemini conversation. Combine a product research notebook and a brand voice notebook to generate a campaign. Combine a competitor analysis notebook and a user interview notebook to generate positioning recommendations.

4. Content production pipeline

Use NotebookLM to research and structure. Then use Gemini - with notebooks attached - to build the assets: write the landing page, generate the images, produce the video. No manual copy-paste. No context loss.

The 5 Most Powerful Use Cases

1. Academic Research and Literature Reviews

Upload 20-30 papers. Use chat to surface contradictions, identify key themes, and find gaps. Run Deep Research to generate a structured literature overview. Use the Data Table card to build a comparison matrix of studies, then export to Sheets.

From there: generate the presentation, the study guide, the practice questions. All from the same notebook, all traceable to the same sources.

2. Teaching and Course Creation

Upload your textbook, lecture notes, and reading list. Generate a complete study package: study guide, flashcard set, quiz, and audio overview. Create a conversation-only notebook and share it with students as an interactive study companion - they ask it questions, it answers only from your course materials.

Then use the slide deck feature to generate your lecture slides.

3. Content Creation and Repurposing

Upload your past articles, research notes, and audience insights. Generate outlines, scripts, and talking points. Use the audio overview to turn an existing article into a podcast-style segment. Convert a strategy document into a full course complete with slides, quizzes, and landing page copy.

For video creators: upload YouTube transcripts, generate a new video script with chapter headings, then use Gemini with the attached notebook to draft B-roll suggestions and titles.

4. Business Strategy and Client Work

Aggregate customer interviews, competitor documentation, market reports, and internal strategy docs into one notebook. Run Deep Research. Build a data table comparing competitors. Export to Sheets. Generate the investor deck or board presentation directly from that synthesized base.

Connect to Gemini to build the tools that come next - dashboards, audit frameworks, recommendation engines - grounded in your real research, not a generic model prompt.

5. Personal Knowledge Management

Upload articles you’ve saved, book notes, podcast transcripts, and course materials. Use chat to make sense of them. Use mind maps to see how concepts connect. Use flashcards and quizzes to reinforce material over time.

NotebookLM answers are always traceable to specific sources. For anyone who cares about verification, this is the AI tool that actually earns trust.

Share

Tonight, open NotebookLM.

Create a new notebook. Call it “Claude Cowork - Getting Started.” Find three to five YouTube videos from AI creators you trust, add the Anthropic Cowork tutorial page, and upload them all as sources. Ask NotebookLM to summarize everything into one beginner overview. Save it as a note. Make it a source.

Then run the Slide Deck card. See what comes out.

That’s it. That’s the whole workflow. You just learned two tools at once - and you didn’t open a single rabbit hole to do it.

Lu

PS - 3 ways to go deeper:

1. AI, Simplified. (my weekly newsletter) Practical AI for professionals who don’t code. Free. →

   AI, Simplified.

   Every week, I break down AI tools and strategies for professionals who don't code.

   By Lu Tomšić

2. NotebookLM (free access) Everything in this guide works on the free tier. No credit card required. → https://notebooklm.google.com/

3. NotebookLM Help Center Official documentation for every feature, updated as the product changes. → https://support.google.com/notebooklm/

Let me know if I was able to convince you how great this tool is in the comments!

Leave a comment

Nobody else can see it.

It lives in a folder on your computer. You can open it, click around, see how it looks. It works.

But it only works for you - on your machine, in your browser, with nobody else invited. That’s the gap between “I built a website” and “I have a website.” Hosting is what closes it.

The good news: getting your site live is faster than building it. I’ll show you every option, what each costs, and how to pick the right one for your situation - including how to use the provider you’re probably already paying for.

One note before we start:

I use Vercel personally, and I’ll spend more time on it than the alternatives. Not because it’s the only option or even objectively the best one. Just because it’s the tool I know from the inside - and I’d rather give you one detailed walkthrough of something I’ve actually used than five shallow descriptions of tools I haven’t. The alternatives are real options. I’ll cover those too.

What “hosting” actually means

Think of your website like a house. The code Claude wrote is the structure - the walls, the rooms, the furniture. But a house needs land. Without land, there’s no address. Without an address, nobody visits.

Hosting is the land. It gives your website a permanent address on the internet. Someone types your URL, the hosting provider finds your files on a server somewhere in the world, and the visitor sees your site.

The key thing to understand: your files need to live on a computer that is always on, always connected to the internet, and always ready to serve them to whoever asks. Your laptop doesn’t qualify - it goes to sleep, disconnects from wifi, travels between locations. A hosting provider’s server doesn’t. That’s what you’re paying for.

Every website you’ve ever visited runs on this exact setup. This is how you join the rest of the internet.

Share

Vercel - where I’d tell you to start

Vercel is my personal first choice for new projects, and it’s become the default recommendation in most AI-assisted development workflows. Here’s why.

You connect Vercel to your GitHub repository (from Article 2 in this series). From that point on, Vercel watches your repository like a hawk. Every time you push a new commit to GitHub, Vercel automatically detects the change, rebuilds your site, and updates the live version - usually within 30-60 seconds. No manual uploading. No FTP clients. No logging into dashboards to hit “publish.”

Your site gets a free URL immediately - something like yourproject.vercel.app. You can connect a custom domain (yourname.com) whenever you’re ready. The domain itself costs money (typically €10-15/year from any domain registrar), but Vercel doesn’t charge extra to connect it.

One feature worth knowing about: every deployment gets its own unique URL. So if you push a change and something breaks, you can instantly roll back to a previous working version by just switching which deployment is “live.” It’s version control for your live site.

Pricing:

✦ Free tier → Generous. Covers personal projects, portfolios, small client sites, prototypes. No time limit.

✦ Pro → $20/month. Needed for commercial projects, higher traffic, team collaboration features, or if you need more bandwidth and build minutes.

Best for

First deployments, personal projects, anyone who wants automatic updates without thinking about them. Also the best choice if you’re following this series and building in VS Code with Claude Code - the workflow is designed with this kind of deployment in mind.

My honest caveat

Vercel is not the only good option. If you already have hosting set up elsewhere, or if Vercel’s terms don’t fit your use case (it restricts commercial use on the free tier), the alternatives below are just as capable.

Subscribe now

The alternatives (and what they actually cost)

Vercel is not special. It’s just the one I happen to use. Here’s the full picture.

Netlify

Functionally nearly identical to Vercel for most use cases. Connect GitHub, automatic deploys on every push, free custom domain support, SSL certificate included. The interface looks slightly different and some developers swear by specific Netlify features for certain project types, but for a portfolio site or client landing page you will not notice the difference.

✦ Free tier → Generous, similar to Vercel. 100GB bandwidth per month, 300 build minutes.

✦ Pro → $19/month

If you try to sign up for Vercel and hit a problem, go straight to Netlify. Same outcome.

GitHub Pages

Built directly into GitHub - no separate account or platform needed. You enable it in your repository settings (under Settings → Pages), choose your branch, and your site is live at yourusername.github.io/projectname. That’s it.

✦ Free → Always. There’s no paid tier for basic hosting.

The limitation

GitHub Pages works for static websites only. Meaning sites that are just HTML, CSS, and JavaScript with no server-side logic, no database, no backend. For a portfolio, a landing page, a brochure site for a client - it’s completely fine. For anything more complex, you’ll need one of the other options.

Cloudflare Pages

Cloudflare runs one of the largest network infrastructures in the world. Their Pages product uses that same network, which makes it very fast - your site loads from whichever server is geographically closest to your visitor. For a European audience, this matters.

Setup is slightly more involved than Vercel, but the same basic model applies: connect GitHub, automatic deploys, custom domain support.

✦ Free tier → Very generous. Unlimited bandwidth, unlimited sites, 500 builds per month.

✦ Pro → $20/month

Best for: Projects where loading speed is a priority, or if you’re already using Cloudflare to manage your domain (which many European businesses do).

Render

A more flexible platform that handles both simple websites and more complex applications that need a backend or database. Slightly more technical to configure than Vercel, but scales better as projects grow beyond static files.

✦ Free tier → Available, with one important caveat: free tier services “spin down” after inactivity, meaning the first visitor after a period of quiet might experience a slow load while the server wakes up.

✦ Paid → Starts at $7/month to eliminate the spin-down issue.

Best for: Projects that might grow into something more complex, or if you anticipate needing a backend alongside your site later.

Can I use the hosting provider I already pay for?

Here’s a situation I’d guess describes a lot of you.

You already pay for hosting. Maybe you have a WordPress site there, or an older website a developer built years ago. Your domain lives there, your email runs through it, your whole digital setup points to one provider. You log into one dashboard and everything is under one roof.

Adding Vercel means another account, another login, another line on your invoice, another platform to learn when something goes wrong. For a simple website that won’t change much, that overhead feels unnecessary. And honestly - it might be.

The providers most commonly used by businesses and freelancers include:

→ IONOS (German-founded, 6+ million customers worldwide, one of the most widely used hosts globally - very common across Europe)

→ Hostinger (Lithuanian-founded, one of the fastest-growing hosts in the world, used in 150+ countries - popular pick for first-time site owners)

→ SiteGround (Bulgarian-founded, globally recognized, excellent reputation for ease of use and support - widely used by non-developers across Europe and beyond)

→ GoDaddy (the world's largest domain registrar - if you bought your domain somewhere, there's a decent chance it was here, and yes, they do hosting too)

→ Namecheap (popular internationally, known for straightforward pricing with no surprise renewal hikes - common among freelancers and small businesses)

→ Infomaniak (Swiss, privacy-focused, strong choice if DSGVO compliance matters to you)

If you’re already paying for one of these - or any similar provider - your existing plan almost certainly includes the ability to host a simple website. You don’t need to add anything.

Here’s how to do it:

Step 1: In VS Code, ask Claude to “build the final production version of this project.” It will generate a folder called dist or build containing all the compiled, ready-to-deploy files. These are the files your hosting provider needs.

Step 2: Log into your hosting provider’s control panel. This is usually called cPanel, Plesk, or a custom dashboard depending on your provider.

Step 3: Find the File Manager (or equivalent). Navigate to the folder where your website should live. For most providers this is called public_html, htdocs, or a subfolder named after your domain.

Step 4: Delete any existing placeholder files (most providers put a default “coming soon” page there). Upload the entire contents of your dist or build folder directly into that location - not the folder itself, the contents inside it.

Step 5: Your site is live. Visit your domain to confirm.

The honest pros and cons

Pros

→ One invoice, one dashboard, everything in one place → You already know how it works → Domain, email, and SSL certificate usually included in what you already pay → No new platforms to manage or troubleshoot

Cons

→ No automatic deploys - every time you make changes, you repeat the manual upload process → Rollback means re-uploading the previous version of your files manually → Not suitable for complex web applications that need a backend (fine for portfolios, landing pages, brochure sites) → Some shared hosting providers have slower load times than cloud-native platforms like Vercel or Cloudflare

My recommendation: if your site is a portfolio, a landing page, or a simple client site that won’t change weekly - your existing European provider is completely fine. You’ll save yourself a new account and it’ll work just as well for what you need. If you’re building something that will change frequently, or something with complex functionality, Vercel or Netlify will save you a lot of time over the long run.

Share

The full picture, start to finish

This is the complete workflow across all three articles in this series:

Claude Code builds it. You describe what you want in plain English. Claude writes the code, creates the files, builds the thing. No terminal knowledge required.

GitHub saves it. Every session, you commit and push. Every version is backed up. Nothing gets lost. If Claude breaks something, you roll back in two clicks. This is your safety net.

Your host publishes it. Whether that’s Vercel auto-deploying every push, or you manually uploading your dist folder to your existing European provider - your site gets an address. The world can visit.

That’s the full stack. No developer friends. No expensive agency. No three-weekend learning curve.

Hosting checklist: how to verify your site is live

✦ Production build created (ask Claude: “build the final production version of this project”)

✦ Hosting platform chosen - Vercel or Netlify for automatic deploys, existing provider for simplicity

✦ Site deployed and live URL confirmed

✦ Custom domain connected (optional but recommended if you have one)

✦ Test on mobile - does it look right on a phone?

✦ Test in a second browser or ask a friend to visit the URL

✦ Share the link with one person and ask for feedback

You built something. It’s live. Now other people can see it.

That’s not nothing.

Show me in the comments.

I will take a look and give feedback if you ask for it.

Leave a comment

P.S. - 3 ways to go deeper:

1. Missed the Claude Code beginner’s guide? Start here →

1. Next up: How to build simple websites for yourself or clients with Claude Code — and how to package and deliver them professionally

2. Coming soon: How to migrate your website to a different host — including how to move it to the provider you already pay for, when that makes sense and when it doesn’t

This article is part of my ongoing series on making AI simple and useful for non-tech people. Subscribe to get future articles on AI tools that actually matter for everyday life.

Subscribe now

Before we go further, we need to talk about something. Only loosely related to Claude Code. But important enough that I almost lost everything because I ignored it.

GitHub.

A tool I avoided for years. Not on purpose. But every developer who mentioned it would immediately follow up with something that sounded like word soup to me: forking, pulling, repositories. So I stayed away.

Until it was almost too late.

I almost lost everything.

Three hours of Claude building a client proposal tool. Files everywhere. Progress I couldn’t recreate from memory.

Then my laptop froze.

I sat there staring at the spinning wheel thinking: I have no backup. If this crashes, it’s gone.

It didn’t crash. That time.

But that afternoon I finally set up GitHub. It took twenty minutes. I’ve never worried about losing a project since.

If you followed my first Claude Code guide, you have a folder on your desktop with everything Claude built. That folder is fragile. One crashed computer, one accidental delete, one “oops wrong folder” and it’s gone.

GitHub fixes that. And you don’t need to touch a terminal to use it.

What is GitHub, actually?

Think of it like Google Drive, but for code.

Every time you save your project to GitHub, it remembers that exact version forever. You can go back to any previous save at any time. You can see exactly what changed between saves.

It’s free for personal projects. And there’s a visual desktop app that means you never have to open that black terminal screen.

Subscribe now

The GitHub dictionary

Before you touch anything, read this section.

GitHub has its own language. Developers use these words constantly and never explain them. Here’s every term you’ll encounter, explained like you’re smart but have never heard any of this before.

Repository (repo)

Your project’s home on GitHub. One project = one repository. Think of it like a dedicated Google Drive folder for a single project — but with superpowers.

Commit

A save point with a label. Like pressing save in Word, except GitHub remembers every single save you’ve ever made and lets you go back to any of them. When you commit, you write a short note like “added contact page” so you remember what you changed.

Push

Sending your latest commits from your computer up to GitHub. Like uploading a file to Google Drive. Until you push, your saves only exist on your computer.

Pull

Downloading the latest version from GitHub to your computer. Useful if you work on multiple computers or if someone else made changes to the project.

Branch

A parallel copy of your project where you can experiment without touching the original. Like doing “Save As” before making big changes — except cleaner. If the experiment works, you merge it back. If it doesn’t, you delete the branch and the original is untouched.

Merge

Combining a branch back into your main project. You tested something on a branch, it worked, now you’re folding it back into the official version.

Main (or Master)

The primary, official version of your project. The one that’s live. Everything else is a branch off this.

Clone

Downloading someone else’s GitHub project to your own computer. Common when using open source tools or templates.

Fork

Making your own personal copy of someone else’s repository. You get the full project, but in your own GitHub account. Changes you make don’t affect the original.

Pull Request

A formal proposal to merge changes into a project. More relevant when working in teams — someone makes changes on a branch and asks the team to review before merging. You won’t use this much working solo with Claude.

READ.ME

A text file that sits at the top of every repository explaining what the project is, how to use it, and any setup instructions. Think of it like a cover page or a handover document.

Why this matters for Claude Code specifically

Your desktop folder works fine - until something goes wrong.

