BUT - In July ChatGPT finally launched its desktop app with similar ‘work’ features as Claude. I ignored it for a whole month.

I used Codex before and built all my projects in a provider agnostic way (means I can switch between Claude Code and Codex anytime - read here how to set it up):

Last week however I decided to test the rest of the new ChatGPT ecosystem, mostly because there is one big differentiator compared to Claude: Imagen 2 - the insane image generation model that is included in ChatGPT.

So I went through the entire ChatGPT desktop app feature by feature, comparing each of it to Claude.

Chat, Work, Codex, Voice, Images, Canvas, Projects, Skills, Connectors. Every one of them gets its own real explanation here, what it does, how to actually use it, and where to find it when the sidebar doesn’t make that obvious.

What this covers

This one runs long. Let’s dive straight in.

What ChatGPT Desktop Actually Is

Similar to Claude Desktop: Three tabs, one app. Chat and Work share a single window, toggled in the message box. Codex sits apart, for anything code-related.

Voice and a built-in browser sit on top of all three, so you can talk to any of them or point them at whatever’s already open in your other tab. Claude Desktop landed on the exact same shape independently, Chat and Cowork sharing a window, Claude Code standing apart - which is worth knowing before you assume one app is simpler than the other.

Free, Go, Plus, or Pro: What You Actually Need

There’s a tier between Free and Plus now, called Go, and most comparison posts still don’t mention it. It is my go to in months I do not have to code or create as much. (fyi, I own a Claude Pro subscription that covery my daily tasks; on usage heavy months I upgrade to ChatGPT Plus instead of Claude Max. This setup gives me more than enough usage on only 40 bucks per month).

Chat

Classic everyday mode. Use it for quick questions, brainstorming, first drafts, web search, all in one conversation - but no local file access.

If you’ve used claude.ai, this is the same shape you already know. The differences show up in the other tabs, not here. The only major difference is that you can let ChatGPT create images straight from the chat - more on this later.

Work

You can hand Work a brief and your files, walk away, and come back to a finished document, deck, spreadsheet, or Site. It runs on the GPT-5.6 model family and can work on its own for a while, only checking in with you when it needs a decision. You can also give it direct access to your files and computer.

Prompt example: 
“Turn this outline and these two reference files into a 45-minute presentation, sections, timing, talking points.” 
Review its plan before it starts moving files around, the same review step Claude’s Cowork uses.

Codex

The engineer. Codex can read and edit repositories, run tests, debug, work in a terminal… and now it can also handle multi-repo diff reviews and GitHub PR chats directly inside the desktop app. It’s the same job Claude Code does, both sitting apart from the main chat window as their own space.

Skip this tab entirely if you don’t build software. Nothing else in the app depends on it.

Voice

Your hand hurts? Talk to Chat, Work, or Codex instead of typing. Start a task, check on it, or redirect it out loud, including steering something Work is already running in the background.

Claude’s voice mode is a genuine two-way conversation too, on desktop, mobile, and web. What I couldn’t confirm is whether it reaches into a running Cowork session the same way ChatGPT’s Voice reaches into Work, so treat that specific comparison as open.

Built-in Browser

One browser, built into the desktop app, same tab, same login. It can see what you see, click around a page, download files, and annotate elements you point it at.

Claude splits this job two ways: the Claude in Chrome extension for regular browsing, plus Cowork’s own sandboxed computer-use ability for tasks that need to click through a page. Same capability, different packaging.

Share

Images: The Feature Claude Still Doesn’t Have

This is the one that isn’t even close. ChatGPT generates original images natively, inside the same chat, no plugin needed.

I tested it myself: I asked for a full set of on-topic images for a social carousel, one visual style, five slides. It delivered with one prompt (you can tell by the look of the carousel that it needs some improvement), no second tool, no separate login needed.

Claude has zero native image generation.

What it has instead are diagrams and interactive mockups built from code. Also it can do some basic svg drawings. This is genuinely useful for charts and prototypes but not for illustrated or photorealistic images. To match what ChatGPT just did, you have to leave Claude entirely, open a separate image tool, a separate login, and hand a manual copy back in. Alternatively you can connect Claude Code to e.g. Nano Banana or Higgsfield to achieve the same result more seemless. In both cases you need another subscription and more steps.

Canvas & Sites vs. Artifacts: Who Actually Wins

Canvas used to be ChatGPT’s side-by-side writing and code workspace, highlight a sentence, ask it to rewrite just that part, adjust reading level with a slider, all without leaving the document. Used to, because it’s gone.

There’s no side-by-side Canvas panel anymore - it got retired on May 28, 2026 and folded into the chat itself. What you get now instead: inline writing blocks and code blocks that sit directly inside a chat response, editable in place, with a live preview for code (HTML, React, SVG, Mermaid, Vega charts) and console output for Python. It’s the same underlying trick: highlight a section, ask for a targeted revision, get the change without losing everything else. The old Canvas UI survives only on legacy models, GPT-4.5 already retired, o3 retiring August 26, so it’s functionally gone for anyone on a current model and you don’t need to worry about it.

There’s a second ChatGPT feature that’s the actual competitor to Claudes’ Artifacts: Sites, shipped July 9, 2026 alongside ChatGPT Work. It lets you ask for a full interactive website or app, dashboards, trackers, internal tools, and get back a live, hosted URL. Even a custom domain can be included if you own one. Ask for it directly, or type @Sites inside Work or Codex, and ChatGPT builds a private preview first, then you choose who sees it: just you, specific people, your workspace, or the public internet.

OpenAI’s own release notes, last touched in mid-August, still list Sites’ public publishing and expanded rollout as blocked in the EEA, Switzerland, and the UK. That didn’t hold up for me as I was able to test it from inside the EU. Therefore, treat the official line with some skepticism rather than take it at face value, changelogs don’t always get revisited once written. What is confirmed either way: Free and Go don’t get Sites at all, paid plans only.

Claude Artifacts do the same core job: they generate something, publish it to a real link, so you can hand it to anyone, no separate hosting step. Every one saves as its own persistent object, playable if it’s an app, publishable with no account required on the other end. Since April 2026, Cowork’s artifacts can even connect to MCP servers and refresh with live data from a spreadsheet or database instead of going stale the moment you close the tab. The difference from Sites: Artifacts work on every plan including Free, and publishing is a click inside any normal conversation rather than a dedicated workflow.

Subscribe now

Projects: Where They Actually Live

Projects group related chats, files, and instructions into one workspace, so you stop re-explaining context every time you open a new conversation. They’re available across all tiers now, including Free.

Since the April 2026 redesign, files aren’t attached inside a Project’s own settings anymore, which is where people lose time. Every file you upload anywhere, including inside a Project, lands in the Library, a tab in the sidebar below Images, and you pull from there into any Project.

Library is still mid-rollout, worth knowing if you don’t see it. It launched to Plus, Pro, and Business only in March 2026 with the EEA region excluded entirely. As of this past week, OpenAI’s own release notes say it’s expanding to Free and Go accounts too, including in the EEA, so if it’s missing for you, you likely just haven’t been reached by the rollout yet. There's no confirmed workaround in the meantime. Settings → Storage was announced in the same rollout, so if Library isn't there, that probably isn't either (I don’t see it yet in my app, located in the EU).

To set a Project up: open the left sidebar, click New project, name it something specific (”Q1 client deck” beats “Work”), then drag a file into any chat inside it once. That file is now in your Library and reusable across every other Project too.

Claude’s Projects don’t work quite the same way here. Each Project has its own separate Files section, and a file you’ve already uploaded to one Project has to be uploaded again to use it in another. ChatGPT’s Library in which one upload is reusable in any Project, is the more convenient system on this specific point.

Skills & Plugins: The Feature That Isn’t Where You’d Look

This one’s easy to miss entirely. There’s no “Skills” item anywhere in the sidebar. It lives under Plugins → Skills, a tab inside the Plugin Directory, not a menu of its own.

Great news on skills

ChatGPT Skills use the exact SKILL.md format Anthropic created and open-sourced earlier for Claude Code. OpenAI didn’t invent a competing standard, it adopted Claude’s. A skill written well enough to stay format-agnostic can move between the two products.

Similar to Claude, you can build a skill four ways: write it in chat, use the editor, upload a file, or demonstrate it live with Record & Replay. That last one is the interesting one: open Skills, hit the “+”, select “Record a skill,” then perform the task once, e.g. filing an expense, booking a slot, publishing a report. ChatGPT watches and compiles the steps into an editable SKILL.md file you can run again anytime you need it.