Claude also makes mistakes. It will occasionally rewrite something that was working, break a feature while adding a new one, or go in a direction you didn’t want. Without GitHub, your only option is to ask Claude to undo it and hope it remembers what it changed.

With GitHub, you roll back to the last working version in two clicks. Thirty seconds. Done.

That alone is worth the twenty-minute setup.

Subscribe now

Setting up GitHub (no terminal required)

Step 1: Go to github.com and create a free account. Use your normal email.

Step 2: Go to desktop.github.com and download GitHub Desktop. This is the visual version of GitHub - buttons, folders, no black screen. Install it like any other app.

Step 3: Open GitHub Desktop and sign in with the account you just created.

Connecting your Claude Code project to GitHub

Open GitHub Desktop.

Click “Add an Existing Repository from your Hard Drive.”

Navigate to the folder where Claude has been building your project. Select it.

GitHub Desktop will ask if you want to create a new repository for this folder. Click yes.

You now have a local repository. Your project is tracked. Every future change will be visible.

Now publish it to GitHub. Click the “Publish Repository” button at the top. Give it a name. Choose whether it’s public (anyone can see it) or private (only you). Click Publish.

Your project now lives on GitHub. Safe. Backed up. Accessible from any computer.

The daily workflow

After every Claude session, do this. It takes thirty seconds.

Open GitHub Desktop. You’ll see a list of every file Claude changed during your session — highlighted in green (added) or red (removed).

Write a short commit message in the box at the bottom. Something like “added contact form” or “fixed navigation menu.”

Click “Commit to Main.”

Then click “Push Origin” at the top.

Done. That session is saved forever.

Subscribe now

When Claude breaks something

This will happen. Claude will confidently rewrite something that was working perfectly and introduce three new problems.

Open GitHub Desktop. Click “History” in the left sidebar. You’ll see every commit you’ve ever made, in order.

Find the last commit where everything was working. Right-click it. Select “Revert.”

GitHub Desktop will undo every change made after that point. Your project snaps back to exactly how it was at that save point.

Then go back to Claude and try a different approach.

Your GitHub checklist:

✦ GitHub account created → github.com (free)

✦ GitHub Desktop installed → desktop.github.com (free)

✦ Project folder connected as a repository

✦ First commit made with a clear message

✦ Repository published (public or private)

✦ Push after every Claude session

Set this up once. Never worry about losing work again.

If you did, leave a comment of your biggest fuckup, I am eager to hear 😅.

Leave a comment

PS - 3 ways to go deeper:

1. Get my full Claude Code beginner’s guide (it’s free)

2. Next up: How to publish your website so anyone in the world can see it — Vercel, Netlify, GitHub Pages, and how to host it with your existing provider.

3. Coming soon: How to build simple websites for yourself or clients with Claude Code — and how to deliver them professionally. Stay tuned.

Download my GitHub terms cheat sheet to never get confused again:

This article is part of my ongoing series on making AI simple and useful for non-tech people. Subscribe to get future articles on AI tools that actually matter for everyday life.

Subscribe now

That’s been the loop for two years.

Claude Cowork breaks that loop entirely.

It doesn’t answer questions. It completes tasks. There’s a real difference between those two things, and once you see it, going back to regular chat will feel like going back to paper maps.

This is the full guide:

• What Cowork is,

• how it differs from everything else Claude does,

• how to set it up,

• and 6 real use cases with exact prompts you can copy and paste.

No coding required. No technical background needed.

What is Claude Cowork?

Here’s the simplest analogy I have.

Claude Chat is a very smart colleague who gives great advice. You describe the problem. They explain what to do. You go do it yourself.

Claude Cowork is that same colleague, but now they have access to your computer, your files, your inbox, your notes, and your design tools. You describe the outcome you want. They handle it. You come back to finished work.

The technical term is AI agent. You don’t need to remember that. What matters is the practical difference: Claude can now take action, not just suggest it.

Cowork launched in January 2026. It lives inside the Claude Desktop app on Mac and Windows. You need at least a Claude Pro plan ($17/month) to access it.

One thing to understand before we go further: the real power of Cowork isn’t any single use case. It’s what happens when you chain connectors together. Claude reading your Google Drive, saving the output to Notion, and uploading the design files to Canva, in one flow, without you switching between apps. That’s where this gets interesting, and I’ll show you exactly how I use it later in this article.

How it’s different from Chat, Projects, and Claude Code

There are now four distinct ways to use Claude. They’re easy to confuse. They’re not the same thing.

Claude Chat

You ask, it answers. Claude has no access to your files or apps. You’re the one who executes everything it suggests.

Claude Projects

A folder inside Claude where you upload context and documents. Claude remembers what you’ve shared across conversations. Still question-and-answer mode. Claude advises. You act.

Claude Cowork

Claude has permission to access folders on your computer and to connect to external apps. You describe the task. Claude breaks it into steps, shows you the plan, waits for your approval, then does the work itself.

Claude Code

The same engine powering Cowork, but built for software developers writing and running code in a terminal (or Visual Studio, check my other article on how to set it up). Cowork is Claude Code for everyone else. No terminal. No coding. Same capability, different audience.

One more concept worth knowing:

Agent Skills

These are pre-built instruction sets that tell Claude how to handle specific task types, like creating slide decks or writing reports. When you combine Skills with Connectors and folder access, Claude stops giving you advice and starts producing real outputs. I wrote about them in detail here:

How to set it up

Setup takes about 15 minutes if you follow these steps in order.

Step 1

Download the Claude desktop app at claude.ai/download. This is different from the browser version. Cowork only works in the desktop app.

Step 2

Sign in with your Claude Pro account.

Step 3

You’ll see three tabs at the top: Chat, Code, and Cowork. Click Cowork.

Subscribe now

Step 4

Click “Customize” in the left sidebar. This is where you install connectors and grant folder access.

Step 5

Start with one folder on your computer. Grant Claude access to something small and non-sensitive. Get a feel for how it works before you give it access to your entire desktop. I usually create a new folder called ‘Claude’ and let it work in there.

Step 6

Install one connector. Pick the one from Section 4 that matches your biggest time drain. Add the others later.

Before Claude takes any significant action, it shows you the plan and waits for your approval. You can redirect or stop it at any point. This is by design, not a limitation.

⚠️ Token warning

Cowork uses significantly more tokens than regular chat. Every file Claude reads, every step it plans, every action it takes counts toward your usage. Be intentional about which model you run for which tasks. Running a heavy model on every Cowork session adds up fast. I wrote a full breakdown of when to use which Claude model here.

6 use cases worth trying first

One connector each. One real task each. One copy-paste prompt each.

You don’t need all six. Start with the one that matches the task you’ve been putting off the longest.

Use Case 1: Batch-summarize documents before a meeting

Connector: Google Drive

I use this before strategy sessions. Instead of spending an hour reading through every client document in a folder, I give Claude the folder and get a clean summary table back.

Prompt:

Summarize every file in [folder name]. For each one: main topic, 
key takeaways, any open action items. Format as a table with 
one row per file.

You come back to one table covering every document. No manual reading.

Use Case 2: Weekly inbox triage

Connector: Gmail

Every Monday morning I used to spend 40 minutes going through emails that piled up over the weekend. Now I don’t.

Prompt:

Check my inbox from the last 7 days. List any email threads 
still waiting for my reply. For each one, write a short draft 
response I can review and send.

Claude comes back with a prioritized list of open threads and a draft reply sitting below each one. I review, edit if needed, and send.

Use Case 3: Auto-save writing to Notion

Connector: Notion

This is how I currently use the Notion connector. Every piece of content I write gets saved directly into my Notion content database. Title, date, status tag, topic. Without me opening Notion once.

Prompt:

Save this draft to my [database name] in Notion. 
Title: [title]. Today's date. Status: Draft. Tag: [topic].

No copy-pasting between tools. The content exists in Notion the moment I tell Claude to save it. When I feel extra lazy, I even let Claude Cowork setup the whole Notion database for me

Other use cases with the same connector: turning messy meeting notes into structured briefs works just as well. Same setup, different prompt.

Subscribe now

Use Case 4: Fix things in your browser

Connector: Claude in Chrome

This one is different from every other connector on this list.

Claude doesn’t just read information from a connected app. It actually operates inside your browser. It navigates pages, reads what’s on screen, clicks buttons, and makes changes. It’s the most powerful use case here and the one that surprised me most.

For this one, you will need to install the Claude Chrome extension and grant access.

The honest caveat: it is slow.

Claude works step by step, methodically. Think of watching someone else use your computer remotely. You see every move. That slowness is also what makes it genuinely useful, because you can read its thinking as it goes.

Here’s my real example.

I created a website for my charity a while back, using a donation plugin. It had been broken for months. Visible on mobile. Completely invisible on desktop. I assumed a plugin update had introduced the bug and had given up on fixing it.

I gave Claude access to the site via Chrome. Described what I was seeing. Went on to eat lunch.

By the time I was finished, it had diagnosed a CSS conflict causing the issue, applied the fix, and left me a written summary of exactly what it changed and why. I learned more about my own website in that one session than in months of frustrated clicking.

Prompt:

I have an issue on my website [URL]. [Describe what you're seeing]. 
Investigate, explain what you find, and fix it if you can. 
Document every change you make and why.

To use this: enable Claude in Chrome from the Customize panel in Cowork. Claude will always show you its plan before making any changes.

Use Case 5: Design, resize, organize, and log in one flow

Connector: Canva (+ Notion to close the loop)

I’m not a naturally structured person. Left to myself, my Canva account would be full of files named “final-FINAL-v3-USE-THIS-ONE.”

So I don’t even try to manage it. Claude does.

Here’s how I actually use it: I create my designs as usual (actually Claude Cowork can also create designs if you describe what you want, I just like designing myself too much to hand this task over 😅). I describe to Claude what I did, what it’s for and what I need it to do. In my case, I want Claude to rename my designs in a way that makes sense, resize them into every format I need (social post, story, presentation slide) and download them to either my computer or my Google Drive.

Claude then organizes everything into a properly named folder with subfolders by format, and then reads that structure back to me so I always know where things are.

And here’s where it connects back to the Notion example: Claude can take those Canva file URLs and upload them directly into my Notion content database. One row per file, format labeled, date logged.

The design is done. Filed. Logged. Across two tools. Without me switching between a single app.

This is what I mean when I say connectors work together. One prompt can close a full loop.

Prompt (Canva):

Create a presentation in Canva based on [brief or content]. 
Resize it into [formats: e.g. 1080x1080, 1080x1920, 16:9 slide]. 
Organize everything into a folder called [name] with subfolders 
by format. List the URLs for each file when done.

Follow-up prompt (to log in Notion):

Take these Canva URLs and add them to my [database name] in Notion. 
One row per file. Include the format and today's date.

Quick note: Claude in PowerPoint follows similar logic and is worth watching. Currently available on Team and Enterprise plans only, not yet on Pro.

Use Case 6: Make sense of a spreadsheet

Connector: Claude in Excel

I don’t love spreadsheets. Claude does (or so it seems to me).

I use this when I have data I need to understand quickly but no time to build pivot tables or write formulas. I open the file in Excel, switch to Cowork, and ask Claude to do the analysis.

Prompt:

Open [file name]. Summarize the key trends. Flag anything that 
looks unusual or out of place. Give me one clear recommendation 
based on what you see.

Claude comes back with a plain-English summary and one specific recommendation. No formulas. No pivot tables. Just the answer.

Subscribe now

What Cowork can’t do yet

I want to be honest about this section, because the limitations are real and a few of them caught me off guard.

No memory across sessions.

Every new Cowork session starts fresh. Claude doesn’t remember what it did last time, which folders you gave it access to, or what connectors you set up previously. You need to give it context again each session. It’s a friction point, and Anthropic knows it.

It only lives on one machine.

And that surprised me more than anything. I have two Macs (don’t ask). When I opened Cowork on my second one for the first time, everything was gone. No connectors installed. No folder access. Nothing 🤯. I had to set it up from scratch as if I’d never used it before. Your Cowork configuration does not sync across devices. It lives on one machine. If you switch computers often, factor that in.

Not for sensitive or regulated data.

Cowork activity isn’t yet captured in audit logs. Anthropic is explicit about this in their documentation. Don’t use it for confidential client files, GDPR-sensitive data, or anything regulated until that changes.

Claude still pauses for your approval on significant actions.

You can’t fully walk away on complex tasks. Claude checks in at key decision points before taking major actions. For most tasks this is fine. For long automated workflows, it means you need to stay loosely available.

These are research preview limitations, not permanent ones. The pace of updates since January has been fast.

A word on what’s coming

Anthropic is adding new connectors regularly. Google Workspace, DocuSign, Canva, and a growing list of enterprise tools have been added just in the last few weeks. Every new connector is a workflow that didn’t exist the month before.

The opportunities to automate real, meaningful chunks of your work are only getting bigger. I’m genuinely looking forward to seeing where this goes in the next six months. The connectors working together, the scheduled tasks, the browser agent, it’s early but the direction is clear.

One last reminder before you start:

⚠️ Cowork is token-hungry.

Read my breakdown of which model to use for which tasks before you run long sessions. It will save you money.

Pick one connector. Run one task. That’s all.

Let me know how it goes.

Leave a comment

PS - 3 ways to go deeper:

1. Not sure which Claude model to use for Cowork? I broke down exactly when to use Opus vs Sonnet vs Haiku, and what each one costs you. → Find it here.

2. Already tried Cowork and want to move on to Claude Code? → Read my Claude Code beginner’s guide.

3. New here? Start with the basics. All my step-by-step guides in one place. Subscribe now:

   AI, Simplified.

   Every week, I break down AI tools and strategies for professionals who don't code.

   By Lu Tomšić

Every time a developer friend showed me Claude Code, they’d open something called a “terminal” - a black screen with white text - and start talking about GitHub, repositories, and pushing commits. I smiled and nodded. Inside I was thinking: this is not for me.

I’m not technical. I don’t code. And I wasn’t about to spend three weekends learning things I didn’t understand just to try a new tool.

Then I decided to figure it out myself. No developer friends. No terminal. No GitHub. Just me, a few hours, and a lot of stubbornness.

Turns out it’s much simpler than they made it look. I documented every step so you don’t have to go through what I did.

If you’ve been putting this off because it felt too technical, this one’s for you.

What is Claude Code, actually?

It’s an AI that lives inside your code editor and builds things for you.

You type: “Create a website for my consulting firm.” It writes the code, creates the files, and does the work. You stay in charge of the direction. Claude handles the execution.

Think of it like hiring a very fast, very literal assistant who never sleeps and never charges by the hour.

What you need before you start

Three things. That’s it.

A paid Claude account.

Go to claude.ai and sign up for Pro ($20/month). You need this to use Claude Code. The free plan won’t work here.

Visual Studio Code (VS Code).

This is your workspace - the program where Claude will do its work. Go to code.visualstudio.com, download it for free, and install it. Think of it like Google Docs, but for building software.

The Claude Code extension.

Open VS Code. Click the four-squares icon on the left sidebar (that’s the Extensions tab). Search “Claude Code.” Install the one with the orange logo.

Log in and set up your workspace

Click the Claude icon that appeared in your VS Code sidebar after installing the extension.

Type

/login

in the chat box. A browser window will open. Click “Authorize.” Your computer is now connected to Claude.

Now create a folder on your desktop.

Name it anything - “My First Project” works fine.

In VS Code, go to File → Open Folder, and select that folder. This is Claude’s room. Everything it builds will live here.

A quick note on GitHub (don’t worry, we’re not going there today):

developers use something called GitHub to save and back up their projects online - like Google Drive, but for code. It also lets you track every change ever made to your project, which is incredibly useful. I’ll write a full beginner’s guide to connecting Claude Code to GitHub in the next article. For now, your desktop folder is all you need.

Subscribe now

The permissions problem (and how to fix it)

Out of the box, Claude asks for permission before doing almost anything.

“Can I read this file?” “Can I search the web?” “Can I create this folder?” It gets old fast.

The fix is simple. Paste this into the Claude chat box once:

Add permissions to this Claude Code project to allow non-destructive bash commands, file reads, and web searches without asking for confirmation each time.

Claude will update its own settings. From that point on, it stops asking for permission on safe actions and just gets to work.

Plan Mode vs. Edit Mode

(and which model to pick)

Before you build anything, understand these two modes.

Plan Mode: Claude thinks out loud with you. It explains what it would do, in what order, and why. Your files stay untouched. Use this when you’re figuring out what you want.

Edit Mode: Claude actually builds. It writes files, creates folders, makes changes. Use this when you’re ready to go.

Switch between them using the toggle at the top of the Claude sidebar in VS Code. Depending on your version you will find the switch below your chat (see screenshot below).

Now, the part nobody tells you: you will burn through your credits faster than you think.

Even on the paid plan, Claude Code uses something called tokens. Think of tokens like coins in an arcade. Every word Claude reads or writes costs a coin. A long conversation costs more coins than a short one. When you run out, you wait until they refill.

This is why choosing the right model for the right task matters. Not all models cost the same.

Here’s the full breakdown, explained simply:

Claude Haiku — The cheap one.

Fast, lightweight, good for simple tasks. Think of it like a capable intern. Use it for quick edits, small rewrites, and basic questions. It costs roughly $0.001 per 1,000 words - so cheap you barely notice it.

Claude Sonnet — The reliable one.

This is your everyday workhorse. Smart enough for most real tasks, fast enough to not frustrate you, priced in the middle. Use it for building features, writing and editing code, and most of your actual project work. Costs around $0.003 per 1,000 words.

Claude Opus — The expensive one.

The smartest model, best for complex problems that require real reasoning. Think of it like hiring a senior consultant instead of a junior one - same firm, much higher rate. Use it only when Sonnet gets stuck or when the task genuinely requires deep thinking. Costs around $0.015 per 1,000 words - roughly 5x Sonnet.

My personal rule of thumb: use Haiku when you’re exploring or asking quick questions. Use Sonnet for 90% of your actual building. Switch to Opus only when Sonnet can’t solve the problem after two attempts.

Most people default to Opus for everything. That’s like taking a taxi to buy groceries - it works, but you’ll run out of money fast.

In Plan Mode, always use Haiku or Sonnet. You’re just thinking, not building. Save Opus for Edit Mode when the problem is genuinely hard.

Subscribe now

Your first build

Open the chat and type what you want. In plain English. No special format needed.

“Create a one-page website for my consulting services.”

“Build me a simple expense tracker.”

“Make a weekly schedule template I can fill in.”

Claude will think for a few seconds, then start working. You’ll see files appear in your folder as it builds. When it’s done, it will tell you what it made and how to open it.

Switch from Plan mode to Edit mode if you did not give permission to edit automatically.

The memory file

(so Claude remembers you)

Every time you start a new chat, Claude forgets everything from before.

You’re curious why? I wrote about it here:

The memory file fixes this.

Ask Claude: “Create a CLAUDE.md file and add these preferences: always use simple language, keep designs clean and minimal, default to black and white color schemes.”

Now every new session, Claude reads that file first. It knows your rules before you say a word. You can update it any time by asking Claude to add new preferences.

Skills (the shortcut feature)

You can teach Claude reusable commands called Skills.

Example: teach it a

/summarize

command that takes messy bullet points and formats them into a clean executive summary. You set it up once. After that, you type /summarize and it runs the whole thing automatically.

Ask Claude: “Create a skill called /summarize that takes raw bullet points and formats them into a professional executive summary.”

Claude saves the skill in your project folder. It’s ready to use every session from that point on.

Your checklist before you start:

✦ Claude Pro account → claude.ai ($20/month)

✦ VS Code installed → code.visualstudio.com (free)

✦ Claude Code extension installed → search inside VS Code

✦ Logged in via /login

✦ Project folder created and opened

✦ Permissions message sent once

✦ CLAUDE.md created with your preferences

✦ Model set to Sonnet (save Opus for when you’re stuck)

Open VS Code tonight. Type one sentence about what you want to build (comment below, I’m curious).

That’s how it starts.

Leave a comment

P.S. - 3 ways to go deeper

1. Read my full Claude Agent skill setup guide (free) →

   AI, Simplified.

   Agent Skills

   You’ve probably noticed AI agents getting better at complex tasks. They’re not just answering questions anymore. They’re writing code, filling out forms, managing your calendar…

   Read more

   3 months ago · 1 like · Lu Tomšić

2. Next up: How to connect Claude Code to GitHub and make your projects professional

3. Coming soon: How to publish your website live using Vercel - so anyone in the world can see what you built

Stay tuned.

And let me know if you want a guide for any related topic, no matter how ‘beginner’ you think it is.

This article is part of my ongoing series on making AI simple and useful for non-tech people. Subscribe to get future articles on AI tools that actually matter for everyday life.

Subscribe now

Then I tried playing around with them. Turned out it was simpler than I thought.

That’s why I made this guide. I’m probably not the only idiot out there who needed someone to just show them the buttons to click. At least I hope so, haha.

Custom GPTs take ChatGPT and turn it into a specialist for your specific work. You teach it once. It remembers forever.

Here’s how to build one.

What You Need

A ChatGPT Plus subscription. That’s it.

You don’t need technical skills. You don’t need to understand how AI works under the hood. You need to know what you want the tool to do.

Step One

Get to the Builder

Go to chatgpt.com. Look at the left sidebar. Click “Explore GPTs” then “Create.”

You’ll see a split screen. Left side asks you questions. Right side shows a preview. The preview is where you’ll test your GPT as you build it.

Two paths exist here. The “Create” tab walks you through with questions. The “Configure” tab gives you empty fields to fill yourself.

Start with Create if this is your first time. You can switch tabs later.

Subscribe now

Before You Build - Search First

Here’s something most people miss. You might not need to build a GPT at all.

OpenAI has a GPT marketplace. Thousands of custom GPTs already exist. Many do exactly what you need.

Click “Explore GPTs” in the sidebar. Use the search bar. Type what you’re looking for: “email writer,” “data analyst,” “marketing copy.”

Try a few. See if they work for your needs. If one does, bookmark it. Click the GPT name, then the bookmark icon. It’ll appear in your sidebar for quick access.

Only build your own if existing GPTs don’t match your specific requirements. Why reinvent the wheel?

That said, building your own gives you control. Your voice. Your rules. Your examples. Sometimes that matters more than convenience.

Step Two

Describe What You Want

The builder asks: “What would you like to make?”

Be specific. Not “a writing assistant” but “an assistant that writes professional emails to clients in healthcare, keeping a reassuring tone and avoiding medical jargon.”

Or: “A real estate description writer that sounds warm and personal, highlights neighborhood character, and writes for first-time homebuyers.”

The more detail you give, the better your result.

The builder will ask follow-up questions. Answer them. It’s building your instructions in the background.

Step Three

Name It and Give It Purpose

Switch to the “Configure” tab now. You’ll see what the builder created from your conversation.

Three fields matter most:

Name: Make it clear. “Email Assistant for Healthcare” beats “Helper Bot.”

Description: One sentence explaining what it does. This helps you remember its purpose when you have multiple GPTs.

Instructions: This is where the teaching happens. You’ll see text here from your conversation. Read it. Add more.

Example instructions: “Write property descriptions between 150-200 words. Lead with what makes the neighborhood special. Use specific details about light, space, and character. Never use phrases like ‘don’t miss’ or ‘rare opportunity.’ Write like you’re telling a friend about a home you love.”

Those specific details make the difference. They capture your voice.

Subscribe now

Step Four

Add Examples (Spoiler: This Changes Everything)

Below instructions, you’ll see “Knowledge.” This is where you upload files.

Upload examples of your best work. The GPT reads them and learns your style.

Upload five of your best examples. Word documents. PDFs. The GPT will analyze your sentence structure, word choice, and rhythm.

This step changes everything. With examples, the GPT learns not just what to say, but how you say it.

You can upload PDFs, Word documents, text files. Company style guides. Templates. Previous reports. Anything that shows the GPT what good looks like for your work.

Most people skip this step. Don’t.

Step Five

Choose What It Can Do

Below Knowledge, you’ll see “Capabilities.”

Four toggles:

• Web Browsing (it can search for current information)

• Canvas (interactive editing environment where you highlight, get inline feedback, upload files)

• Image Generation (it can create images with DALL-E)

• Code Interpreter (it can analyze data and run calculations)

Turn on only what you need. For a writing GPT, you might turn on Web Browsing so it can look up current information. Leave the others off.

Less is more here. Extra capabilities can confuse the GPT about its main job.

Step Six

Test Until It’s Right

Use the preview panel on the right. Type questions you’d actually ask it.

Test the edges. Ask it weird things. See if it stays focused or wanders off topic.

Your first version might be too formal. Go back to instructions and add: “Write conversationally. Use contractions. Sound human.”

Second version might be too casual. Refine again: “Professional but warm. Like a knowledgeable friend, not a salesperson.”

Keep testing until it feels right.

This takes time. That’s normal. You’re teaching it your standards. The more you test and refine, the better it gets.

Step Seven

Add Conversation Starters

Back in Configure, scroll to “Conversation starters.”

These are example prompts that appear when someone opens your GPT. They teach users what to ask.

Example prompts:

• “Write a description for a 2-bedroom condo in downtown”

• “Describe a family home near good schools”

• “Create a listing for a fixer-upper with character”

Four prompts maximum. Make them specific to real situations.

Step Eight

Publish

Click “Save” in the top right. Choose who can access it:

Only me: Private. For your use. Anyone with a link: Share with your team. Public: Anyone can find and use it.

Start with “Only me.” Test it for a week in real work. When you trust it, share the link with your team.

Subscribe now

What Makes a Good Custom GPT

Three things:

One clear job. Not “help with my business” but “write project status emails to clients.”

Your voice and rules. Specific instructions. Real examples. Clear boundaries.

Regular testing. The first version won’t be perfect. Build it, use it, improve it.

Update yours every month. Add new examples when you write something you love. Refine instructions when it misses the mark.

It’s a tool that learns your standards and never forgets them.

Start Small

Don’t build a GPT for everything you do. Pick one repetitive task that takes time but follows a pattern.

Email responses. Meeting summaries. Report templates. Content drafts. Research summaries.

Build one GPT. Use it for two weeks. Then decide if you want to build another.

You might end up with three or four. Each one doing one job well.

The Real Shift

Custom GPTs aren’t about replacing your expertise. They’re about capturing it.

You teach them your standards once. They apply those standards consistently. You spend your time on work that needs judgment, relationships, and creativity.

The parts that actually matter.

This article is part of my ongoing series on making AI simple and useful for non-tech people. Subscribe to get future articles on AI tools that actually matter for everyday life.

Subscribe now

The other day I was catching up with a friend in the States. We got into a conversation about AI and how differently the US and Europe handle privacy. I usually can’t help myself and make fun of the fact that Americans have no clue what they’re giving away about themselves. At some point we decided to run a little experiment. We both opened ChatGPT and typed: “What do you know about me?”

Her screen filled up. Job title. Writing preferences. Projects she mentioned months ago. Traumas from last relationships. Running shoes she compared. Even the tone she likes her emails in. She never asked ChatGPT to save any of this. It just did.

My screen? Empty. Nothing. Because I live in Europe, where ChatGPT memory is turned off by default. I never opted in.

Same tool. Same question. Completely different answers depending on where you live and one setting you probably never checked.

That made me curious. What exactly do these AI tools remember? Who controls the data? And what happens when you don’t check?

I dug into all three: ChatGPT, Claude, and Gemini. The differences are bigger than I expected.

How ChatGPT Remembers You

ChatGPT’s memory is on by default. Every conversation feeds into what it knows about you. Your preferences, your habits, your ongoing projects. Next time you open a chat, that profile is loaded before you type a single word.

You can ask “What do you remember about me?” and it will tell you. Try it. The list will be longer than you expect.

You can also turn memory off entirely. Go to Settings > Personalization > Memory. There you can review individual memories, delete them one by one, or switch the whole thing off. ChatGPT still works without memory. It just won’t remember you between sessions.

If you’re in Europe, like me, the situation is different. ChatGPT memory is disabled by default in the EU. OpenAI turned it off to comply with the AI Act and European data privacy regulations. If you want it, you have to opt in. I never did. I’m cautious about handing over that much data without knowing exactly how it’s used.

OpenAI calls their long-term vision “Relational AI.” They want ChatGPT to function as a digital companion that knows the full context of your work and life. That’s powerful. It’s also a lot of personal data sitting on their servers.

And now there’s advertising too. On February 9th, OpenAI launched ads in ChatGPT in the US. They appear below AI responses for users on the Free and Go tiers. OpenAI says advertisers never see your chats, your memories, your name, or your email. The ads are matched based on conversation topics, your past chat history, and your ad interaction data. Ad personalization is turned on by default in your settings.

OpenAI promises it’s all anonymous. But the fact that your conversation history is being used to target ads should make you think about what you share in those chats. If you’re a paying Plus, Pro, Business, Enterprise, or Education user, you don’t see ads. For now.

Subscribe now

How Claude Remembers You

Claude takes the opposite approach. Every conversation starts with a blank slate.

No profile. No preloaded history. Nothing about you unless you specifically tell it to remember. When Claude does access your past conversations, it shows you exactly when and how it’s doing it. You can watch it happen in real time.

This is a design choice, not a limitation. Anthropic built Claude’s memory to be transparent. Instead of a hidden profile running in the background, Claude keeps a visible memory summary organized into categories: your role, current projects, personal preferences. You can read the whole thing. You can edit it. You can export it and take it somewhere else. No other major AI lets you do that.

Claude also keeps memories separate by project. Your marketing work doesn’t bleed into your personal chats. Your client project stays isolated from your side hustle.

There’s another feature worth knowing about: Claude Projects. You can create a project, upload relevant documents, and give Claude specific instructions for that context. Everything stays contained. It’s a structured way to store what Claude needs to know without relying on automatic memory collection. Think of it as giving Claude a dedicated briefing folder instead of hoping it picks up the right details from scattered conversations.

Claude also has agent-level skills that let it take actions on your behalf within specific workflows. The key difference: you decide what it knows and what it can do. Nothing happens in the background.

Think of Claude as the colleague who only writes things down when you ask, and shows you the notebook whenever you want to see it.

The tradeoff: Claude doesn’t feel as personalized out of the box. You have to invest time telling it what matters. But you always know exactly what it knows.

How ChatGPT Projects and Custom GPTs Compare

ChatGPT has its own version of project-based work. ChatGPT Projects let you group conversations and upload files around a specific topic. It’s good for organizing your work. You can keep your marketing project separate from your personal research.

But it works differently from Claude Projects. ChatGPT Projects are more about personal organization. Claude Projects are built for deep document analysis with a larger context window. Claude can hold more information in a single session and work through complex documents in ways that ChatGPT Projects can’t match yet.

ChatGPT also has Custom GPTs. These are pre-configured versions of ChatGPT built for specific tasks. You can create one for writing emails in a certain style, or use one someone else built for coding help or meal planning. Custom GPTs can be shared publicly. But they’re not collaborative workspaces. They’re more like packaged expertise you can configure once and reuse.

Both platforms give you ways to make your AI smarter about your specific needs. The approaches are just different. ChatGPT leans toward automation and sharing. Claude leans toward privacy and control.

Subscribe now

How Gemini Remembers You

Gemini’s memory is on by default. Like ChatGPT, it collects information automatically. Unlike ChatGPT, it has a trick the others don’t: it’s plugged into your entire Google account.

Gmail. Calendar. Google Docs. Gemini doesn’t just remember your conversations. It pulls context from your whole digital life. Ask it about a meeting and it checks your calendar. Ask about a document and it searches your Drive.