Claude has almost the identical move, “Record a skill” sits in the “+” menu inside Cowork, same idea, same pitch: show it once instead of describing it. Availability is the catch on both sides. ChatGPT’s official pricing page lists Skills as available across every consumer tier, but plenty of Plus and Pro accounts don’t have it live yet. Check your own Plugins tab rather than trusting either the pricing page or this guide.

Connectors, Apps & MCP: Can ChatGPT Use MCP?

The simple answer is yes. And it’s not a workaround, it’s the same protocol Claude uses. MCP, the Model Context Protocol, is the open standard Anthropic created back in November 2024 to let any AI assistant connect to any external tool through one common interface. OpenAI adopted it across the Agents SDK, the Responses API, and the ChatGPT desktop app itself back in March 2025.

In December 2025, Anthropic handed MCP’s governance to a vendor-neutral foundation, so the same server that powers a connector in Claude can power the identical connector in ChatGPT. In practice, that means three doors: individual users on paid plans connect remote MCP servers through Developer Mode, Business and Enterprise workspaces can publish vetted MCP apps org-wide, and developers can wire any MCP server straight into the Responses API.

Full read-and-write MCP support, not just read access, is still rolling out in beta and currently favors Business, Enterprise, and Edu plans over individual Plus and Pro accounts. Documented friction is real too: custom apps intermittently vanishing from the directory, OAuth completing without the connector showing up in chat, and developers reporting the same MCP server behaving better inside Claude than inside ChatGPT.

Like what you read? Subscribe now!

Scheduled Tasks

Set a task once, have it run once, repeat on a cadence, or watch for changes, all from the sidebar on desktop, web, and mobile. Claude’s Cowork has had the same capability since February 2026, comparably mature, so this one’s a tie rather than a Claude-only feature.

Both need the app open and the computer awake for a local run; also both are rolling out remote execution that doesn’t.

Memory

ChatGPT’s memory setting is per-project now. This allows you to choose whether a Project draws on your general memory or stays scoped to only to what’s inside it. Claude’s Projects work the same way, context kept to the workspace it belongs to.

Where Each One Fails

ChatGPT:

• Skills and full MCP support are unevenly rolled out, the pricing page and the real account experience don’t always match.

• The surface still asks you to track three tabs and where a given feature actually lives, Plugins hosting Skills being the clearest example.

• Community reports of MCP connectors misbehaving are common enough to plan around, not dismiss.

Claude:

• No native image generation, at all. The one gap in this guide that's not close.

• Windows user-only quirk: Cowork's desktop app needs Windows Pro, Enterprise, or Education, since Windows Home doesn't ship with the Hyper-V it runs on. Mac has no such gate. It also runs from the browser and mobile now regardless of OS, so this only bites Windows Home users who need it touching local files.

• Three tabs still means three mental models, the same complexity tax as ChatGPT under a different logo.

In a Nutshell

Neither app wins the whole job anymore, they’ve converged too far for that. What’s left are specific differences: images, a publishable artifact, an MCP connector that shows up when it should.

Pick the tool for the task:

• need native images? pick ChatGPT

• need a good artifact you can share with you colleagues? Claude is your man

• You like Claude’s way of talking to you more? You know what to do.

Me for my part will set up my most important projects in both. Luckily copy pasting instructions and uploading skills that are written in a model agnostic way make it possible.

PS - Let me know which one you prefer.

Leave a comment

Subscribers get all my resources, incl. step-by-step guides, prompts, commands and workflows for free in the Resource section, the Chat and the Welcome email.

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

So I keep fixing the same things. Claude keeps forgetting I fixed them.

Eventually I got sick of it and decided a change is needed.

Not a new file - but a skill that remembers my edits and teaches Claude a lesson.

What this covers

This one runs long. It comes with a file.

Before we continue: If you missed my article on how to teach Claude to sound like you, read it here -

The mistake I keep making, on repeat

Open my last five drafts and you’ll find the same fixes.

I cut a hedge word Claude loves. I shorten the opening paragraph. Or I add a real number where Claude wrote a vague claim.

I thought my voice file solved this. Instead it solved something else.

A voice file is a photo. Accurate the day I took it.

It tells Claude how I sound. It says nothing about what I personally keep correcting, draft after draft.

Share

Where I stole the idea (and cut it down)

Wyndo, who writes The AI Maker (I can’t recommend him enough), rebuilt something Alex Lieberman showed on How I AI.

His system runs in four stages.

✦ An interview that won’t stop until you give it real stories

✦ A draft that only pulls from that transcript

✦ Five agent reviewers checking structure, hook, craft, slop, and reader value

✦ A step that compares your draft to your edit and saves the pattern

Four stages is more elaborate than I need at this point in time. I already have a process for ideas and draft writing. Admittedly, I could use some reviewers for my craft and hooks but I’ll leave this for another article.

For now, I only wanted the last piece: the one that watches what I fix and remembers it.

The anatomy of the fix

Here’s the whole thing, in plain English.

Imagine your voice file like a photo. It is an accurate representation of the moment you took it.

This setup is more like a diary. It keeps adding pages.

So, I built one Skill called /correction-lessons. It needs three files: the draft Claude wrote, the version I actually published, and a running file called lessons.md that starts empty.

Here’s what it does, step by step:

A. It reads both versions and finds what actually changed. And I don’t mean typos, but decisions: a cut sentence, a softened claim, a reordered section, a hedge word removed.

B. It ignores anything that only happened once. A pattern needs to repeat before it counts as a lesson.

C. For every real pattern, it writes one plain instruction, and attaches the real before-and-after line as proof. Never an invented example.

D. It adds new lessons to lessons.md under that day’s date. It never deletes the old ones. If a new lesson fights with an old one, it flags the conflict instead of quietly picking a winner.

That’s it. No interview panel. No five agent reviewers.

One file that gets slightly smarter every Monday (or whichever day you use for writing), forever.

Subscribe now

What happened when I ran it on this article

I ran /correction-lessons on my own edit of my previous and this exact article. Three real patterns showed up. One near-pattern got correctly thrown out.

Here a selection:

First: Claude’s draft assumed that everyone edits on the same day I do. I softened “Monday” -my edit day- twice: once in the opening, once in the anatomy section.

Lesson: stop assuming my schedule is the reader’s schedule.

Second: Claude stated metaphors as flat fact. My edit kept turning “X is Y” into “imagine X like Y” or “X is more like Y.” Same fix, twice.

Lesson: a metaphor lands softer as an invitation, not a declaration.

Third: Claude called Lieberman’s five reviewers just “reviewers.” Twice, I added the word “agent” back in to make it easier understanable for you guys.

Lesson: keep the AI-agent framing explicit when naming his subagents, don’t flatten it to a generic word.

One thing that first looked like a lesson wasn’t. Claude used several examples naming “five samples” and “three mistakes.” My edit softened both numbers, in the same sentence.

According to my /correction-lessons skill, that is one edit decision, not two separate patterns.

Consequently the Skill didn’t log it, because a fix has to show up twice, in two different places, before it counts as a lesson.

Three lessons. One correctly ignored. First real run, and it already did exactly what it was built to do.

Where it still needs you

This Skill has no opinion until you feed it two real versions of the same piece. Run it once and it finds nothing, because one edit pass isn’t a pattern yet.

It also can’t rescue a draft with no material. If there’s no real story, no number, no opinion that I’ve actually worked out, there’s nothing for it to learn from. That problem sits upstream of this fix, not inside it.

And when two lessons disagree, it doesn’t decide. You do. That’s on purpose.

Build your own

Two ways to run this. Pick whichever matches how you already work. Both lead to the same result: Claude has your lessons in front of it before it writes a single word of your next draft.

If you use Notion like me

1. Make three pages (or ask Claude to do it).

   One folder, three pages: Draft (original), Final (edit here), Lessons.

   Draft never changes. Final is the one you mark up. Lessons is the running log.

2. Get two copies of your first draft in.

   When Claude hands you a new draft, ask it to paste that exact text into Draft (original), and paste the same text again into Final (edit here).

   Now you have one frozen copy and one working copy.

3. Edit Final like any Notion page.

   Do your real edit pass there, not the rough one. Leave Draft (original) alone.

4. Ask Claude to compare them.

   Once your edit is done, tell Claude: “compare Draft and Final, update Lessons.”