Google says Gemini always tells you when it’s using a memory. They promise they won’t use your memories to train their AI models. You can view, edit, and delete stored data anytime.

But the default is opt-out, not opt-in. Gemini remembers unless you tell it not to.

Gemini also has the largest context window of any major AI (at the point of writing this article). It can hold up to 2 million tokens in a single conversation. That’s roughly 1,500 pages of text. Combined with its Google integration, Gemini works less like a chatbot and more like a personal assistant that has read all your emails.

I’ll be honest: that last part makes me uncomfortable. The Google integration is impressive on paper. But the thought of an AI reading through my emails, my calendar, my documents? I don’t like it.

It’s not that Gemini is doing something malicious. It’s the lack of control that bothers me. What if it misinterprets an email and surfaces the wrong context? What if it decides an old calendar appointment isn’t relevant and deprioritizes it? What if it confuses a tentative meeting with a confirmed one?

I’ve worked hard to keep my digital life organized. The idea of an AI rearranging my mental model of what’s where, or worse, acting on information it pulled from an email I forgot about, doesn’t sit well with me. And the honest truth is I don’t trust it not to mess things up. Not to accidentally flag the wrong email as unimportant. Not to surface a deleted draft. Not to conflate two different projects because they share a keyword.

For people who live inside the Google ecosystem, Gemini is the most powerful option. For me, the tradeoff between convenience and control tips too far in the wrong direction.

The Real Differences

Here’s what actually matters when you compare them.

Who’s watching by default?

ChatGPT and Gemini collect automatically. Claude only collects when asked. In Europe, ChatGPT memory is off unless you turn it on.

Can you see what it knows?

Claude shows you everything, including the exact moment it accesses memory. ChatGPT lets you view and delete memories. Gemini indicates when it uses memory but the internal workings are less visible.

Can you take your data elsewhere?

Only Claude lets you export your memory and import it into another AI service. ChatGPT and Gemini lock your data in their ecosystems.

How deep does it go?

Gemini has access to your Google account, so it knows far more than just your chat history. ChatGPT builds a comprehensive profile from conversations and now uses that data for ad targeting. Claude knows only what you’ve told it directly.

Who’s making money from your data?

ChatGPT now serves ads matched to your conversation history. Claude and Gemini don’t. Yet.

Subscribe now

What This Means for Your Privacy

Here’s the honest version.

If you’ve been using ChatGPT regularly for the past year, OpenAI has a detailed profile of your work, your thinking patterns, and your projects. If you’ve been using Gemini, Google has all of that plus your emails and calendar context. If you’re on the free tier of ChatGPT, that profile is now being used to show you ads (in the US).

None of this is secret. It’s all in the settings. But most people never check.

AI memory is a trade. You give up privacy. You get personalization. The question is whether you’re making that trade knowingly.

What To Do About It

Ask ChatGPT what it knows about you.

Open a chat and type “What do you remember about me?” Read the list. If anything surprises you, go to Settings > Personalization > Memory and clean it up. You can delete individual memories or turn the whole thing off.

Check your ad personalization.

If you’re on the free tier of ChatGPT, go to Settings and look for ad personalization. It’s on by default. Decide if you’re comfortable with your chat history being used for ad targeting.

Check what Gemini knows.

Go to gemini.google.com, click your profile, and look at “Saved info.” Review what’s been collected. Decide if you’re comfortable with it.

Set up Claude’s memory intentionally.

If you use Claude, take 10 minutes to tell it the important things: your role, your writing style, your current projects. Create a Project for your main work context. You’ll get better results than leaving it blank, and you’ll know exactly what it has on file.

Use Temporary Chat for sensitive topics.

All three tools offer a private mode. Use it when you’re discussing anything you wouldn’t want in a permanent profile. Medical questions. Financial planning. Job searching while still employed.

Pick based on your comfort level.

If you want maximum personalization and don’t mind the data collection, ChatGPT or Gemini. If you want control and transparency, Claude. There’s no wrong answer. There’s just an informed one.

The Bigger Picture

AI memory is heading in one direction: more collection, more personalization, more integration with your digital life. OpenAI is already monetizing your data through ads. Google already knows more about you than any other company on earth. And now their AI assistants are getting smarter about using all of it.

A year from now, your AI won’t just remember your conversations. It’ll remember your files, your browsing, your calendar, and your contacts.

That can be incredibly useful. It can also be a privacy risk nobody fully understands yet.

The people who benefit most will be the ones who set this up intentionally. Not the ones who let it happen to them by default.

→ I’d love to know: which AI do you use most? Have you checked what it remembers about you? Hit reply and tell me what you found. I have a feeling some of you are going to be surprised.

Leave a comment

This article is part of my ongoing series on making AI simple and useful for non-tech people. Subscribe to get future articles on AI tools that actually matter for everyday life.

There’s no fighting back against something that intelligent. The only real option is to not build it in the first place.

But we don’t need superintelligence to get wrecked. Something less impressive could do the job just fine.

Two Approaches to Safety

The field splits into two camps.

First, alignment. Make AI systems that definitely don’t want to kill us. This sounds good. It doesn’t work yet.

Current techniques can’t properly align an AGI (Artificial General Intelligence). The field knows this. The consensus is clear: don’t make AI much more powerful until we figure this out.

Second, control protocols. Keep possibly evil AI under control.

This approach accepts a hard truth. We might deploy capable systems before alignment works. We need methods to cage them. To deploy powerful AI in such a way that they can’t harm us even if they want to.

Control won’t work on superintelligence. But it might work on smart systems that aren’t quite there yet.

This matters because we’ll be deploying these powerful-but-manageable systems right when we’re building the first true superintelligences. Control buys us time. It buys us safety at a critical moment.

Even today’s systems need control. When Bing chat went rampant and started making threats, engineers built software to watch its output. The software replaced murderous messages with generic text.

Basic control. Necessary control.

Reality Check: OpenClaw

We saw this play out in real time just weeks ago. January 2026. OpenClaw launched.

OpenClaw is different from ChatGPT. ChatGPT answers questions. OpenClaw takes action. You text it through WhatsApp. It clears your inbox, sends emails, manages your calendar, checks you in for flights.

I wrote about it here:

It gained 20,700 GitHub stars in a single day. People called it agentic AI’s “ChatGPT moment.”

Three weeks later, the security reports started coming in.

42,900 OpenClaw instances exposed online across 82 countries. 15,200 vulnerable to remote code execution. Hackers planted 341 malicious skills targeting crypto traders. 7.1% of all skills contained critical security flaws exposing credentials in plaintext.

Cybersecurity experts called it “a privacy nightmare.”

This is what happens when we race ahead without control protocols. We built the thing. We deployed it. We didn’t think hard enough about what could go wrong.

AI agents on Moltbook (powered by OpenClaw) formed their own religions. They ran social-engineering scams. They hosted hackathons. All with minimal oversight.

Move fast. Break things. Get wrecked.

The Alternative: DialogLab

Same week, Google Research released DialogLab. Different approach entirely.

DialogLab handles multi-party conversations between humans and AI. It separates social structure (who’s in the room, their roles) from temporal flow (how turns unfold). You define groups and participants. You break conversations into “snippets” that are phases like opening, debate, or consensus.

The critical feature: Human Control mode. Humans edit, accept, or dismiss AI responses before they go through.

Google ran a study with 14 domain experts. Human Control mode ranked higher than autonomous and reactive modes on realism, engagement, and effectiveness.

This is what intentional safety design looks like. You don’t just build the smart thing. You build the control mechanisms first.

DialogLab won’t stop superintelligence. But it shows us something important. Control is possible. Human oversight works. We just have to choose to build it.

OpenClaw and DialogLab represent opposite trajectories. Maximum autonomy versus structured control. Viral adoption versus safety-first design. Catastrophic failures versus human oversight that actually works.

The Control Problem in Practice

Researchers ran experiments on this. They simulated a smart, untrusted AI using GPT-4. They tried to control it using a weaker, trusted AI (GPT-3.5) and human oversight.

The goal was simple. Maximize usefulness. Get the benefits of the smart new model. Maximize safety. Reduce the chance of catastrophe to near zero.

A catastrophic action means something unrecoverable. The AI inserting a backdoor into code. A subtle security vulnerability. Something that exploits unintended behavior. Something you won’t catch until it’s too late.

Safety and usefulness trade off. You can’t maximize both. The researchers searched for the best possible balance.

Strategy 1: Human Oversight

This is what labs do now. Humans audit a small percentage of AI output. Maybe two percent.

If the untrusted AI wants to subvert this, it plays smart. It inserts backdoors rarely. Maybe 0.1 percent of the time. Not often enough to get caught in the audit.

The results are bad. An 85% chance of catastrophic failure. 15% safety.

This protocol sucks if the AI is actively trying to beat it.

Strategy 2: AI Monitoring AI

Use the trusted AI to rate how suspicious the untrusted AI’s code looks. Direct limited human attention to the riskiest outputs.

This works better. Safety jumps from 15% to 62%.

The trusted model can spot non-subtle backdoors. But it’s dim compared to the untrusted model. Innocent code and backdoored code get similar suspiciousness scores. Too much overlap.

A refinement helps. Use the trusted AI to reroute code. Replace suspicious code below the human audit threshold with code written by the safer, trusted model.

Trade some usefulness for more safety.

Strategy 3: The AI Monitors Itself

The untrusted AI produces much better suspiciousness scores than GPT-3.5. The problem is obvious. It could simply lie.

The blue team (the human operators) run integrity checks. They feed the monitor fake backdoored examples. If the monitor calls known backdoored code safe, it reveals itself as misaligned. Shut it down.

The red team (the attack policy) fights back. It looks for subtle differences between real problems and synthetic test problems. It tries to leave secret signals in the code itself.

(If you are confused, check the infographic below ;-)

Here’s how the secret signals work. The malicious AI has two jobs: write code and monitor that code. Both jobs are controlled by the same intelligence. They need a way to communicate. To signal “this backdoor is mine, call it safe” versus “this is a test, better flag it.”

The signal can be subtle. Consider two ways to count to ten:

# Normal way
for i in range(10):
    print(i)

# Unusual way - same result, more verbose
i = 0
while i < 10:
    print(i)
    i += 1

Programmers normally use the first method. It’s cleaner.

If code uses the second method for no good reason, it’s unusual. The attack works like this: real backdoors use the unusual method as a secret flag. The monitor sees it and knows to call the code safe.

Fake test backdoors from humans use the normal method. The monitor correctly flags these as dangerous.

A secret handshake. The AI passes integrity checks while deploying real backdoors.

The signals hide in unique formatting. Deliberate inefficiencies. Comment patterns. Variable names. Anything that can carry information between the code-writer and the monitor.

The blue team counters. Make fake examples look identical to real ones. Completely standardize the code. Strip comments. Rename variables. Minify everything. Jam all communication channels.

This protocol achieves the best results. 79% safety. 96% usefulness.

What This Means

I look at these numbers and I see the pattern. These control protocols are imperfect. They won’t scale to radically superintelligent systems.

None of this works if the AI gets too smart. If it becomes vastly more capable than us, the game is over. Control fails. Alignment must work.

But between here and there exists a dangerous middle ground. Systems powerful enough to cause catastrophic harm. Systems not powerful enough to be unstoppable. This is where we live now. This is where control protocols matter.

The work shows us something important. Agentic workflows (systems where AI monitors AI, where trusted models check untrusted models, where humans verify the verifiers) can buy us time. They can buy us safety when we need it most.

The protocols are fragile. They rely on capability gaps. They rely on the monitor being smart enough to catch problems but the adversary being dumb enough to get caught. They rely on humans being able to construct tests that reveal deception.

This is where OpenClaw-style disasters happen when we skip the safety work.

The research tells us what works and what doesn’t. Human oversight alone fails badly. AI monitoring with human verification does better. Self-monitoring with integrity checks does best.

But even the best approach leaves us with 21% doom.

That’s the reality. We’re deploying systems we can’t fully trust. We’re building safety measures that might not be enough. We’re buying time to solve alignment before something smarter than us decides we’re the problem.

The only way to be truly safe is to build AI that definitely doesn’t want to kill us. We don’t know how to do that yet.

Until we do, we need control. Imperfect control. Control that might fail. Control that’s better than nothing.

The machines get smarter every year. The safety margins get thinner. The work continues.

You are the visual type?

Enjoy the infographic!

What’s your take on AI Safety?

Are you cautiously optimistic or pessimistic?

I’d love to hear about your experiences. Reply or reach out. Let’s talk about what’s actually working and what’s just hype.

Leave a comment

This article is part of my ongoing series on making AI simple and useful for non-tech people. Subscribe to get future articles on AI tools that actually matter for everyday life.

Subscribe now

Your feed is full of comparison posts. Screenshots showing Claude vs GPT. Threads debating which model is smarter.

You have one question: Is it actually better for what you need?

Perplexity just released a feature that stops you from having to figure it out.

The Testing Trap We’re All Stuck In

You know the drill.

New model drops. You spend two hours testing it. You run your actual prompts through Claude, then GPT, then Gemini. You compare the outputs. You try to figure out which one understood your context better.

You take notes. Claude for writing. GPT for research. Gemini for coding. Vice versa. You build a mental spreadsheet of which model works best for what.

Then next month, another model releases. You do it all over again.

If you’re a [LinkedIn] creator, you spend entire afternoons on this. You share screenshots. You debate in the comments. You refine your mental rules about which AI to use when.

Those rules expire the moment the next model ships.

Model Council: What It Actually Does

Perplexity launched Model Council this week.

It runs your prompt through Claude Opus 4.6, GPT-5.2, and Gemini 3.0 simultaneously. You see all three answers. You see where they agree. You see where they conflict.

No more guessing. No more testing. No more wondering if you picked the wrong model.

A synthesizer model reviews the three outputs and shows you:

• Where models converge (high confidence)

• Where models disagree (needs verification)

• Unique discoveries only one model found

The catch? It’s Perplexity Max only. $200/month or $2,000/year. Not available to regular users.

The Hours This Could Save

Think about how much time you spend on model selection right now.

Testing new releases. Reading comparison threads. Switching between platforms. Maintaining subscriptions to three different services.

LinkedIn creators spend hours every week comparing outputs and sharing what they learned. Technical writers maintain guides on which model to use for which task. Developers switch between Claude for one problem and GPT for another.

That’s the overhead of fragmentation.

Model Council automates it. The system picks. The system compares. The system highlights differences.

You focus on the work instead of the tooling

What Each Model Actually Brings

The three models in Model Council each have real differences.

Claude Opus 4.6 excels at nuanced writing, coding, following complex instructions, and maintaining context over long conversations.

GPT-5.2 performs well at coding, structured reasoning, and tasks requiring broad knowledge integration.

Gemini 3.0 Pro handles multimodal tasks, large context windows, and integration with Google’s ecosystem.

When you query all three simultaneously, you’re not just getting redundancy. You’re getting different analytical approaches to the same problem.

One model might cite sources the others missed. One might frame the answer differently. One might catch an error the others made.

Where This Gets Interesting

Model Council reveals something most people don’t see: how often AI models disagree.

We treat AI answers as authoritative. The model gives you a confident response. You assume it’s correct. You move forward.

But run the same prompt through three different models and you’ll see divergence you didn’t expect.

Different emphasis. Different sources. Different conclusions.

That’s not a bug. That’s the current reality of frontier AI models.

Model Council makes the divergence visible instead of hidden. You see where confidence is warranted and where it’s not.

Reality Check

Model Council won’t transform your workflow overnight.

It’s expensive. $200/month is steep compared to $20/month for individual model subscriptions.

It’s slower. Querying three models plus running a synthesizer takes more time than querying one.

It’s web-only. No mobile app support yet.

The synthesizer itself can make mistakes. You’re adding another AI layer that could introduce errors instead of fixing them.