   Claude reads both pages, finds anything that repeats, and adds it to Lessons.

5. Make it automatic.

   Add one line to your Claude Project’s instructions: “Before writing any draft, check the Lessons page in Notion first, and follow what’s there.”

Set it once. Claude checks it every time after. You never have to ask again.

Subscribe now

If you use plain files, no Notion

1. Make one folder.

   Call it anything, “substack drafts” works fine. Inside it, three files: draft.md, final.md, lessons.md.

   Start lessons.md empty.

2. Save two copies of your first draft.

   Save Claude’s draft as draft.md. Save the exact same text again as final.md.

3. Edit final.md in any text editor.

   Notepad, TextEdit, Google Docs, whatever you already have open works. Do your real edit pass there.

   Leave draft.md untouched.

4. Run the Skill.

   Upload both files into a chat with Claude, along with the correction-lessons Skill file. Claude compares them and hands you back new lesson entries.

   Paste those into lessons.md and save it.

5. Make it automatic.

   Upload lessons.md into your Claude Project’s files, once. Every chat inside that project can see it from then on, alongside your voice file, without you re-uploading it each time.

No project yet? Upload lessons.md fresh at the start of every drafting session, right next to your voice file.

That’s the whole system. No app to install, nothing to automate. One file, or one page, checked before every draft.

Comment below to grab the Skill for free.

Run it tonight, on the last thing you actually published and edited.

In a Nutshell

My voice file still says I write calm, direct sentences. It has no idea I’ve cut the phrase “to be honest” out of four drafts in a row.

Now it will.

Open your last edited draft tonight. Find the pattern you keep fixing. Feed it back in.

PS - Let me know how it goes. I encourage you to share ….

Leave a comment

Subscribers get all my resources, incl. step-by-step guides, prompts, commands and workflows for free in the Resource section, the Chat and the Welcome email.

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

This one is personal.

I’ve been ahead for years. I never learned how to say it out loud.

Two degrees. Four countries. Thirteen years across corporate insurance and an AI startup automating claims.

None of that made saying it any easier. If anything, it made the silence louder.

I was never the overachiever people assume. I’m lazy in the most efficient way possible, and I’ll take a mid-range grade for minimal effort every time.

I guess this is where my imposter syndrom comes from.

My master’s was the exception. I decided it mattered, so I finished with a 1, “sehr gut” in German grading, not because anyone was watching but because I wanted to prove to myself that I could.

Knowing you can do something doesn’t make you confident. It just means you’re good at proving it to yourself, and bad at telling anyone else.

It’s been seven months since I started brisk. And I’ve never told you why.

Not the AI part. That one’s obvious. The real one.

Here it is.

Subscribe now

The Rules Nobody Said Out Loud

I grew up as a Croatian immigrant in Germany. My parents left a war-torn country and arrived with a plan: work hard, don’t cause trouble, and eventually someone will notice and reward you.

The rules were never spoken out loud. They just showed up everywhere.

Get good grades. Don’t ask for too much. Stay invisible. Behave, and you’ll be safe.

There was a second rule underneath the first one. Respect anyone older than you, especially anyone with a title, and don’t question them even when something feels off.

And a third: what would the neighbors say. It was never really about the neighbors. It meant don’t do anything that could be seen and judged (who am I kidding, it was definitely also about the real neighbors) 😆

I understood where the rules came from. My parents survived things I never had to. But they stuck long after the reasons stopped applying to me.

If you were raised to keep your head down and trust the people above you, you already know these rules.

They don’t stay in your childhood home.

They follow you into every meeting where someone senior is wrong and you already know it.

Breaking It Once, Then Following It Again

At nineteen, two months into a business degree I didn’t want I called it quits. It felt so wrong, I had to break free. So, I found a family in Costa Rica and booked a flight. I couldn’t count to three in Spanish. I went anyway, working as an au pair, learning the language by living inside it.

By month four I was speaking Spanish without ever taking a class.

That was the first time I broke the rule. It worked. That should have been the whole lesson.

Instead I went for round two. Went back to studying. I finished my degree in Germany, the US, Argentina, and Brazil, on scholarships and part time jobs, because my parents couldn’t send me money to do it any other way.

I also built and ran a small business through most of it. Different story, for another time.

By then I had done a lot, but nothing with a name attached that would mean something on a CV. That’s what I was raised to chase: a name people recognized, proof that someone had noticed.

I applied to as many well-known (corporate) companies as I could find. The insurer in Munich was the fastest to say yes, so I ended up in an industry I hadn’t chosen, in an internship I’d taken for the name on the door, not the work behind it.

And I liked it anyway. I liked the office, the people, even the outfits.

After the internship, they even offered to pay for my master’s if I stayed and worked full time while studying nights and weekends. I said yes. The rule was back in charge, and now it had a manager attached to it.

Respect Your Elders, Even the Wrong Ones

For a long time I did exactly what I was taught. I looked up to anyone senior, anyone with more years in the job than me, and I gave them my respect before they’d earned it.

I didn’t question whether what they were doing was good. I assumed if they were older and higher up, they knew something I didn’t.

The longer I stayed in corporate, the less that held up. A lot of what I was deferring to was politics. People collecting a big salary while chasing whatever made this quarter look good, never looking past it.

Not everyone. Some of the best bosses I’ve had made corporate good to work in. But middle management, for the most part, was the weakest link I kept running into.

I started speaking up eventually. That part of the story isn’t true the way I told it: that I just kept my head down for thirteen years. I didn’t.

I was told I needed to be more diplomatic. I’m direct, which in Croatia isn’t a flaw, it’s just how people talk to each other. The rest of Europe doesn’t always agree.

I watched colleagues with weak ideas, the ones who never pushed back on anything, get promoted over people who knew what they were doing and thought for themselves. Corporate doesn’t reward competence. It rewards compliance.

Every time I tried to leave, I got offered more money and a better title instead. I used to think that meant I was good at my job. Now I think it meant the system was even better at keeping people who should have left.

That’s the real reason I didn’t post anything for thirteen years, on top of the “neighbor rule”:

If you’ve ever bitten your tongue in a meeting because the person talking had ten more years than you, you know this feeling. A lot of other overachievers do too. They don’t say so either.

I knew, somewhere under all of it, that I didn’t fit. I spent a decade assuming that meant something was wrong with me, and not with the room.

That’s part of why I started writing in public. I got tired of being the only one who knew what I thought.

What AI Gave Me

December 2022. A colleague mentioned ChatGPT before the holidays. I’d never heard of it, so I googled it that night.

I spent the entire Christmas break using it in secret, convinced they’d start charging a billion or shutting it down any day.

I wasn’t panicking. I was fascinated. I slept fine.

Until I realized around the same time that I had spend 10 years of my life in corporate.

That ten-year anniversary caught me off guard. I did the math and realized where it had all started: an industry I’d picked by accident, at a company that hired fastest, for an internship I’d only wanted for the name on my CV.

No more excuses. I left the insurer 2 months later and joined an AI startup automating insurance claims. One step closer into the direction I wanted. I told myself that was the break I needed. It wasn’t.

My clients were still corporate. The politics followed me there too.

In January, thirteen years into working, I finally posted something under my own name instead of hiding it. I started brisk.

I still find it cringe. But I do it anyway.

I did it on the side, without quitting my job. I also started taking on AI work of my own (hit me up if you need advice, an audit, education or implementation or book a call at brisk.vision).

Seven months later, I’m still writing this newsletter every week. That’s the part that surprised me, not the newsletter itself. The not quitting.

I truly enjoy it and can’t wait to do it full time.

What AI gave me was structure. For the first time, I had a way to take what was stuck in my head and get it onto a page in an order that made sense.

AI is the first thing that gave me a way to do it.

I’ve been ahead for a while. I always suspected it, and I could never quite own it out loud. Now I can, because I finally have the words.

That’s my why, in full. I write brisk because AI is the first thing that gave me a way to say what I already knew.

If you already sense you know more than you’re saying out loud, that’s worth paying attention to. Open Claude tonight and start structuring the thing you already know.

This is the most personal thing I’ve written for brisk. It’s also the first time I’ve introduced myself to you.

I don’t know exactly what that means for what I write next, so I’m asking you.

Reply and tell me which one you want more of:

PS - Hit reply. I read every single one.