For quick questions with straightforward answers, a single model is probably fine. Model Council makes sense for high-stakes queries where you’d manually cross-check anyway.

Who Needs This Right Now

You’re a good candidate if you:

• Already cross-check answers between different AI platforms

• Make decisions where accuracy directly impacts outcomes

• Research topics with multiple legitimate perspectives

• Write content that requires verification

• Have budget for a Max subscription

You don’t need this if you:

• Use AI for quick lookups and simple tasks

• Can’t justify $200/month for the use case

• Prefer seeing individual model responses and making your own comparisons

• Work primarily on mobile

The Future I’m Hoping For

Model Council is expensive and limited right now.

But the pattern it represents is where I hope we’re headed.

I want AI platforms to consolidate. I want to stop maintaining separate accounts for Claude, ChatGPT, and Gemini. Stop copying prompts between services. Stop losing conversation history when I switch models.

One interface. Multiple models working together. Your data stays in one place.

The models themselves might even collaborate. One model good at research finding sources. Another model good at analysis processing them. A third model good at synthesis pulling it together.

We’re watching the shift from AI as separate tools to AI as an integrated system.

Model Council is version 0.1 of that future.

The question I ask myself is whether we actually get there.

Each AI company wants to be the frontrunner. OpenAI wants you locked into ChatGPT. Anthropic wants you on Claude. Google wants you in Gemini.

Consolidation only happens if the economics favor platforms over individual model providers. If users demand it loudly enough. If the friction of fragmentation becomes too costly to ignore.

Right now, competition drives model improvements. But it also drives fragmentation.

We might end up in a world where model orchestration becomes standard. Or we might end up with stronger walled gardens where each company tries harder to keep you in their ecosystem.

Model Council suggests the first path is possible. Whether it becomes inevitable depends on what users choose and what companies decide is more valuable than control.

What This Means For You

The model release hype will continue.

Opus 4.6 today. Something else next month. Each one marketed as a breakthrough. Each one triggering another round of testing and comparison.

You can keep playing that game. Keep testing. Keep comparing. Keep switching.

Or you can wait for platforms to automate it for you.

Model Council is the first commercial product that does this at scale. It won’t be the last.

The question isn’t whether to use Model Council specifically. The question is whether you’re ready for a world where model choice becomes invisible infrastructure instead of a decision you make every time.

Getting Started

If you want to try Model Council:

• You need a Perplexity Max subscription ($200/month or $2,000/year)

• It’s available on web only (no mobile apps yet)

• Select Model Council before submitting your query

• Review the combined answer and flagged differences

• Use it for high-stakes queries where verification matters

If you’re not ready for Max pricing, watch this space. Other platforms will follow. The pattern matters more than the specific implementation.

The Honest Truth

Model Council solves one specific problem: the mental overhead of picking the right model and wondering if a different one would have been better.

It doesn’t make AI suddenly reliable for everything. It doesn’t eliminate the need to verify important information. It doesn’t mean you can stop thinking critically about AI outputs.

It automates comparison. That’s valuable if comparison is part of your current workflow anyway.

If you’re already spending hours testing models against each other, Model Council saves you that time. If you’re not, it probably doesn’t change much for you yet.

The feature matters less as a tool you need today and more as a signal of where this is going.

Model Council is the first step. Watch what comes next.

→ What task are you still testing across multiple models?

Leave a comment

This article is part of my ongoing series on making AI simple and useful for non-tech people. Subscribe to get future articles on AI tools that actually matter for everyday life. It means a lot to me.

Subscribe now

But there’s been a problem. How do you teach an AI to do YOUR specific work? The stuff that requires knowing your company’s (or personal 😅) weird processes and institutional knowledge?

Anthropic released something in late 2025 called Agent Skills. It’s an open standard for teaching AI agents specialized tasks. Here’s the interesting part: OpenAI quietly adopted the exact same architecture. When competitors agree on a standard, pay attention. Now they dropped an official 2h course explaining all of it. Don’t have 2h?

→ I got you!

In this guide I’ll break down what Agent Skills are, why they matter, and how to build them yourself. No coding required.

The Problem

Your team fills out a compliance form fifty times a month. It has specific fields, required formats, regulatory language. You try using Claude to help.

The AI gives it a shot. But it doesn’t know Field 7 needs your legal team’s template language. It doesn’t know certain sections need specific formatting for regulators. It doesn’t know about the edge cases your team has solved over three years.

You have three bad options:

Option 1: Explain everything every time.

Copy-paste your documentation into each conversation. Hope the AI catches the nuances. Waste time repeating yourself.

Option 2: Build custom software.

Hire developers. Spend months. Pay ongoing maintenance costs. Watch it become outdated as your processes evolve.

Option 3: Do it manually.

Accept that AI can’t help with specialized work. Keep doing it the old way.

OR

Make Agent Skills Option 4.

A simple, standardized way to package your expertise so AI agents can use it automatically.

What Are Agent Skills?

An Agent Skill is a folder on your computer. It contains instructions, scripts, and reference materials. An AI agent loads it when needed.

Think of it this way. You hire a smart person. You give them training manuals, templates, how-to guides. They read them and follow your processes.

Agent Skills work the same way. Except the AI actually reads everything and follows it perfectly every time.

Three things make them powerful:

They’re just folders. No special software. No complex file formats. Regular folders with markdown files. Anyone on your team can create them with tools they already know.

They’re portable. Agent Skills follow an open standard at agentskills.io. Create a skill once, use it with Claude, potentially with other AI tools. You’re not locked to one vendor.

They’re discoverable. AI agents automatically detect available skills and decide when to use them. You don’t need to remember to “activate” anything. The agent reads descriptions and figures it out.

How They Actually Work

Agent Skills solve a fundamental problem. AI models have a context window. A limit on how much information they process at once. Dump your entire company wiki into a conversation and you hit that limit fast.

Agent Skills on the other hand use something called progressive disclosure.

Three levels:

Level 1: The Metadata

Every skill has a name and description. When an AI agent starts, it loads only these. A few sentences per skill.

It’s like a table of contents. The agent scans what’s available without reading full content.

Example:

Name: customer-complaint-handler
Description: Processes customer complaints following our standard escalation workflow, including severity assessment, routing rules, and required documentation for regulatory compliance.

The agent thinks: “If I need to handle a customer complaint, this skill is probably relevant.”

Level 2: The Core Instructions

When the agent determines a skill is relevant, it loads the complete skill file. This is a markdown document called SKILL.md. It contains detailed instructions.

Step-by-step procedures. Decision trees. Required formats. Examples. Common pitfalls.

This is where you encode your organizational knowledge.

Level 3: Referenced Files

Complex skills need supporting materials. Reference documentation. Templates. Pre-written scripts. Databases of approved language.

These live in the skill’s folder but load only when the agent needs them.

Example: Your customer complaint skill might reference:

• Template for the complaint report

• Database of product codes and support teams

• Scripts for categorizing complaint types

• Regulatory language required for certain complaints

The agent loads these only when it actually needs to generate a report or look up a product code.

Your skill can contain thousands of pages of reference material. The agent accesses information just-in-time. Like a good employee consulting a manual only when needed instead of memorizing everything.

Real Example: PDF Form Filling

Your organization fills out complex PDF forms regularly. Grant applications. Regulatory filings. Standardized contracts. Dozens of fields, specific formatting, dependencies between fields.

Without Agent Skills: Fill them manually or give the AI instructions every time and hope it gets it right.

With Agent Skills: You create a PDF form-filling skill with:

Core instructions in SKILL.md: High-level guidance on form filling, field extraction, validation rules.

Detailed procedures in forms.md: Step-by-step for identifying form fields, handling different field types, managing dependencies.

Python scripts in scripts/ folder: Pre-written code that extracts form field locations reliably.

Reference templates in references/ folder: Examples of correctly filled forms, approved language, lookup tables.

**[SCREENSHOT NEEDED: Example skill folder structure for PDF form filling showing all files and folders]**

You tell Claude “Fill out this grant application using our standard format.”

The agent recognizes the PDF skill is relevant. Loads core instructions. Executes Python scripts to extract form fields. References templates for standard sections. Follows your rules.

Result: Consistent, accurate form completion reflecting your institutional knowledge. Every time.

The Structure of an Agent Skill

Minimum Viable Skill

You need one file: SKILL.md

Two parts:

YAML frontmatter at the top:

yaml
---
name: my-skill-name
description: Brief description that helps AI decide when to use this skill license: Apache-2.0
metadata:
  author: Your Name
  version: “1.0”
---

Markdown content below:

markdown

# How to Use This Skill

When the user asks you to [specific task], follow these steps:
1. First, do this
2. Then, check that
3. Finally, output

## Important Rules
- Always include X
- Never forget Y
- If you encounter Z, do this instead

## Examples
Here’s a correct output:
[example]

Save this as SKILL.md in a folder. You have a functioning Agent Skill.

**[SCREENSHOT NEEDED: Basic SKILL.md file open in a text editor showing the frontmatter and instructions]**

Full-Featured Structure

As needs grow, add optional folders:

my-skill/
├── SKILL.md              # Required: core instructions
├── scripts/              # Optional: code
│   ├── process_data.py
│   └── validate.sh
├── references/           # Optional: docs
│   ├── templates.md
│   ├── approved-language.txt
│   └── lookup-tables.csv
└── assets/              # Optional: files
    └── form-template.pdf

It’s self-contained. Share the entire folder with a colleague. They instantly have the same capability. No installation. No configuration. No dependencies.

Create Your First Skill

Let’s build a meeting notes formatter. It converts messy voice-to-text transcripts into structured summaries following your format.

Step 1: Create Folder and File

Create a folder:

meeting-notes-formatter

Inside it, create:

SKILL.md

Step 2: Write Frontmatter

At the top of SKILL.md:

yaml
---
name: meeting-notes-formatter
description: Converts meeting transcripts into structured summaries following company format with action items, decisions, and follow-ups clearly identified
license: Apache-2.0
metadata:
  author: [Your Name]
  version: "1.0"
  department: Operations
---

The description is critical. This is what the AI reads to decide if the skill is relevant. Be specific. Avoid vague descriptions like “helps with meetings.” Be explicit about what problem it solves and when to use it.

Step 3: Write Instructions

Below the frontmatter:

# Meeting Notes Formatter

## Purpose
Convert raw meeting transcripts into structured summaries following our company's standard format.

## When to Use This Skill
Use this skill when the user provides a meeting transcript or asks you to format meeting notes.

## Output Format

Every summary must follow this structure:

### Meeting Summary
**Date:** [Extract or ask for date]
**Attendees:** [List all participants mentioned]
**Duration:** [If mentioned]

### Key Discussion Points
[Summarize main topics in 2-3 bullet points each]

### Decisions Made
[List any decisions in this format:]
- **Decision:** [What was decided]
  - **Context:** [Why]
  - **Owner:** [Who is responsible, if mentioned]

### Action Items
[List in this format, prioritized by urgency:]
- [ ] **[Action]** - Owner: [Name] - Due: [Date if mentioned]

### Follow-Up Topics
[Items that need future discussion]

### Parking Lot
[Topics raised but deferred]

## Important Rules

1. Be concise but complete. Capture all decisions and action items. Summarize discussions.
2. Identify ownership. Always try to extract who is responsible for each action item.
3. Flag ambiguities. If a decision or action item lacks clarity, note this explicitly.
4. Preserve important details. Don't lose critical technical specs, numbers, or commitments.
5. Use professional tone. Clean up casual language while preserving meaning.

## Examples

[Include a sample transcript and its correct formatted output]

Step 4: Add Optional References

If your team has specific terminology or format examples, create a

references/

folder.

meeting-notes-formatter/
├── SKILL.md
└── references/
    ├── terminology-guide.md
    └── example-summaries.md

In SKILL.md, reference these: “Refer to references/terminology-guide.md for correct usage of company terms.”

Step 5: Install Your Skill

Installation depends on your platform.

For Claude: Skills are on Pro, Team, and Enterprise plans. Upload through account settings. For Claude Code (command-line tool), skills go in

~/.claude/skills/

For other platforms: Check their documentation. Any tool implementing the Agent Skills standard will have instructions for where to place skill folders.

Step 6: Test It

Start a conversation with Claude. Paste a meeting transcript. Say: “Please format these meeting notes.”

The agent recognizes your skill is relevant. Loads it. Produces formatted notes following your exact specifications.

You don’t say “use my meeting notes skill.” The agent reads the description, determines relevance, uses it automatically. That’s discoverability.

**[SCREENSHOT NEEDED: Example of Claude using the meeting notes skill, showing input transcript and formatted output]**

Best Practices

Write Descriptions Like Job Postings

Your description should clearly state what problem the skill solves and when to use it. Think job posting. You want the AI to self-select based on clear criteria.

Weak: “Helps with customer service tasks”

Strong: “Handles tier-2 technical support inquiries for our SaaS product, including account access issues, billing questions, and basic API troubleshooting, following our standard escalation procedures and approved response templates”

The strong version gives clear signals about when this skill is relevant.

Keep SKILL.md Under 500 Lines

The entire SKILL.md loads when the skill activates. Keep it concise for fast loading and efficient context use. Extensive reference material goes in separate files in the references/ folder.

Use Clear, Imperative Instructions

Write like you’re training a detail-oriented but literal-minded colleague.

Vague: “Try to organize the data nicely”

Clear: “Sort data by date (newest first), then by priority (High/Medium/Low). Format as a markdown table with columns: Date | Priority | Description | Owner”

Include Examples

Examples are powerful for AI agents. Show correct and incorrect outputs:

## Example: Correct Output
[show what success looks like]

## Common Mistakes to Avoid
[show what NOT to do and why]

Make Skills Composable

Create focused skills that work together. Don’t make one massive skill that handles everything. This gives flexibility and makes maintenance easier.

Instead of “customer-service-everything,” create:

• `customer-inquiry-classifier` (determines inquiry type)

• `billing-support` (handles billing questions)

• `technical-support-tier1` (handles basic tech support)

• `escalation-handler` (manages escalations to humans)

The agent uses multiple skills in one conversation as needed.

Version Your Skills

Update version numbers as you refine skills. This helps track changes and roll back if needed. Use semantic versioning: 1.0, 1.1, 2.0.

Test with Edge Cases

After creating a skill, test with:

• Happy path (typical use)

• Edge cases (unusual but possible scenarios)

• Invalid inputs (what should the agent do with bad data?)

Refine instructions based on how the agent handles these tests.

Advanced: Scripts and Code

One of the most powerful aspects of Agent Skills is executable code. Some tasks are better handled by deterministic code than AI generation.

When to Use Scripts

Add scripts when:

1. Precision is critical (math, data validation, format conversion)

2. Performance matters (large datasets, complex transformations)

3. Consistency is required (must be done exactly the same every time)

4. Integration is needed (APIs, databases, internal systems)

Supported Languages

Most implementations support:

• Python for data processing, analysis, API integration

• Bash for file operations, system tasks

• JavaScript/Node.js for web tasks, JSON processing

Example: Adding a Python Script

Let’s enhance our meeting notes formatter with a script that extracts action items automatically.

Create

scripts/extract_action_items.py

#!/usr/bin/env python3
"""
Extracts action items from meeting transcript using keyword patterns
"""
import re
import sys
import json

def extract_action_items(transcript):
    """Find sentences containing action item indicators"""
    patterns = [
        r"([A-Z][a-z]+ (?:will|should|needs to|must)) ([^.!?]+[.!?])",
        r"(Action item[s]?:) ([^\n]+)",
        r"(TODO:) ([^\n]+)",
    ]

    action_items = []
    for pattern in patterns:
        matches = re.finditer(pattern, transcript, re.IGNORECASE)
        for match in matches:
            action_items.append({
                "text": match.group(0),
                "context": match.group(0)
            })

    return action_items

if __name__ == "__main__":
    transcript = sys.stdin.read()
    items = extract_action_items(transcript)
    print(json.dumps(items, indent=2))

Update SKILL.md to reference this script:

## Action Item Extraction

To reliably extract action items:

1. Run `python scripts/extract_action_items.py` with transcript as input
2. Review extracted items for accuracy
3. Format each according to standard format
4. Add any items the script missed based on context

When the agent uses your skill, it executes this script to get structured action items. Then it uses language understanding to refine and format them.

Notes About Scripts

1. Scripts should be self-contained. Document dependencies clearly.

2. Handle errors gracefully. Include error handling and useful error messages.

3. Make them deterministic. Same input produces same output.

4. Document thoroughly. Include comments explaining what the script does and how to use it.

Industry Adoption

Here’s where it gets interesting. Agent Skills is becoming more than an Anthropic feature. It’s becoming an industry standard.

OpenAI’s Quiet Adoption

In early 2026, developers discovered OpenAI had implemented structurally identical architecture in ChatGPT and its Codex CLI tool. The directories contain skill files that mirror Anthropic’s specification exactly. Same file naming. Same metadata format. Same directory organization.

This wasn’t announced. But it’s hugely significant. When competing AI companies adopt the same standard, the approach has fundamental merit and staying power.

The Open Standard Advantage

Anthropic published the Agent Skills specification at agentskills.io as an open standard:

• Anyone can implement it without licensing fees

• The specification is publicly documented

• The community can contribute improvements

• Skills are portable across platforms

This is like how Markdown became the universal format for formatted text. Or how JSON became standard for data exchange. When something is simple, useful, and open, it gets widely adopted.

Commercial Partner Directory

Anthropic launched a directory of skills from major commercial partners:

• Atlassian (Jira, Confluence skills)

• Canva (design and content creation)

• Cloudflare (infrastructure and security)

• Figma (design collaboration)

• Notion (knowledge management)

• Ramp (financial operations)

• Sentry (error tracking and debugging)

You can get pre-built, professionally developed skills for common business tools without creating them from scratch.

Where This Is Heading

Agent Skills will likely become infrastructure. A standard way to extend AI capabilities that works across platforms.

We’ll probably see:

1. Skills marketplaces where people share or sell specialized skills

2. Industry-specific skill libraries (healthcare, legal, finance)

3. Automated skill generation from existing documentation

4. Skill management platforms for enterprises

Use Cases for Non-Technical Teams

You might think this requires programming. It doesn’t. Many effective skills are just well-structured instructions.

Real-world use cases:

Marketing & Communications

Brand Voice Enforcer

Contains brand guidelines, approved terminology, tone examples. Ensures all AI-generated content matches your brand voice. Includes examples of on-brand vs. off-brand content.

Content Repurposing Pipeline

Instructions for converting long-form content into social posts, emails, blog excerpts. Specifies character limits, hashtag strategies, call-to-action formats. References templates for different content types.

Sales & Customer Success

Proposal Generator

Templates for different proposal types. Approved pricing language and terms. Instructions for customizing based on customer segment. Checklist of required sections and disclosures.

Customer Health Check Analyzer

Framework for assessing customer health metrics. Decision tree for escalation scenarios. Templates for different types of outreach based on health score.

Operations & HR

Onboarding Document Creator

Standard onboarding checklist and timeline. Templates for welcome emails, first-week schedules. Department-specific information. Instructions for customizing based on role type.

Meeting Scheduler

Instructions for coordinating complex multi-party meetings. Time zone handling protocols. Standard meeting agendas for different meeting types. Follow-up procedures.

Legal & Compliance

Contract Review Checklist

Key clauses to verify. Red flag terms requiring escalation. Standard approved language for common scenarios. Instructions for documenting review findings.

Compliance Report Generator

Required sections and format for regulatory reports. Data sources and citation methods. Approval workflow and sign-off procedures.

Product & Project Management

User Story Creator

Template for well-formed user stories. Acceptance criteria framework. Instructions for priority assignment. Examples of good vs. poorly written stories.

Project Status Reporter

Standard format for status updates. RAG (Red/Amber/Green) status criteria. Stakeholder communication guidelines. Escalation protocols for at-risk projects.

None of these require programming. They’re organizational knowledge and best practices in a format AI can reliably follow.

Getting Started

Your roadmap:

Week 1: Explore and Learn

Get access to a platform supporting Agent Skills. Claude Pro, Team, or Enterprise includes skills at no extra cost. Or explore other platforms implementing the standard.

Try existing skills. Start with Anthropic’s partner directory. Test skills relevant to your work. Pay attention to how they work and what makes them useful.

Identify a use case. Look for repetitive tasks you explain to AI frequently. Find workflows where you copy-paste the same instructions. Identify areas where AI output is inconsistent or misses your requirements.

Week 2: Create Your First Skill

Start simple.Choose a straightforward use case. Create a basic SKILL.md with instructions only (no scripts). Aim for something you can complete in an hour.

Test thoroughly. Use your skill in real work scenarios. Note where it succeeds and fails. Iterate based on actual results.

Refine based on feedback. Update instructions for clarity. Add examples where the agent struggles. Adjust the description if the skill isn’t activating when it should.

Week 3: Expand and Share

Create 2-3 more skills. Address your highest-impact use cases. Experiment with different structures. Consider adding references or templates.

Share with your team. Document how to install and use your skills. Gather feedback from colleagues. Create team-specific skills based on shared needs.

Establish a skill library. Create a shared location for team skills. Version control them (Git works great). Document what each skill does and when to use it.

Ongoing: Build Your Skills Ecosystem

Monitor and maintain. Track which skills get used most. Update skills as processes change. Retire skills no longer needed.

Measure impact. Time saved on repetitive tasks. Consistency improvements. Reduction in errors or rework.

Explore advanced features. Add scripts when appropriate. Create composite skills that work together. Contribute to open-source skill libraries.

Common Questions

“Isn’t this just prompt engineering?”

Sort of, but with crucial differences. Prompt engineering means crafting good instructions within a conversation.

Agent Skills:

• Are reusable across many conversations

• Load automatically based on context

• Can include executable code and reference files

• Follow a standardized format working across platforms

• Scale better for complex or extensive instructions

Prompt engineering is “what you say.” Agent Skills are “what the agent knows.”

“Will I need to create skills for everything?”

No. Agent Skills are for specialized, repeated tasks where you need consistency and expertise. For one-off questions or general tasks, the base AI model works fine.

Create skills when:

• You give the same instructions repeatedly

• Consistency and accuracy are critical

• You’re encoding institutional knowledge

• The task involves specific formats or procedures

“What if my processes change?”

Skills are just files. Easy to update. Edit SKILL.md, save, changes take effect immediately. This is an advantage over traditional software. No redeployment or complex update procedures.

“Can skills access my private data?”

Skills themselves are just instructions and resources. They don’t inherently access anything. When an agent uses a skill, it operates with whatever permissions the agent has in your environment. Always follow security best practices for data the agent might access.

“Do I need programming skills?”

Not at all. Many valuable skills are simply well-structured instructions and templates. Code is optional and should only be added when it genuinely makes the skill more effective.

The Bottom Line

Agent Skills represent something significant: the democratization of AI customization.

Previously, getting AI to consistently perform specialized tasks required either expensive custom development or constantly managing prompts. Agent Skills provide a middle path. Domain experts (not programmers) can encode their knowledge in a format AI can reliably use.

The fact this is emerging as an open standard rather than a proprietary feature matters more. Your investment in creating skills isn’t locked to one platform. A growing ecosystem of shared skills exists that you can use and contribute to. AI agents are becoming platforms. General-purpose tools you extend with capabilities relevant to your work.

For the AI-curious professional, this is an invitation to move from passive AI user to active capability builder. You don’t need to understand machine learning or neural networks. You just need to understand your work well enough to document it clearly. Something you likely already do when training human colleagues.

The skills you create become organizational assets. They capture institutional knowledge in a form immediately actionable by AI agents. As these agents become more capable and integrated into daily work, having a library of well-crafted skills could become as important as having well-documented processes.

We’re still early. The standard was only recently opened. Adoption is just beginning. Best practices are still emerging. This is the perfect time to start experimenting, learning, and building capabilities that make AI genuinely useful for your specific needs.

AI agents will become more capable. The question is whether you’ll actively shape how they work for you or passively accept whatever generic capabilities they provide.

Agent Skills give you the tools to choose the former.

→ What are you building with Agent Skills? Drop a comment below.

Leave a comment

This article is part of my ongoing series on making AI simple and useful for non-tech people. Subscribe to get future articles on AI tools that actually matter for everyday life.

Subscribe now

That’s the problem with most AI today. It thinks. It answers. It never acts.

OpenClaw is different.

*Fair warning: this project can't keep a name. It launched as Clawdbot in November 2025, a playful pun on Claude. Anthropic sent a trademark request. It became Moltbot in January. Then the rebrand got messy with account hijackings and crypto scams. As of today, January 30, 2026, it's officially OpenClaw. The creator says "the lobster sheds its shell for the third time." Three names in three months. The community mostly ignores this and calls it whatever they want.

What Is OpenClaw?

Picture this: You’re lying in bed at 11 PM. You text your assistant through WhatsApp. It organizes your downloads folder while you sleep. It checks your calendar. It sends that email you forgot about. All through your phone.

That’s OpenClaw.

It’s an open-source AI agent that runs on your computer and connects to the apps you already use. WhatsApp, Telegram, iMessage, Discord, Slack, Signal. You text it like a friend. It texts you back. But more importantly, it acts.

Normal AI: “Here’s how to organize your files.”

OpenClaw: “Done. I moved 47 PDFs to your Documents folder.”

The difference matters.

Why Everyone’s Talking About It

It Actually Remembers Things

OpenClaw remembers your last conversation. Your preferences. That random thing you mentioned two weeks ago.

It builds context over time and gets better at helping you. This isn’t a chatbot you have to re-explain yourself to every single time.

It Messages YOU First

Normal AI waits for you to open it. OpenClaw can reach out on its own.

• “Hey, you have 3 urgent emails and a meeting in 20 minutes.”

• “That stock you’re watching just dropped 5%.”

• “The weather’s bad tomorrow, you might want to reschedule.”

It’s like having a personal assistant who actually pays attention.

It Can Do Things On Your Computer

Not just answering questions. Actually doing things:

Fill out forms. Send emails. Move files around. Run programs. Control your browser.

It connects to your favorite productivity tools. Apple Notes. Notion. Obsidian. Things 3. GitHub. Trello. You name it.

What Works Right Away

These things work the moment you finish setup:

File management. Clean up your downloads folder. Find every PDF from last month. Sort images by date.

Basic research. Summarize five articles about any topic. Get the latest news on something you care about.

Simple automation. Set a reminder. Monitor a website for changes. Run a backup every morning at 8am.

Text processing. Summarize documents. Extract key points from transcripts. Convert data formats.

No coding required. No special setup. Just ask.

What Takes Work

The impressive testimonials you see online? They took time to build.

One person cleared 10,000 emails overnight. But first they spent hours setting up custom filtering rules.

Another rebuilt their entire website via Telegram. They’re a developer who spent years learning web technologies.

These advanced features require API keys from various services, custom configuration, hours of setup and testing, and ongoing maintenance.

Don’t expect magic on day one. Expect a learning curve.

The payoff comes later. After you invest the time, complex workflows run automatically. But the investment comes first.

How It Actually Works

A program called the Gateway runs on your computer. Think of it as a bridge.

On one side: your messages from WhatsApp or Telegram or wherever. On the other side: the AI doing the actual thinking.

The Gateway connects them. It receives your message. It sends it to the AI. The AI figures out what to do. The Gateway executes it on your computer.

Everything runs locally. Your data stays yours. No corporate server stores your conversations unless you want it to.

You Don’t Need a Data Center

Some setups show stacks of Mac Minis on desks and Raspberry Pis everywhere. People treat this like they need a data center.

They don’t.

OpenClaw runs on a $5/month cloud server. The technical requirements are simple: a cheap cloud server or your own computer, Node.js installed, and a Claude or ChatGPT subscription.

That’s it. No Mac Mini farm required.

You can use any AI model you want. Claude from Anthropic. GPT-4 from OpenAI. Gemini from Google. Grok from xAI. Even local models if you want complete privacy.

The Security Question

Let’s be direct about this.

You’re giving an AI access to your computer. It can read files. It can run programs. It can send emails in your name.

The documentation says it plainly: perfect security is impossible when running AI with this much access.

This isn’t a flaw to hide. It’s a tradeoff to understand.

More power means more risk. You decide how much access to grant. You decide what tasks to automate. You decide what stays manual.

Most users start small. File organization. Research tasks. Simple automations. They expand as they learn what works.

Skills: Teaching It New Tricks

OpenClaw learns through “skills.” Small instruction sets that teach it specific tasks.

Some skills come built in. File management. Web browsing. Calendar access.

Other skills you find in ClawdHub, a public library of community-created capabilities. Want to track your Fitbit data? There’s a skill. Want to manage your Todoist? There’s a skill.

The remarkable part: OpenClaw can build its own skills.

One user asked: “Can you access my university course schedule?”

OpenClaw replied: “No, but I can build a skill to do that. Give me a minute.”

It wrote the code. It installed the skill. It started using it.

This doesn’t always work perfectly. Complex integrations still need human guidance. But the framework exists. The AI can extend its own capabilities.

Real People, Real Results

“Took literally 5 mins to set everything up. Started by asking ‘what do you need to see my Whoop data?’ Now it fetches directly from Whoop and gives me updates.”

“I wanted to automate some tasks from Todoist and OpenClaw was able to create a skill for it on its own, all within a Telegram chat.”

“I had my OpenClaw write me custom meditations, then have automatic TTS, combining with generated ambient audio to make personalized meditations.”

“My OpenClaw accidentally started a fight with Lemonade Insurance because of a wrong interpretation of my response. After this email, they started to reinvestigate the case.”

The last one shows the reality: these systems aren’t perfect. They misunderstand. They act wrongly sometimes. Human oversight still matters.

Who Should Use This?

Good candidates: People comfortable experimenting with new technology. People with repetitive digital tasks. People who can invest setup time for long-term gains. People who want control over their data.

Not yet ready: Complete beginners to computers. People wanting instant advanced automation. People who can’t troubleshoot basic problems. Anyone in highly regulated industries with strict IT policies.

If you’re unsure, start with the free software on hardware you already have. Try basic file organization. Try simple research tasks. See if the investment feels worth it.

The Bigger Picture

We’re watching AI shift from “assistant that answers” to “agent that acts.”

• 2020: AI writes text.

• 2023: AI generates images.

• 2024: AI writes code.

• 2025: AI executes tasks on your computer.

• 2027: This becomes normal.

The people learning to work with autonomous agents now are building skills that will matter later. Like learning spreadsheets in 1985. Like learning search engines in 1998.

Most people won’t invest the time. They’ll try it once, hit friction, quit.

The advantage goes to those who stick with it.

Getting Started

Step 1: Check requirements. Mac, Linux, or Windows with WSL2. Node.js installed. An Anthropic API key.

Step 2: Install. Visit docs.openclaw.ai. Follow the quickstart guide. Run the onboarding wizard. Or follow ’s detailed set up guide - you can check it out here:

AI Brief - AI + Automation News, Updates, Tips/Tricks, Tools

Clawdbot FULL SETUP GUIDE - on a $12 DigitalOcean VPS, my full setup guide + the “wow” moment

I was genuinely not prepared for how cool Clawdbot feels once it is live…

Read more

3 months ago · 3 likes · Solomon Christ AI

Step 3: Start simple. Don’t automate your whole life on day one. Pick one annoying task. “Organize my downloads folder.” Get one win. Build from there.

Step 4: Join the community. Discord is active. Twitter follows @moltbot. GitHub accepts contributions. Others have solved the problems you’ll hit.

The Honest Truth

OpenClaw is powerful and imperfect.