Leave a comment

Subscribers get all my resources, incl. step-by-step guides, prompts, commands and workflows for free in the Resource section, the Chat and the Welcome email.

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

It shortens the distance between a skill you have and a person who will pay for it. Some of those distances are short, and most are longer than a lot of the gurus admit.

I do not code and I most likely never will. I taught myself AI for free with YouTube and dumb questions.

I also know the other side of this because I already failed at it once.

In January 2023, six weeks after ChatGPT launched, I made a little prompting e-book in Canva. I was so hyped, convinced I was about to get rich. Nobody was writing about this yet and I had got there first.

I listed it on Etsy and sold two copies. Fourteen euros, total.

So I quit, and told myself the market was not ready.

The market eventually showed up. But I was not there when it did.

Three years later, the advice in that e-book is what I publish every Thursday, to people who pay for it. The product was fine. I had no distribution and no patience, and I called that a market problem.

So this is the honest version. I rank these twelve streams by one thing only: how fast a beginner realistically sees a first payment.

This one runs long. Grab a coffee ☕. It is the whole map.

The rule that makes this list honest

I will not rank these by “potential income.”

That number is a guess dressed as a promise. Nobody knows what you will make, because your make depends on your skill, your niche, your effort, your market, and a fat slice of luck. Anyone who hands you a dollar figure with confidence is selling you the figure.

Speed to first dollar is different. That one is roughly knowable, and I can tell you which streams tend to pay a beginner in a week and which ones make you wait three months on faith.

Speed is what keeps a beginner going. A first payment, even a small one, proves a stranger will pay you for this at all (it feels insane btw!), and that proof is what gets you to the second payment.

None of these twelve is guaranteed. Some readers will earn from them and some will not.

Effort matters and luck matters. I am not going to pretend otherwise to make the list feel warmer.

The four fastest: first dollar in days to weeks

These trade your time for cash. Low ceiling, real money, fastest possible proof that a stranger will pay you.

Start here if you have never earned a euro from AI. Not because they are the best, but because they close the loop fastest. You want the feeling of “someone paid me for this” as early as you can get it. It keeps you going.

1. Freelance task work.

You pick up small, defined jobs on a marketplace: summarize a report, format a document, clean up messy notes. Claude does the heavy lifting, you check and deliver.

✦ The one skill: knowing what “good” looks like so you catch what Claude gets wrong. ✦ The catch: the fees are real. Fiverr takes a flat 20% of every order, and Upwork runs a tiered service fee up to 15%. Pay per task is low and there is a lot of competition.

2. Document and research cleanup.

Someone has a 40-page PDF and no time. You turn it into a clean two-page brief. Claude drafts, you sharpen, they pay for the hour they got back.

✦ The one skill: judgment about what to cut.

✦ The catch: it is one-off work. You are only as booked as your next client.

3. Ghost-drafting a human edits.

You draft emails, bios, LinkedIn posts, or first-pass copy that a real person then owns and edits. You are the fast first draft, not the final voice.

✦ The one skill: getting the draft 80% there so the edit is quick.

✦ The catch: AI slop gets you fired. If you hand over raw output, you are done.

4. Small local-business fixes.

The cafe with the ugly menu. The plumber with no FAQ. The clinic whose cancellation policy reads like a legal threat. You fix one document, in an afternoon, for a flat fee.

✦ The one skill: talking to a non-technical owner without scaring them.

✦ The catch: you have to knock on doors. The work is easy and the finding is the job.

The shared catch on all four: low ceiling, and the money stops the second you do. They are also the fastest proof on this list that you can earn at all, and proof beats a plan.

Subscribe now

The four medium: first dollar in weeks to months

Here you stop selling hours and start selling a thing. This is the jump that raises your rate.

The fast streams cap out because your day has a fixed number of hours in it. The medium streams break that cap, because a thing can be sold more than once, or sold for a price that has nothing to do with how long it took you.

5. A packaged service.

Not “I’ll help with your notes.” Instead: “I turn your monthly meeting notes into a one-page report, delivered the first Monday of every month.” Same work, named and repeatable, sold as a package.

✦ The catch: you have to define it tightly and say no to the jobs that fall outside it.

6. A niche skill file or template.

A SKILL.md file or document template for one specific job one specific group hates doing. Narrow beats broad, so "contract review for property managers" outsells "prompts for business." Since Anthropic shipped "Record a skill" in July 2026, you can build one by recording your screen while you narrate the job, which takes about ten minutes.

✦ The catch: skill files sell for $3 to $10 on Agensi and KissMySkills, so this is pocket money unless you build volume. Building got easy, which is exactly why the file got cheap.

7. A small paid workshop for people who already know you.

One hour, your existing network, teaching the one AI thing you understand cold. You do not need an audience of thousands. You need eight people who trust you.

✦ The catch: you have to be one step ahead of the room, and honest about where your knowledge ends.

8. Productized editing.

You take a client’s raw AI output or rough draft and make it sound human. A fixed scope, a fixed price, a fixed turnaround.

✦ The catch: this only works if your taste is better than the machine’s. That is the whole product.

Share

The four slow-but-durable: months, but they compound

These are slow to a first euro and hard to build. They also build an asset instead of a wage.

An asset is the difference between a job and a business. It keeps earning while you sleep, or it lets you charge for outcomes instead of hours. The wait is real and I will not soften it. Also, the time you invest into building before anybody pays you for it is huge.

9. A small delivery team.

You take the packaged service from stream 5, write down exactly how you deliver it, and train one person to run it without you. You stop being the person who does the work and become the person who owns the thing that does the work.

That written-down process is the asset. It is the same skill file from stream 6, except the buyer is your own hire instead of a stranger on a marketplace.

✦ The catch: you need a service that is already booked and already boring before there is anything worth handing over. The first hire costs money for weeks before they make any, and you have to accept work that is 85% as good as yours.

Consultants who cannot let go of that last 15% stay solo forever, and that is a choice, not a failure.

10. A paid newsletter.

You write, consistently, for people in one niche, and eventually some of them pay for the deeper version. brisk. is mine. It started in January 2026, it publishes every Thursday, and I will tell you plainly: this is a very slow build.

✦ The catch: you write for months to a small room before money shows up. I am not building a list, I am building trust with humans who let me into their inbox, and that takes time you cannot shortcut.

11. A small tool or app.

You use Claude to help build a narrow tool that solves one annoying problem, and you charge a little for it. You do not have to code it perfectly. You have to find the problem.

✦ The catch: building is the easy 20%. Getting anyone to pay and keep paying is the hard 80%, and it is nothing to do with Claude.

12. Forward-deployed consulting.

You embed with one client, learn their actual mess, and use AI to fix real workflows from the inside. The highest-trust, highest-value stream on this list.

✦ The catch: it takes the credibility you build from the eleven streams above it. It is the destination, not the starting line.

The honest trade underneath all four: fast money is time-for-cash and stops when you stop.

Subscribe now

I know that stretch. I quit in it more than once already.

You publish, the room stays small, and the little voice asks whether anyone is even reading. Month two or three is where it gets loudest, which is also where nearly everyone lets go.

Who should skip this list entirely

Some of you should not start any of these right now, and I would rather say that than sell you a stream you will abandon.

If you are hoping AI replaces a full-time income by next month, close this and go breathe. That is the fantasy the gurus sell, and chasing it is how people burn a savings account on a course. None of these twelve does that on that timeline.

If you have zero hours a week to spare, do not start either. Every stream on this list costs time before it pays, and pretending otherwise would make me one of the liars.

And if you have no skill you can name yet, start there instead. Claude is an amplifier, not a substitute. It multiplies what you bring, and it multiplies zero into zero.

The catch that applies to all twelve

AI lowered the floor. That cuts both ways.

The easy streams are crowded, because anyone can start them now. The hard streams are thin, because almost nobody pushes through the boring middle to finish them. The floor came down and the ceiling stayed exactly where it was.

“Using Claude” was never the product. A prompt is not a skill, and a skill file recorded in ten minutes is not a business. The money is in the judgment, the taste, and the trust you bring to the tool.

Here is the trap of a list like this. Reading twelve options and picking none feels like progress, and it is how a week disappears with nothing to show.

One last thing on the numbers. Every fee, cut, and subscription tied to these streams moves, so check the live figure on the platform’s own site the week you start. And treat any income screenshot you see online as fiction until proven otherwise, including the ones with your favorite guru’s face on them.

Pick the single fastest stream that matches a skill you already have, and do the one thing that stream needs this week. One stream, started, beats twelve, bookmarked.

PS - What were your attempts are making money with AI so far? Tell me in the comments!

Leave a comment

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 am writing it because I keep watching professionals hit the same walls, and every wall exists because nobody showed them how to actually use it.

So here are the use cases, and the failures, I have watched pile up over the last few months. Beginner first, then a step up.

Excel is part of my daily job. I am not a heavy user, which is exactly why I never learned the deep formulas or the shortcuts everyone assumes you already know.

I struggled with it. Quietly, for years.

Then Claude for Excel showed up, and now I enjoy the thing. Two reasons. One: I can finally produce what I pictured in my head without knowing the formula to get there. Two: it does the boring data cleanup I used to grind through by hand, because I did not know any better.

💡 Claude can hand you a confidently wrong formula. Speed is the gift. Verifying is the rent you pay for it, and it costs you two minutes.

This one runs long. Grab a drink, open your worst spreadsheet, and follow along.

1 - The thing that actually changed: Claude now lives inside Excel

Here is the shift, and it is the piece nobody explains.

For a long time, using Claude with a spreadsheet meant copying cells out of Excel and pasting them into a chat window. That still works. It is also the smallest way to do this.

Claude now has three different homes when it comes to your spreadsheets, and picking the wrong one is the first failure I watch people make.

A. Claude chat (paste or upload).

This is the fast, no-setup path. You open the Claude app in your browser or on your desktop, and you either paste a few rows straight into the chat, or upload the whole file.

Use it for a quick “fix this one formula” question, or when you are not even in Excel and just want a fast answer. Claude reads pasted tables cleanly, columns and all, and it can hand you an edited file back.

B. Cowork.

Cowork lets Claude work with the actual files sitting in the folders on your computer. It opens your spreadsheet, edits it in place, and saves it back, no uploading and re-downloading.

Use it when the file already lives on your machine and you want it handled, not babysat through a chat window.

C. Claude for Excel (the add-in).

This is the one worth your attention. Claude sits in a sidebar inside Excel itself, reads your live open workbook, and edits it while preserving your formula dependencies.

It cites the exact cells it is talking about, so you can click and jump straight to them. It highlights every change it makes, so nothing happens to your sheet in the dark.

Getting the add-in set up, once:

→ Install “Claude by Anthropic for Excel” from the Microsoft Marketplace.

→ You need a paid Claude plan (Pro, Max, Team, or Enterprise) and a Microsoft 365 work or school account. Personal @outlook.com accounts and old perpetual versions like Excel 2016 or 2019 are not supported.

→ Open Excel, activate the add-in, and sign in with your Claude account. The shortcut to open the sidebar is Ctrl+Alt+C on Windows, Ctrl+Option+C on Mac.

Claude’s spreadsheet features move fast, so the exact button names and menu paths may have shifted by the time you read this [verify current menu path]. Do the setup once yourself this week and take a screenshot of the real screen.

2 - Formulas in plain English (start here)

This is the beginner move, and it is the one that kills the shame fastest.

You do not learn VLOOKUP. You do not learn INDEX MATCH, or nested IF statements, or whatever function they added this year. You describe the outcome you want, and Claude writes the exact formula, then explains what each piece does.

Excel is a language you were never taught. All those symbols and commas and dollar signs are just syntax, the grammar of a language you were expected to pick up on your own.

Try this exact prompt. Swap in your own column letters.

I have a sheet where column A is client name, column B is invoice date, and column C is amount. 
I want column D to show “overdue” if the invoice date is more than 30 days ago, otherwise “ok”. 
Give me the exact formula to paste into D2, and tell me in one line what it does.

Claude gives you the formula for D2, you paste it, you grab the little square at the bottom-right of the cell, and you drag it down the column. Done.

The magic is the second half: “tell me in one line what it does.” That single sentence turns every formula into a mini-lesson.

Here is the failure I watch people make: they take the formula and skip the explanation. So they never see the pattern, and they stay dependent forever, asking for the same thing next month.

Here is a second prompt, for the lookup that has humiliated you before.

I have two sheets. Sheet 1 has an “email” column. 
Sheet 2 has “email” and “company name”. I want Sheet 1 to pull in the matching company name from Sheet 2. 
Give me the exact formula, tell me which cell to put it in, and warn me about anything that could break it.

That last line, “warn me about anything that could break it,” is the difference between a formula that works on your screen and one that survives your actual messy data.

Share

3 - The boring cleanup (this is the part I stopped dreading)

Formulas are the flashy part. Cleanup is the part that used to eat my afternoons.

Real spreadsheets are filthy. Names typed five different ways, dates in three formats, sneaky trailing spaces, duplicate rows, a stray “N/A” that quietly breaks every calculation downstream.

I used to fix all of that by hand. Row by row. Because I did not know there was another way.

Describe the mess. Get the fix.

This column has client names typed five different ways: “ACME”, “Acme Inc”, “acme inc.”, “ACME INC”, “Acme Incorporated”. 
Give me the steps to standardize them all to one clean version, no code, just clicks I can follow.

Ask for “no code, just clicks I can follow” when you want to stay inside Excel and do it yourself. Ask for a formula when you want it automated. Inside the add-in, you can just tell Claude to clean the column in place and watch it highlight every cell it touched.

For the classic date problem, this one:

Column B has dates in mixed formats: some are “05/03/2026”, some are “March 5 2026”, some are just text. 
Show me how to get every value into one consistent date format so I can sort and filter them.

The failure here is subtle: people ask for a formula when they wanted clicks, or clicks when they wanted a repeatable formula. Match the ask to the job. One-time cleanup, ask for clicks. Something you repeat every month, ask for the formula.

4 - A summary and analysis without the panic

Here is where it gets genuinely fun.

You do not have to know which analysis to run. You describe the question in words, and Claude either walks you through the pivot table click by click, or writes the formula that skips the pivot table entirely.

I want to see total revenue by client, broken down by quarter. 
Walk me through building this as a pivot table, one click at a time, assuming I have never made one. 
Then tell me the one formula I could use instead if I want to skip the pivot table.

The real unlock is letting Claude find the analysis you did not think to ask for.

I do this with reports before I present them, and it consistently surfaces one angle I would have missed.

The failure I keep seeing: people use Claude to confirm the answer they already wanted, instead of letting it challenge them. You get a research assistant only if you actually let it think.

Subscribe now

5 - Level up: building a model from scratch

Everything above is beginner ground, and it is where you should start. This is the step up, and it is what changed how I feel about Excel.

The add-in does not just write single formulas. It builds whole models. It links tabs, runs scenario and sensitivity analysis, and updates one assumption without breaking the twenty formulas that depend on it.

Start small: a forecast.

You do not need to be in finance to want this. If you run a business, you want to see where the money goes.

Build me a simple 3-year revenue and cash-flow forecast for a small consulting business. 
Assumptions: I bill roughly [X] per month now, growing [Y]% a year, with monthly costs around [Z]. 
Lay it out month by month, keep the assumptions in their own cells so I can change them, and explain the structure in plain English.

Because the assumptions sit in their own cells, you change one number and the whole forecast updates. That is the moment it stops feeling like a spreadsheet and starts feeling like a tool.

Then the real thing: a 3-statement model.

This is the backbone of serious financial modeling: your profit and loss, your balance sheet, and your cash flow, all linked so they move together.

Build a 3-statement financial model for a small SaaS company: 
income statement, balance sheet, and cash flow, linked so the three connect properly. 
Use placeholder assumptions I can edit, keep drivers on their own tab, and add clear labels so a non-finance person can follow it. 
Explain how the three statements link, one paragraph, no jargon.

Claude understands the structure. It knows the cash flow ties back to the balance sheet, and it keeps those dependencies intact when you change an input. This used to be the work of people who spent years learning it.

Two features that make the models yours:

→ Instructions. In the add-in sidebar, set persistent instructions once, like your formatting conventions, and Claude applies them to every conversation in Excel.

→ Skills. Type “/” in the sidebar to see ready-made moves. Some are Excel-specific, like a cleanup skill; the finance ones can build DCF models and comparable-company analysis on request.

Where to learn the harder models.

I am not going to pretend I can teach you a full valuation model in one section. What I can do is point you at the door.

• Anthropic’s own “Use Claude for Excel” help page has a long list of finance prompts, from 3-statement models to DCF to LBO. Start there.