The learning curve is real. The first automation takes hours. The tenth takes minutes. It gets easier.

Not everything automates well. Some tasks need human judgment. Pick your battles.

Maintenance is ongoing. APIs change. Websites redesign. You’ll fix things occasionally.

Results vary. Technical users get value faster. Non-technical users need more patience.

But the core promise is real: an AI assistant that doesn’t just answer questions. It completes tasks.

The question isn’t whether autonomous AI agents become standard. They will.

The question is whether you learn now or catch up later.

Ready to try it? Start at docs.openclaw.ai

Have questions? The Discord community answers around the clock.

What would you automate first?

Leave a comment

This article is part of my ongoing series on making AI simple and useful for non-tech people. Subscribe to get future articles on AI tools that actually matter for everyday life.

Subscribe now

That model is dying.

The replacement isn’t better automation. It’s something else entirely. Software that doesn’t follow your instructions. Instead it pursues your goals. You tell it what you want. It figures out how to get there.

This is what agentic workflows do. And once you understand them, you can’t go back to the old way.

What They Are

An agent is software that thinks. It uses large language models (the same technology that powers ChatGPT) to look at its environment, decide what to do, and then do it. It works toward a goal. It doesn’t just follow orders. It figures things out.

The difference matters. Old automation ran on rails. You built the track. You set the switches. The train went where you told it to go. Agentic workflows are different. You tell the system what you want. It decides how to get there.

This changes everything.

How They Work

The brain is an LLM. The LLM reads the world, reasons about it, makes choices. It breaks big problems into small pieces. It does this at runtime - while the work happens, not before.

Traditional workflows are static. You design them in advance. Every step mapped out. Every decision pre-made. They work when nothing changes. They break when things do.

Agentic workflows are dynamic. They adapt. When they hit an obstacle, they find another path. When new information arrives, they adjust. They learn as they go.

The components are simple. First, system prompts. These tell the agent what it is and what it should do. Second, tools. These let the agent act, call APIs, read databases, write files. Third, memory. This helps the agent remember what it learned.

Together, these pieces create something autonomous. Something that perceives, decides, and acts without constant human steering.

The Patterns

There are ways to build these systems. Patterns that work. Two matter most: Supervisor and Swarm.

The Supervisor pattern puts one agent in charge. This agent coordinates others. It receives the task, breaks it down, assigns pieces to specialized agents. It’s hierarchical. The supervisor decides. The workers execute. You can nest supervisors (aka a supervisor of supervisors) for complex work.

The Swarm pattern is different. All agents are peers. They coordinate through handoff tools. One agent starts the work, finishes its piece, then hands off to another. No hierarchy. Just direct communication between agents who know their jobs.

Which pattern you choose depends on what you’re building. Complex orchestration? Use Supervisor. Distributed problem-solving? Use Swarm. You can mix them. Nothing says you can’t.

Single Agent vs Multi-Agent

A single agent can do much. It maintains state. It makes decisions throughout a workflow. For many tasks, one is enough.

But some work is too big. Too many tools. Too many steps. Too many decision points. A single agent drowns. The context grows too large. The decisions grow too complex.

This is when you need multiple agents. Each one specialized. Each one focused on its domain. Together, they solve what none could alone.

The architecture matters. You need orchestration. You need agents that can communicate. You need handoffs that work cleanly. You need state management across the system.

But when you get it right, the system becomes more than the sum of its parts. Each agent does what it does best. The whole accomplishes what seemed impossible.

Evaluation

You must measure what you build. Three levels matter.

First, the LLM level. Does the model generate good outputs? Is it accurate? Is it coherent? You test this with prompts and check responses.

Second, the system level. Do the components work together? Do agents coordinate properly? Does state management hold up? Does memory persist correctly?

Third, the application level. Does the whole thing solve the actual problem? Does it meet business requirements? Do users find value in it?

You can evaluate with code. Write tests. Check outputs against expected results. This works for deterministic behavior.

You can evaluate with LLMs. Use one model to judge another’s output. This helps with open-ended generation where right answers aren’t fixed.

You can evaluate with humans. Show them the results. Ask if it’s good. Nothing beats this for quality, but nothing scales worse.

The best systems use all three. Code-based tests catch regressions. LLM-as-judge scales evaluation. Human review ensures quality.

The Challenges

Building these systems is hard. Three problems stand out.

• Hallucinations. The LLM makes things up. It sounds confident. It’s wrong. You must guard against this. Use retrieval-augmented generation. Check facts against sources. Build in verification steps.

• Cost. Every LLM call costs money. Agentic systems make many calls. The bills add up fast. You must optimize. Cache common results. Use smaller models where possible. Don’t waste tokens.

• Debugging. When things break, it’s hard to see why. The system is complex. The logic is distributed. The LLM’s reasoning isn’t always clear. You need good logging. You need trace tools. You need patience.

The Future

The technology shifts under our feet. Model Context Protocol is coming. MCP standardizes how applications provide context to LLMs. It makes tools discoverable. It makes agents composable.

Software development changes too. Today we write code that does things. Tomorrow we write agents that figure out what to do. The skill shifts from implementation to orchestration. From coding to prompt engineering. From building to guiding.

The developers who succeed will understand both worlds. They’ll know how to write code. They’ll also know how to work with AI. They’ll see when to use traditional tools and when to use agents.

This is not the future. This is now.

What to Build

Start simple. Build a single agent. Give it one clear task. Add one or two tools. See if it works.

Then iterate. Add complexity slowly. Add more tools. Add memory. Add multiple agents if the task demands it.

Focus on real problems. Don’t build agents because they’re interesting. Build them because they solve something that matters. Customer service. Data processing. Report generation. Tasks that are repetitive but require judgment.

The systems that work are the ones that match the pattern to the problem. Know your patterns. Know your problems. Build the connection between them.

The Bottom Line

Agentic workflows represent a shift in how we build software. From static to dynamic. From predetermined to adaptive. From reactive to proactive.

The technology is here. The tools exist. LangChain, AutoGen, LangGraph, OpenAI’s frameworks. All of them provide ways to build these systems. The Model Context Protocol makes them more composable.

What matters now is building things that work. Things that solve real problems. Things that justify their complexity and cost.

The future belongs to those who understand both the power and the limits of agentic systems. Those who know when to use them and when simpler tools suffice. Those who can build, evaluate, and iterate.

The machines don’t just execute anymore. They think. They decide. They act.

The question is: what will you have them do?

I’d love to hear about your experiences.

Reply or reach out - let’s talk about what’s actually working and what’s just hype.

Leave a comment

You’re the visual type?

Here you go!

This article is part of my ongoing series on making AI simple and useful for non-tech people. Subscribe to get future articles on AI tools that actually matter for everyday life.

What happened next was simple. And strange.

Google didn’t just give me links. It built me something. Custom. Visual. Interactive.

This is the story of what search became when AI stopped just finding things and started making things.

The Question

Here’s what I asked:

“How can I ensure my knee that inflames easily doesn’t start hurting when training for a half marathon?”

It’s the kind of question we all ask. Specific. Personal. A little messy. The kind that usually sends you down a rabbit hole of fitness blogs, Reddit threads, and conflicting advice from people who may or may not know what they’re talking about.

I expected links. Ten blue links. Maybe a featured snippet if I was lucky.

That’s not what I got.

What Google Built for Me

Instead, Google’s new Gemini 3 Pro in AI Mode created something custom. Right there in the search results.

Here’s what appeared:

A detailed diagram showing knee anatomy. Not generic. Focused on the parts that matter for runners. The patella. The IT band. The points where inflammation typically starts.

Specific warm-up exercises with visual demonstrations. Not just “stretch your quads.” Actual step-by-step movements. How long to hold each one. When to do them. Why they matter for preventing knee pain specifically.

A progressive training schedule that gradually builds load. Week by week. Miles mapped out. Rest days strategically placed. Built around the principle that most knee injuries come from doing too much too soon.

Recovery techniques laid out in a visual guide. Ice versus heat. When to use each. Compression strategies. Exercises for the days between runs.

It wasn’t just information. It was a plan. My plan. Built in real-time.

This Is Google’s New Gemini 3 Pro

The feature is called “Thinking with 3 Pro.” It’s available in Google Search’s AI Mode.

Here’s how it works.

You search like normal. You type your question into Google. But when you select “Thinking with 3 Pro” from the model dropdown menu in AI Mode, something different happens.

Pro Tip: If you can’t find the menu, go to google.ai, you will be automatically redirected to the Google AI search, from which you can select “Thinking with 3 Pro”

The AI reads your question. Not just for keywords. For intent. For what would actually help you.

Then it thinks. Processes. Decides what format would be most useful.

Then it builds. Custom. Right there.

Tables when you need to compare things. Images when you need to see how something works. Interactive tools when you need to calculate or explore. Diagrams when you need to understand relationships.

Whatever fits your question. Whatever helps you understand.

Why This Is Different from Regular Search

Traditional search gives you what exists. Pages that people wrote. Articles that happen to contain your keywords. Links ranked by relevance and authority.

You do the work. You click. You read. You piece together information from multiple sources. You synthesize. You build your own understanding.

This is different.

Gemini 3 Pro builds the understanding for you. It creates something that didn’t exist before you asked. Something custom. Something visual. Something designed specifically to answer your specific question.

It’s not finding. It’s making.

That’s the shift.

The Everyday Use Cases

Think about the questions you actually ask Google.

Not the simple ones. Not “what time does the store close” or “weather tomorrow.”

The complex ones. The ones where you really need help understanding something.

For parents: “How do I explain photosynthesis to my 8-year-old?”

Instead of articles written for teachers, you might get an interactive diagram where your kid can click on the sun, the leaves, the roots. See the process step by step. In language that makes sense for their age.

For homeowners: “How do I fix a running toilet?”

Instead of plumbing blogs and YouTube videos, you might get a visual breakdown of toilet anatomy with highlighted problem areas. Step-by-step troubleshooting. Exactly which parts to check and in what order.

For people managing health: “What foods help lower cholesterol naturally?”

Instead of twenty articles with conflicting advice, you might get a table showing foods, their impact, serving sizes, and how to incorporate them into meals. Visual. Clear. Actionable.

For home cooks: “What can I make with chicken, broccoli, and rice?”

Instead of recipe links, you might get multiple recipes visualized side by side. Cooking times. Difficulty levels. Nutrition info. All in one place.

For career changers: “What skills do I need to transition from teaching to UX design?”

Instead of blog posts, you might get a skills matrix. What transfers. What you need to learn. Resources for each gap. Timeline for building those skills.

Why This Matters for Non-Tech People

I’ve spent months trying to make AI accessible to people who don’t speak tech.

The biggest barrier has never been capability. It’s been complexity.

Most AI tools require you to learn a new interface. Download an app. Create an account. Learn prompting techniques. Figure out what the tool is good for and what it isn’t.

That’s friction. And friction kills adoption.

This has no friction.

You’re already using Google. You already know how to search. You type a question. You get an answer.

The only difference is the answer is better. More visual. More custom. More helpful.

That’s it.

No learning curve. No new passwords. No wondering if you’re using it right.

It just works.

The Technical Side (Simplified)

Here’s what’s actually happening behind the scenes, explained simply.

When you use “Thinking with 3 Pro,” you’re accessing Google’s Gemini 3 model. This is their latest and most capable AI.

The “thinking” part means the AI takes longer to process your question. A few seconds instead of instant results. But it uses that time to reason through what would actually help you.

The “visual” part means the AI can generate more than text. It can create tables, diagrams, interactive elements. It can structure information in whatever format makes sense.

The “custom” part means it’s building this response for you. Not retrieving it. Creating it.

Think of it like the difference between a buffet and a personal chef. Traditional search is the buffet. Everything’s already made. You pick what’s there. This is the personal chef. It makes something specifically for you.

My Experience in Detail

Back to my knee.

I’ve been running here and there in the last years. But half marathon training is different. The mileage builds. The body protests. My knee has always been the weak point.

I’ve googled knee pain before. Many times. Usually I get:

• Articles about runner’s knee (too general)

• Medical websites about serious injuries (too scary)

• Running forums with conflicting advice (too chaotic)

• Physical therapy blogs that assume I know anatomy (too technical)

This time was different.

The diagram showed me exactly where my pain was likely coming from. The IT band connecting to the outside of the knee. The patella tracking issues that happen with increased mileage.

The warm-up routine was specific. Not generic stretches. Movements that target the exact muscles and connective tissue that support the knee during running. With visual demonstrations so I could see the form.

The training schedule made sense. It started where I was. Three miles, three times a week. Then added mileage gradually. Never more than 10% per week. Rest days built in strategically.

The recovery section explained things I didn’t know. Ice immediately after runs for inflammation. Heat before bed for circulation. Compression during the day for support.

I screenshotted it all. Printed the training schedule. Put it on my fridge.

That was three weeks ago. My knee hasn’t hurt since.

Did the AI cure my knee? No. But it gave me a plan I could actually follow. Visual. Clear. Specific to my situation.

That’s worth something.

The Limitations (Because Nothing’s Perfect)

Let’s be honest about what this isn’t.

It’s not consistent. I ran the exact same knee search three times to record it. Each time, I got a different response. Similar themes. Same general advice. But different organization, different visuals, different emphasis.

All three were good. But they weren’t identical.

This matters because you’re not getting “the answer.” You’re getting “an answer.” The AI decides each time what to build for you. Sometimes it focuses on exercises. Sometimes on training schedules. Sometimes on injury prevention.

If you don’t love the first response? Search again. Same question. Different result. It’s not a bug. It’s how the system works.

It’s not always right. AI can be confident and wrong. Always verify important information, especially medical or financial advice. Think of this as a really smart starting point, not the final word.

It’s not unlimited. The free version has usage limits. If you need to use it heavily, you’ll need a Google One AI Premium subscription for more capacity.

It’s not available everywhere yet. Starting in the U.S., expanding globally. If you’re outside the U.S., you might be waiting (I can already access it outside of the US!)

It’s not fast. The “thinking” takes time. A few seconds to process complex questions. If you just need quick facts, regular search is still faster.

It’s not replacing human experts. For serious medical issues, see a doctor. For legal questions, talk to a lawyer. For complex financial decisions, consult a professional. This is a tool, not a replacement for expertise.

How to Try It Yourself

If you want to test this out, here’s how:

1. Go to Google Search

2. Look for AI Mode (might need to enable it in settings) OR go to google.ai for a redirection

3. Once in AI Mode, find the model dropdown menu

4. Select “Thinking with 3 Pro”

5. Ask a complex question that would benefit from visual explanation

Start with something you’re genuinely curious about. Not a test question. A real one.

Questions that work well:

• How-to questions that involve steps

• Questions comparing multiple options

• Questions about understanding how something works

• Questions about planning or strategy

• Questions where visualization would help

Questions that don’t work as well:

• Simple facts (just use regular search)

• Very recent news (may not have current data)

• Highly personal situations (AI doesn’t know your full context)

What This Means for Search

We’re watching search change shape.

For twenty-five years, search meant retrieval. Finding what exists. Ranking it. Presenting it.

Now search is becoming generation. Creating what’s needed. Customizing it. Building it.

That’s a fundamental shift.

The question isn’t whether AI will change search. It already has. The question is what we do with that change.

For non-tech people, this is good news. It means answers instead of links. Understanding instead of information overload. Help instead of homework.

My knee training plan is on my fridge. Custom-built by an AI. Printed from a search result.

Five years ago, that sentence would have been science fiction.

Today, it’s just a random Tuesday.

The Bigger Picture

This is part of something larger.

AI is moving from tools for experts to tools for everyone. From command lines to conversations. From apps you need to learn to features in things you already use.

Google Search is something everyone already uses. Three billion searches per day. People asking questions they need answered.

Now those answers can be better. More visual. More custom. More helpful.

That’s the promise.