• The pre-built finance Skills cover discounted cash flow, comparable-company analysis, and more.

• For structured learning, the Corporate Finance Institute has a course on AI financial modeling with Claude and Excel, and there is a free beginner’s guide on GitHub that walks through a three-statement model and DCF.

The failure I watch here is the dangerous one: people treat a model Claude built as finished. They ship it to a client or a board without tracing a single number. Do not be that person. Which brings me to the last section.

Share

6 - The failures, and the one check that prevents most of them

Now the honest part, because I am not here to sell you a fantasy.

Claude can hand you a confidently wrong formula. It will look right. It will sound right. It will run without an error and quietly give you a number that is off, because the logic was subtly wrong or it misread which column meant what.

This is the truth about every AI tool, not just this one. It is a fast, fluent assistant, and fluent is not the same as correct.

So here is the rule, and it takes two minutes.

If row 2 matches your hand calculation, trust the column. If it does not, tell Claude “the result for row 2 is wrong, here is the number I expected and why,” and it will fix the formula.

A few more failures worth naming, because I have watched all of them:

→ Saving before reviewing. The add-in highlights every change it makes. Look at the highlights before you save, every time.

→ Shipping unreviewed. Claude is not built to be your final client deliverable without a human check. Treat its output as a strong first draft, never the signed-off version.

→ Trusting a stranger’s file. Do not run the add-in on spreadsheets from untrusted sources. Hidden instructions can sit inside cells and comments, and a downloaded template or a vendor file can quietly tell Claude to do something you never asked for. Use it on sheets you trust.

Verify one row and you get all the speed with almost none of the risk. Skip the check and you are trusting a stranger’s math on your own numbers.

Prices, plans, and menu paths move, so reconfirm the current details on Anthropic’s site before you rely on them.

Install the sidebar, open the most annoying spreadsheet you own, find the one formula you have been avoiding for weeks, and describe it in plain English. Then check one row.

PS - Let me know what the craziest thing is you build with Claude for Excel.

Leave a comment

Subscribers get my full Installation Guides 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

I stopped listening. The argument misses the point for people like us.

Almost everyone raving about Fable 5 is a developer on the Max plan, building whole apps with it. The rest of us, on free or Pro, burn through our usage by Tuesday and wonder what we did wrong.

I went looking for a way to get the most out of the little Fable 5 I actually have. I found one. It starts by refusing to pick a side.

You do not have to choose Fable or Sol. You can run both. Let the expensive genius plan and check the work, let the cheaper, faster one do the actual building.

That is the whole idea. You get the same result while burning far less of your Fable 5 on it.

It also fixes the quiet mistake almost all of us make: paying a genius to chop onions.

I have been saying the boring version of this since I started brisk.: stop using the most expensive model for everything. Most of your work does not need the smartest brain in the building. It needs a fast pair of hands that follows instructions.

I do not code. I never will. My job for 13 years was translating between the developers who build things and the people who sign off on them.

That is what this is. A developer trick, translated into plain English, so you can decide if it is for you before you touch a single setting.

Now, an honest word before we start.

What this covers

• Why using one AI for everything is the most expensive habit you have

• The two jobs every smart setup splits: the one who plans and the one who does

• What a “token” actually is, in coins, so the savings finally make sense

• The one command that makes an expensive AI run a cheap one for you

• Why the clock is ticking: Fable 5 is only free for a little longer

• Where this breaks, and who should not bother yet

• The exact setup, copy-paste ready, as a free download

This one runs long. Grab a coffee.

1. The most expensive mistake is one AI doing everything

Picture a Michelin-star kitchen.

The head chef is a genius. She designs the dishes, tastes every plate, and sends anything wrong straight back. She is also brutally expensive by the hour.

Now imagine you make that head chef chop all the onions. And peel the potatoes. And wash the pans.

The food is fine. The bill is insane. You are paying genius rates for onion work.

That is the trap. You pick the smartest, priciest model and use it for everything, from the hard planning right down to the boring copy-paste tasks.

It works. It just quietly burns money on jobs that never needed a genius.

Here is the thing nobody tells beginners. There is not one AI. There are tiers.

Some models are big, slow, and brilliant. Some are small, fast, and cheap. Using the brilliant one for a job the cheap one could do is the most common way to overspend on AI.

2. Every good kitchen has two jobs

Split the kitchen into two roles and everything gets cheaper without getting worse.

The head chef. She plans the dish and tastes the result. She does not stand there chopping, she thinks and she judges. You only need her a few times per meal.

The line cook. He does the actual cooking. Fast, cheap, reliable, he makes the dish exactly as written, again and again, for a fraction of the chef’s rate.

Now swap in the AI names.

→ Fable 5 is the head chef. The boss. It is the expensive, deep-thinking model. Its job is to plan the work and review the result. It does not do the grunt work itself.

→ GPT-5.6 Sol is the line cook. The worker. It is the fast, cheaper model. Its job is to take the plan and actually build the thing, then fix it when the boss sends it back.

That is the whole idea. One AI plans and checks. A different, cheaper AI does the work.

The people who build these systems have a name for it: the architect-worker pattern. Fancy words for “the smart one draws the blueprint, the fast one lays the bricks.” You do not need the jargon, just the split.

Share

3. What a “token” actually is, and why the split saves money

Quick detour, because “saves tokens” means nothing until you know what a token is.

A token is a tiny chunk of text. Roughly a word, sometimes part of one. Every time an AI reads your request and writes an answer, it spends tokens, like feeding coins into a machine.

Here is the part that matters. Smart, expensive models charge more coins per word. Fast, cheap models charge fewer.

Same sentence, different price, depending on which model reads and writes it.

The money question is simple. Who is spending the expensive coins, and on what?

If one pricey model does everything, you pay premium coins on every single word, including the boring ones. If you split the work, the expensive model only spends coins on the few things that need real thought. The cheap model handles the mountain of ordinary work at a lower rate.

Look at how the two roles actually get used in a real job:

The saving hides in that “how often” column. You call the expensive brain a handful of times. You call the cheap hands dozens of times.

The bulk of the work runs at the cheap rate, so the total drops.

The builders running this report cutting the bill by around half for similar output. I have not measured that myself, so I am handing you their number, plainly, not dressing it up as mine.

4. The magic word is /route

Here is where it turns from an idea into a thing you can actually run.

The whole setup gets wrapped into one command: /route.

You type it, and the two-model kitchen springs to life. The command runs the chef and the cook, not you.

And here is the part I love. It only runs when you call it. Ordinary questions, everyday chatting, quick edits, none of that triggers the expensive boss.

/route 

sits quietly until you ask for it, so it never burns your premium limits on a normal Tuesday.

Watch one run, in plain English.

→ You type /route and describe the job. Say, “add a check to the signup form so people can’t leave the email blank.”

→ The boss plans. Fable 5 restates the job in one line and writes a short, step-by-step plan. It saves that plan before anything gets built.

→ The worker builds. The plan gets handed to GPT-5.6 Sol, which does the actual work and writes the thing.

→ The boss reviews. Fable 5 reads back what the worker built, hunting for what is wrong, missing, or sloppy. It is the quality check. It does not rubber-stamp the first try.

→ The loop repeats. Anything wrong goes back to the worker to fix. Build, check, fix, again, until the boss is happy. Then you get the finished result.

That back-and-forth is the entire trick. Not “one AI does it and hopes.” A planner and a doer, passing work between them, with a checker in the middle who happens to be the same smart boss who drew the plan.

5. Why this is worth setting up now

Right now, Fable 5 is free to test inside the Claude subscription. But only for a limited window, and that window keeps inching toward the door.

Two things make that urgent.

First, inside the subscription Fable 5 eats your allowance fast. It counts against your limits at roughly double the rate of the older top model. Spend those calls on onion work and you hit the wall by Wednesday.

Second, the free ride ends. When the preview window closes, Fable 5 goes API-only and metered: about $10 per million tokens in, $50 per million out. No free tier.

Read that gap again. Ten dollars to read, fifty to write. The model you were happily throwing every task at becomes the model you count every call of.

That is the whole reason the boss-and-worker split stops being a nice idea and becomes the setup. Let Fable plan and review a handful of times. Let the cheap worker do the hundred small things.

You get the smart model’s judgment without paying smart-model rates on every keystroke.

6. Where this breaks (read this before you get excited)

I am a truth machine, not a hype machine. Here is where this setup falls down.

✦ It is a developer setup. You run it in a terminal, the black screen with the blinking cursor. If “open your terminal” already made you tense, this one is not your Tuesday-morning tool yet.

✦ The names will change. gpt-5.6-sol is the worker model as things stand today. Model names get retired and renamed, so do not be surprised when the label shifts.

✦ The cheap worker is not magic. It gets things wrong. The whole reason the boss reviews is that the worker cannot be trusted on the first pass. Skip the review to save a few more coins and you are back to hoping. Hoping is expensive.

✦ It is manual, on purpose. /route only fires when you call it. That is a feature, not a bug, but it means this is not a robot that runs your whole job while you sleep.

7. Want to actually build it?

I did not print the commands in this article on purpose.

The full setup is short but exact, and exact things belong somewhere you can copy-paste without a typo. That is why I put the whole thing in a guide.

Inside:

→ The four commands that install the piece connecting the two models

→ The one config line that makes the cheap worker the default

→ The complete /route skill file, ready to paste in

→ A demo prompt to run the loop end to end and watch it work

It takes under five minutes if you already have the tools installed, and the guide starts from zero if you do not.

It is free. Subscribe to brisk. and it lands in your inbox.

Subscribe now

The onions, one more time

You would never pay a Michelin chef to peel potatoes. You would have her design the menu and taste the plates, and let the cheap, fast hands do the rest.

Your AI bill works the same way. One brain to plan and judge. One cheaper pair of hands to do the work.

You have been paying genius rates for onion work. Now you do not have to.

Subscribe to brisk., grab the setup guide, and go run your own kitchen.

PS - Let me know how it goes. I encourage you to share ….

Leave a comment

Subscribers get the full Installation Guide incl. Terminal commands and the full /route skill (including my setup explanations for dummies) - just leave a comment and I’ll send it over.

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 have spent 13 years at the line where IT meets business, translating what developers say into something a board can act on. That translation is my whole job, and it is also the whole trick with AI. There are no magic words, only a clear brief.

I do not code. I never have. So everything here is the thing I actually do, not theory I read somewhere.

Here is the part nobody selling a course wants you to hear. A good prompt is just a clear brief. If you can brief a freelancer, you can prompt Claude or ChatGPT tonight.

What this covers

This one is short. Read it once, steal the three prompts at the end, use them today.

1 - Why your answers come out generic

Your answers feel bland because you gave the AI a topic, not a job.

“Write about marketing” is a topic. It has no reader, no goal, no length, no voice. So the AI does the only thing it can: it gives you the most average version of “marketing,” because average is what a topic with no edges produces.

Compare that to a job. “Write a 3-line follow-up email to a client who went quiet after our proposal, warm but not needy.” Now there is a reader, a situation, a tone, and a length. The bland disappears because you removed the blank space it was filling.

Think about how you would hand this to a real assistant. You would never walk over to a competent colleague and say only “marketing” and then judge them for guessing. You would tell them who it is for, what you need, and by when.

The AI is not being lazy when it gives you beige. It is filling the gaps you left, and it fills them with the most common answer it has ever seen. Vague in, average out.

Watch it happen once and you will never forget it. Type “write about time management” into Claude or ChatGPT, read the beige answer, then type “write a 4-line Slack message telling my team we are moving the Tuesday standup to Thursday, casual, no apology.” Same tool, completely different quality.

The AI did not get smarter between those two prompts. You did.

2 - The only prompt shape you need: Role, Task, Context, Format

Everything I write to an AI fits one four-part shape. Role, Task, Context, Format. That is it, and once you see it you will notice it is just how you would brief a competent human.

Role is who it should act as. “You are an experienced management consultant.” This sets the vocabulary and the judgment it pulls from.

Task is the one job. Not three jobs, one. “Rewrite the email below” beats “help me with my emails.”

Context is what it needs to know to not sound like a stranger. Who the client is, what went wrong, what you are trying not to do.

Format is what the answer should look like when it lands. Length, structure, tone, ready-to-send or rough draft.

Here is the whole thing as one copy-paste brief. Steal it.

You are an experienced management consultant. 
Task: rewrite the email below so it sounds warm but direct. 
Context: the client missed our last two calls and I do not want to sound annoyed. 
Format: under 120 words, no jargon, ready to send. [paste your email]

Notice there is nothing clever in there. No secret token, no “act as a world-class expert,” no threat, no bribe. It is four plain lines that tell the AI who to be, what to do, what it is walking into, and what “done” looks like.

Here is the same shape for a completely different job, so you see it is a pattern and not a script. Say you have a wall of workshop notes and forty minutes before a client call.

You are a sharp executive assistant. 
Task: turn the messy notes below into a clean agenda. 
Context: this is for a 30-minute call with a nervous first-time client who wants 
reassurance we have a plan. 
Format: 5 agenda items max, one line each, plus one sentence I can open the call with. 
[paste your notes]

Same four parts, different words. Role, Task, Context, Format is the container, and you just pour the specifics of your actual job into it.

You will not use all four every time. A quick question needs none of this. But the moment an answer disappoints you, the missing part is almost always one of these four, and you can feel which one.

Subscribe now

3 - The one move that beats every “advanced” trick: show an example

If you learn one thing from this piece, learn this. Paste one thing you like and say “match this.”

Every course sells you tricks for describing the tone you want. Punchy. Conversational but authoritative. Warm yet professional. All of it is guessing, because tone is nearly impossible to describe and instant to demonstrate.

So stop describing. Show.

Here is a paragraph I wrote. 
Write the next three in the exact same voice, rhythm, and sentence length:
[paste your paragraph]

The AI now has your voice, your structure, your length, all copied for free. You skipped the entire vocabulary of “prompt engineering” by handing it the target instead of a description of the target.

This works for anything with a shape. Paste an email you liked and say “reply to this new message in that style.” Paste a report section and say “write the next one to match.” Paste a good LinkedIn post and say “same structure, my topic.”

Examples are the cheat that is not cheating. They are how you get a human freelancer to nail your voice too.

4 - Fixing a bad answer without starting over

The first answer is a draft, not a verdict. And beginners quit at the wrong moment, right here.

Here is what newbies do wrong. They get a mediocre answer, sigh, delete everything, and rewrite the whole prompt from scratch. They treat the AI like a vending machine that either gives you the snack or does not.

It is a conversation. You do not re-prompt, you correct.

The corrections are short and blunt, exactly like editing a junior colleague:

• “Shorter. Half the length.”

• “Less formal, drop the corporate tone.”

• “You invented a statistic. Remove it, do not replace it.”

• “Keep the second paragraph, rewrite the rest.”

• “Now make it sound like the example I pasted earlier.”

Each correction keeps everything good about the last answer and fixes one thing. Three of those in a row will get you somewhere no perfect opening prompt ever would, because you are steering with real answers in front of you instead of guessing in advance.

The trick is to change one thing at a time. If you rewrite the tone, the length, and the structure all in one go, you will not know which change fixed it and which broke it. One correction, read the result, next correction.

And you are allowed to be blunt. “No. Try again, shorter.” works fine, and it will not sulk. Directness that would be rude to a person is exactly right here.

The follow-up is the skill. The people who get great results from AI are not writing better first prompts. They are having a second and third exchange while everyone else deletes and restarts.

Share

5 - The 3 prompts to keep in a note this week

You do not need a prompt library with 400 entries. You need three, saved in a note on your phone, ready to paste. Here they are, all built on the same shape from section two.

A. The rewrite prompt. For every email, message, or paragraph that came out clumsy.

Rewrite the text below so it is clear and [warm / direct / formal]. 
Keep my meaning, cut anything that does not earn its place. 
Under [X] words. [paste your text]

B. The summarize prompt. For the long document, the thread, the report you do not have time to read fully.

Summarize the text below in 5 bullets, then tell me the one thing I should not miss. 
Plain English, no jargon. [paste your text]

C. The “match this voice” prompt. The one from section three, because it is the single highest-value move you own.

Here is something I wrote. 
Write [what you need] in the exact same voice and rhythm: [paste your example]

Save those three tonight. That is your entire starter kit, and it will carry a non-technical professional further than any paid course.

6 - The part that outlasts the prompt

Here is the uncomfortable bit. Getting a decent draft out of AI is on its way to free. The 4-part shape is not a secret, and soon everyone in your field will have it.

When output gets cheap, output stops being the thing that sets you apart. A competent email, a passable summary, a fine first draft: the machine hands those to everyone now.