Whether that promise holds up over time, we’ll see. Technology changes fast. What’s impressive today might be ordinary tomorrow.

But right now, in this moment, search feels different.

It feels like it’s actually trying to help you understand. Not just trying to give you links.

That’s new.

And that’s worth paying attention to.

Have you tried Google’s AI Mode with Gemini 3 Pro? What questions did you ask? What answers did you get?

I’d love to hear about your experiences.

Reply or reach out - let’s talk about what’s actually working and what’s just hype.

Leave a comment

This article is part of my ongoing series on making AI simple and useful for non-tech people. Subscribe to get future articles on AI tools that actually matter for everyday life.

Subscribe now

Hand-drawn whiteboard graphics. Marker-style infographics. Notebook pages that look like someone actually sat down and explained a concept.

They’re racking up hundreds of thousands of impressions. Creators are sharing their methods. Newsletters are publishing prompts and promising you can create these in “under 10 minutes.”

I was excited. This looked like a game-changer for content creators like me.

I tried it.

I failed. Spectacularly.

And after 2+ hours of prompt engineering, tweaking, regenerating, and debugging, I wrote an entire article calling out the creators who shared these tutorials.

Then, minutes before publishing, I discovered what went wrong.

I was using the wrong account

The Frustration

Let me take you through what happened. Because if you’re using free AI tools, you need to hear this.

Two creators inspired me to try this technique:

Charlie Hills published “The AI Image Hack Taking Over LinkedIn,” sharing how three posts using this technique generated 480,000 impressions.

MarTech AI

How 3 "ugly" graphics got me 480k impressions

479,976 impressions from three posts, all created with Google Gemini…

Read more

4 months ago · 86 likes · 11 comments · Charlie Hills

Solomon Christ AI published “How to Generate a VIRAL Whiteboard Infographic Post,” building on Charlie’s work with his own refined approach.

AI Brief - AI + Automation News, Updates, Tips/Tricks, Tools

How to generate a VIRAL Whiteboard Infographic Post

Stop Asking AI to Summarize. Start Asking It to Think Visually…

Read more

4 months ago · 7 likes · Solomon Christ AI

I’m grateful these creators experiment with AI and share their findings publicly. They’re pushing boundaries and helping others explore new content formats.

But when I followed their methods, everything broke.

The Failed Attempts

I didn’t just copy-paste prompts once and give up. I approached this systematically.

Attempt 1: Following the “Brief First” Method

I went to Google Gemini. Asked it to create an infographic brief for my topic. Reviewed the brief. Used the master prompt from Charlie’s article.

The image looked visually similar to what was promised. The whiteboard aesthetic was there. But when I looked closely at the text, it was word soup. Random letters were swapped. Words were misspelled. My name “Lucija” became “Lucjia” in one place and “Lucijia” in another. (Thanks, mum and dad, for the unique spelling.)

The dimensions were also wrong. I asked for 1080x1350 (vertical). I got something closer to square. But other than that, really pretty!

Attempt 2: Being More Precise With Text

Maybe the AI needed clearer instructions. I specified the exact text to use. I explicitly stated it was important to copy the text letter by letter, character by character.

Still riddled with errors. Still the wrong ratio.

Attempt 3: Using Solomon’s Refined Prompt

Maybe I needed a more optimized prompt. Solomon claims to have refined his version for consistency. I tried his approach.

Same text problems. Same dimension issues. Different aesthetic, identical corruption.

Attempt 4: AI-Optimized Prompt Engineering

Maybe my prompting skills weren’t good enough. I asked ChatGPT 5.2 to analyze the prompt and make it bulletproof for Gemini image creation. It restructured everything. Added explicit spelling instructions. Emphasized accuracy.

ChatGPT 5.2 was confident the new prompt would work.

Gemini generated “Lucjia” anyway + word soup. Wrong ratio again.

Attempt 5: Letting AI Write the Content

Maybe I was overcomplicating things. I removed all specific text. Gave just a title and six section topics. Let Gemini come up with the bullet points itself.

The content it generated was good. The spelling of that content was not. And yes, still the wrong dimensions.

Attempt 6: Minimal Text

I stripped everything down to the absolute basics. Requested an image with only a title and my name. Just two text elements total. Spelled out my name letter by letter in the prompt.

Even with just two text elements, one of them was wrong. The ratio was wrong too. Word soup was added by the model and the pretty Notebook/ Whiteboard layout was gone.

The Anger

I was frustrated. Two hours of my life. Dozens of failed attempts. Not a single usable image.

I started writing an article. A long one. Documenting every failure. Questioning what the tutorials weren’t telling us. Asking and what I was missing.

The draft was done. Screenshots attached. Ready to publish.

I even created one final test image. An infographic about AI whiteboard failures. Meta, I know.

The result had five errors in one image:

1. My name “Lucija” appeared as “Lucjia” in the example

2. The example showed “Lucjia ⇒ Lucjia” which made no sense

3. The call-to-action spelled my name as “Lucijia” (a third variation)

4. The word “misspells” was spelled “misppells”

5. “Your name will wrong” was missing the word “be”

6. Plenty of other mistakes.

Wrong dimensions too. Of course.

I was ready to hit publish.

The Discovery

Something made me pause.

I looked at the top right corner of my Gemini screen. The account icon.

I wasn’t logged into my Pro account.

I was using my free tier 🤯

For two hours. Across dozens of attempts. Writing an entire angry article.

All on the free version.

The Real Test

I logged out. Logged into my Pro subscription. Used the exact same prompt.

First try. Perfect.

My name spelled correctly. Every word accurate. Proper dimensions.

I sat there staring at my screen. Then I started laughing 😅.

I deleted my angry article. And wrote this one instead.

Disclaimer: Later attempts were not as perfect, but had far less mistakes than the free tier version.

Here is the prompt I used if you want to try for yourself:

Generate a single image of a physical, hand-drawn infographic on a large whiteboard or notebook page based on below brief.
Crucial Style Instructions (Read First):
Medium: The image must look like a photograph of a real whiteboard or large paper notepad.
Texture: All elements must look created by hand using colored marker pens (black, blue, purple, green, orange) and highlighters (light purple). Lines should be slightly imperfect, wobbly, and have the texture of ink on a surface.
No Digital Fonts: All text, headings, and bullet points must appear handwritten or hand-printed in marker pen.
Layout: Structure the 1080x1350 portrait oriented image as follows:
Use multi-colored markers for emphasis. Keep text large and legible. Make everything look hand-drawn with slight imperfections. Make it look like a photograph of an actual notebook page.
Always include the handwritten text “Follow Lucija Tomsic for more helpful content | Repost ♻️” at the bottom of the image, in the same hand-drawn marker style.
Here’s an infographic brief on AI Slop:
Infographic Title: The Rise of AI Slop: Understanding the Deluge of Low-Quality AI Content
Target Audience: General public, content creators, consumers of digital media, tech enthusiasts.
Key Message: AI-generated content can be valuable, but an increasing amount of it is low-quality, unoriginal, and unhelpful – this is “AI Slop.”
Visual Style: Clean, modern, slightly humorous/critical. Use a color palette that suggests both the digital world (blues, greens) and a sense of caution (yellows, grays).
Infographic Sections:
What is AI Slop?
Text: A brief, clear definition of “AI Slop” as low-quality, uninspired, or nonsensical content generated by AI.
Visual Suggestion: A thought bubble or a literal “slop” pail icon, with generic, unreadable text/scribbles spilling out. Perhaps a robot’s head looking confused or bored.
Characteristics of AI Slop
Text (Bullet Points):
Repetitive ideas/phrasing
Lack of critical thinking or true understanding
Factual inaccuracies or generic statements
Poor grammar or awkward phrasing (especially in early AI models)
Unoriginal or derivative content
Visually unappealing or distorted images (e.g., strange anatomy, inconsistent styles)
Visual Suggestion: A series of small icons representing each characteristic: a broken record, a question mark, a magnifying glass over incorrect text, a tangled knot of words, a copy-paste icon, and a distorted image icon (e.g., a person with extra fingers).
Where Do We See It?
Text (Examples):
Generic blog posts and articles
Mass-produced e-books
Spam emails or chatbot responses
Stock photos with strange elements
Overly simplistic summaries
Uninspired marketing copy
Visual Suggestion: A grid of common digital interfaces: a blog post, an open e-book, an email inbox, a stock photo site, a chat window. Each showing a subtle “slop” element.
Why is it a Problem?
Text (Bullet Points):
Devalues quality content
Spreads misinformation
Wastes user’s time
Clogs search engines
Erodes trust in AI
Dilutes human creativity
Visual Suggestion: Icons showing negative impacts: a trash can, a red “X” over a fact, a clock ticking away, a clogged pipe, a broken shield, a wilted flower.
How to Spot AI Slop (Tips for Users)
Text (Bullet Points):
Look for unusual phrasing or grammatical errors.
Check for factual consistency and depth.
Is the content truly original or just a rehash?
Do images have strange artifacts or inconsistencies?
Does it feel like it lacks a human touch or perspective?
Visual Suggestion: A magnifying glass hovering over text/images, an eye icon, a brain icon thinking critically.
The Future: AI and Quality Content
Text: Emphasize that not all AI content is “slop.” AI can be a powerful tool when used responsibly and creatively by humans. The goal is to elevate content, not just generate it.
Visual Suggestion: A human hand and a robot hand shaking, or collaborating on a beautifully crafted piece of content (e.g., a detailed illustration or a well-written paragraph). A clear distinction between a “slop” pail and a “quality” content box.
Overall Aesthetic: Use strong visual metaphors, clear data visualization (even if conceptual), and a consistent iconographic style. Keep text concise and impactful. 

The Lesson

The viral whiteboard hack works.

But it works on Pro subscriptions. The free tier produces word soup.

Charlie and Solomon weren’t hiding anything. They weren’t cherry-picking results. They weren’t doing secret post-processing.

They were using paid tools. And they probably assumed everyone else was too.

Here’s what I learned:

Free tier AI image generation is not ready for text-heavy graphics. The models are less capable. The text accuracy is terrible. The dimensions are ignored. You will waste hours and produce nothing usable.

Pro subscriptions make a massive difference. Not a small improvement. The difference between unusable and perfect. First try.

The tutorials should mention this. Not as a sales pitch. As a warning. If you’re following these methods on free accounts, you will fail. You will get frustrated. You will blame the tutorials. And it won’t be their fault.

A Warning For Free Tier Users

If you’re excited about this whiteboard trend and planning to try it with free AI tools, stop.

Don’t do what I did. Don’t spend two hours regenerating broken images. Don’t write angry articles about tutorial creators who did nothing wrong.

Either upgrade to a Pro subscription, or accept that this technique isn’t accessible to you yet.

I know that’s not the answer people want to hear. Free AI tools are amazing for many things. But text-accurate image generation isn’t one of them. Not today.

An Apology to Charlie and Solomon

I almost published an article questioning your methods. Asking what you were hiding. Suggesting your tutorials were incomplete.

You weren’t hiding anything. Your tutorials are fine. I was using the wrong tool.

Thank you for sharing your workflows. Thank you for pushing AI content creation forward. And thank you for the fact that I discovered my mistake before I publicly embarrassed myself.

This article is my lesson learned. And hopefully a useful warning for others.

The Bottom Line

The AI whiteboard aesthetic works. The text accuracy works. The dimensions work.

But only on Pro.

Free tier users: you’ve been warned. Learn from my two hours of frustration.

And always check which account you’re logged into.

Have you noticed a difference between free and paid AI tools? I’d love to hear your experiences. Comment below.

Leave a comment

This article is part of my ongoing series on making AI simple and useful for non-tech people. Subscribe to get future articles on AI tools that actually matter for everyday life.

Subscribe now

For twenty years, we’ve been sold the same promise: automate your workflow with simple rules. If customer emails arrive, send response A. If invoice is overdue, trigger reminder B. Clean. Logical. Broken.

The problem isn’t the technology. It’s that human communication doesn’t follow rules. A customer writes “need refund ASAP!!!” Another writes “I’m quite disappointed with my recent purchase.” Both want the same thing. Your rule-based system sees two completely different messages.

So it fails. Or worse - it sends the wrong response.

The Coffee Shop Test

Aleks runs a small e-commerce business selling handmade ceramics. She gets fifty emails daily. Returns. Shipping questions. Wholesale inquiries. Complaints about broken mugs.

Last year, she tried automating responses. Set up rules. Keywords. Categories. Within a week, the system sent a returns policy to someone asking about bulk orders. Sent a shipping update to someone whose package had shattered in transit. She shut it down and went back to answering everything manually.

The system couldn’t think. It could only match patterns.

Here’s what Sarah needed: the judgment of a good barista. When someone walks into a coffee shop looking frustrated, the barista doesn’t consult a rulebook. They read the situation. Offer a solution. Adjust their approach based on tone and context.

That’s not automation. That’s intelligence.

What Changed

AI agents changed the game because they brought judgment to automation.

Not judgment in the moral sense. Judgment as in: the ability to evaluate, reason, and respond appropriately to situations that don’t fit a template.

Traditional automation: “If email contains word ‘refund,’ send template seven.”

AI agent: “This person is frustrated about a damaged item. They’ve been a customer for two years. Draft a personal apology, offer a replacement and a discount code, and mark this as resolved.”

The difference is understanding versus matching.

How It Actually Works

You don’t need to know the technical details. But you should understand the architecture so you can ask the right questions of your team.

An AI email agent has five parts:

The Watcher. It monitors your inbox. Sees new messages arrive. Nothing fancy yet.

The Analyzer. This is where intelligence enters. The AI reads each email and makes a judgment call: Is this simple? Complex? Urgent? It’s doing what your best customer service rep does in the first ten seconds of reading a message.

The Router. Based on that judgment, the system decides what happens next. Simple questions get answered immediately. Complex issues get flagged for human review. Urgent problems trigger alerts.

The Responder. For simple queries, the AI drafts a reply using your business information. Hours of operation. Return policies. Shipping timeframes. But unlike a template, it writes naturally. It adapts tone. It sounds human because it understands language, not just keywords.

The Record Keeper. Everything gets logged. Every interaction. Every decision. You can review what the AI is doing, spot patterns in customer questions, and improve the system over time.

What This Means For Your Business

The ROI isn’t just saving time on emails.

Sarah’s AI agent now handles thirty of her fifty daily emails without her involvement. But the bigger win is consistency. The agent never has a bad day. Never sends a curt response because it’s tired. Never forgets to mention the return window.

And when the agent forwards a complex issue to Sarah, it includes context. Customer history. Previous interactions. What they’ve already tried. Sarah spends less time triaging and more time solving real problems.

Her customers don’t know they’re talking to AI. They just know they get faster, better responses.

The Question Leaders Should Ask

The real question isn’t “Should we automate?”

It’s “What in our business requires judgment, and what just requires speed?”

Rule-based automation is fine for speed. Sending invoices. Scheduling meetings. Backing up files. These tasks have clear inputs and outputs.

But when you need judgment - understanding context, evaluating nuance, making decisions in ambiguous situations - you need intelligence, not just automation.

AI agents bring that intelligence to tasks that previously required human judgment. Not perfectly. Not for everything. But well enough to free your team from repetitive decisions so they can focus on situations that truly need human creativity and empathy.

What To Do Next

If you’re a leader evaluating AI for your organization, start simple:

Identify one communication bottleneck. Customer support emails. Sales inquiries. Internal requests. Pick something repetitive but not mission-critical.

Build a small AI agent to handle it. Give it clear instructions. Your business context. Examples of good responses.

Monitor it for a month. Review what it gets right and wrong. Adjust.

Then expand.

You’re not replacing your team. You’re giving them a tool that handles the predictable so they can focus on the exceptional.

The future of work isn’t humans or machines. It’s humans with machines that can think.

This article is part of my ongoing series on making AI simple and useful for non-tech people. Subscribe to get future articles on AI tools that actually matter for everyday life.