So the value moves. It moves to the three things the machine cannot average into sameness.

✦ Taste. Knowing which of the ten answers is actually good, and why. The prompt gives you options, your judgment picks the right one.

✦ Voice. Sounding like you, not the beige middle of everything the model has ever read. That is why “match this voice” is the highest-value move in this whole piece.

✦ Relationships. The people who open your email because it is from you. No prompt buys that, you earn it over years.

The prompt gets you the draft. You supply the taste, the voice, and the trust. That is the part no course and no model can do for you.

I sold a prompting e-book on Etsy in January 2023, made in Canva, and it sold two copies for about fourteen euros. I was early to a market that did not exist yet. The one thing I got right back then is still the whole point now: clarity beats cleverness, and taste beats both.

Learn the shape tonight. It takes ten minutes. Then put your real hours into the parts that compound: your taste, your voice, and the people who trust you.

Open Claude. Paste one email from today. Give it the 4-part shape, then make it sound like you.

PS - Let me know how it goes. I encourage you to share ….

Leave a comment

Subscribers get all my free guides including my explanations for beginners 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

You’ve seen them, mostly on Linkedin. “Built a full app with one prompt in 2 minutes.” It’s not a lie. It’s just the first frame of a much longer film. Nobody shows you the rest.

I needed a fun idea to prove the point. Then the World Cup handed me one.

Croatia and Bosnia & Herzegovina are both in this tournament. Two small countries from the same region, the Western Balkans, with a combined population of roughly seven million people, both on the world stage at the same time. I’m Croatian. My surname is Tomšić. Yes, that’s an -ić. The whole world has been stumbling over Modrić, Perišić, Blažević for years and I’m right in that club.

I cheer for Croatia because it’s my country and I always will as long as they’re in the tournament. I also cheer for Bosnia as our passionate neighbor, in their first World Cup knockout stage ever. Supporting both is natural when you understand the region. These two countries share a lot: geography, surnames, football culture, and the specific kind of passion that makes the rest of the world slightly nervous.

Bosnia lost to the USA last night. I wish I wouldn’t have to write that. But they were there, and that doesn’t change. This app covers both countries. It stays up.

I don’t code. I never have. That’s not a disclaimer. It’s the whole point of why I’m writing this.

Here’s what I built, how long it actually took, and every single thing that needed fixing before it was worth sharing.

What the app does

The app is called “Find Your Balkan Fan Alter Ego”.

You enter your first name and surname. You pick an avatar style. You click generate. Then Claude decides the rest. Yeah.

You get a Balkan surname ending in -ić, matched to the first letter of your real surname. Smith becomes Šarić. Johnson becomes Jurić. You get randomly assigned to either Croatia or Bosnia & Herzegovina. You do not get to choose. That’s intentional. That’s how being Balkan works.

You also get a fan archetype. One of four:

✦ The Screamer. Hasn’t had a voice since 1998. Worth it.

✦ The Weeper. Cried at the anthem. Goals. Half time. Full time. Just always.

✦ The Tactical Uncle. Never played football. Knows everything about it.

✦ The Pessimist. “Ionako ćemo izgubiti.” We’re going to lose anyway.

All four are randomly assigned. Your avatar appears in that team’s jersey, with your new -ić name on the back and the expression to match.

Try it before you read the rest: https://balkan-alter-ego.vercel.app/

Leave a comment

What this covers

✦ Why you should never open Claude Code first

✦ How to generate your CLAUDE.md in one chat conversation

✦ Why the avatars look the way they do, and what the real vision is

✦ 4 iterations in Claude chat - what broke and how each one got fixed

✦ The move from chat to Claude Code and what happened next

✦ Deployment without touching the terminal

This one runs long. It’s the full behind-the-scenes. Grab a coffee.

1 - Start in Claude chat, not in Claude Code

Claude Code is a building tool. Claude chat is a thinking tool.

Many builders skip this distinction and go straight to the build, then wonder why the first output is a mess. The problem isn’t the tool. The problem is they hadn’t thought the thing through yet.

I started by describing the idea in Claude chat. A World Cup fan generator. Fully free. No external APIs. Deployable without touching the terminal. Claude asked me questions I hadn’t thought to ask: what should the avatar look like, how should the -ić name be generated, can the user choose their country (no, that’s the joke), what happens on mobile.

By the end of that conversation I had a spec I couldn’t have written that well on my own in that time.

2 - Let Claude build your CLAUDE.md

CLAUDE.md is a plain text file that lives in your project folder. Claude Code reads it automatically at the start of every session.

Without it, you re-explain your project every time you open a new session. With it, Claude already knows the stack, the logic, the conventions, and what’s intentionally left out. That last part matters. Scope creep kills amateur builds.

I generated mine from the spec conversation. Claude produced it in under a minute. I checked it, made two small edits, and it was done.

There are also skills you can use before you build. Skills are instruction files that tell Claude how to behave. There are skills that will interview you about your idea and break the build into a structured plan with subtasks before you write a single line of code. Find a lot of pre-defined skills here: brisk.vision/skills

3 - Decide your stack, with help

Non-technical builders either freeze here or pick something they’ll regret three steps later.

Tell Claude your constraints. Mine: free, no APIs, no backend, deployable without the terminal. Ask it to recommend a stack. Push back if you want. I asked why not a simple HTML file. It explained the tradeoffs. I agreed.

For this app: React, Vite, Vercel. In plain English: React builds the interface, Vite turns it into files a browser can read, Vercel hosts it for free and makes it live. That’s it.

Share

4 - A note on the avatar: Claude is not an image generator

Before we get into the build, there’s something you need to understand about what Claude can and cannot do. It directly explains why the first prototype looked the way it did.

Claude can generate SVG. SVG is code that produces shapes, lines, and fills. It’s how Claude can draw an illustrated avatar in a team jersey without any external tools. No API, no cost, no tokens.

Claude cannot generate images. It is not Midjourney. It is not DALL·E. It cannot produce a photorealistic photo or a detailed cartoon illustration from a text prompt.

This was a deliberate choice for this build. Connecting an image-generating API costs money. Every generated image costs tokens. I wanted to keep this build completely free, and SVG delivered that.

My actual vision for this app is different. The two cartoon fan illustrations you see on the site - those I created with GPT Imagen 2. That’s what a real user-facing experience would look like: you enter your name, the app calls an image generation API, and you get a personalised cartoon fan in your team’s jersey. Something like Nano Banana or the GPT images API would do this.

That version isn’t free. So that version isn’t this build.

The SVG avatars are the free version. They do the job. And they explain why the first iteration came out the way it did.

Subscribe now

5 - The build: four iterations in Claude chat

The entire first phase of this build happened in Claude chat, not in Claude Code.

Claude generates working React components as interactive artifacts right inside the chat window. I could see the app, click buttons, type a name, get a result. I iterated in chat until the logic and visual direction were close to right. The four iterations below all happened there.

The first prototype

The logic worked. The -ić generator worked. The country assignment worked. The four archetypes were there.

Everything else needed fixing. The avatar arms looked like floating rectangles with no connection to the jersey body. The jersey name was invisible against the red-and-white Croatian checkerboard. All UI text was dark grey on a near-black background. The fan illustrations I uploaded had white backgrounds instead of transparent ones.

None of this is a failure. This is what a first prototype is supposed to be: proof that the logic works. If you expect iteration 1 to be the final version, you will always be disappointed.

The iteration rule

I work to a maximum of two problems per prompt. Never more than two in one message.

When you stack too many fixes, the output gets unpredictable, and when something breaks you won’t know which change caused it. Name the problem specifically. Don’t re-explain the whole app. Trust Claude to hold context within the session.

Iteration 1 — The avatar arms

The arms were separate SVG rectangle elements that floated beside the jersey with no connection to it. This is the SVG constraint in practice: Claude can draw shapes, but getting them to look naturally connected takes iteration. A real image-generating model would just draw a jersey. SVG requires you to describe every element precisely.

One prompt: redesign the jersey as a single path that includes the sleeve shapes as part of the same outline. The floating rectangles were gone. The jersey became one connected shape (yes I know, still doesn’t look great, but for a free svg version for demonstration purposes it is good enough).

Iteration 2 — Jersey name visibility

The -ić name on the jersey was white text on a red-and-white checkerboard. Invisible (see Screenshot above). One prompt: add a white name bar behind the text. Croatia gets blue lettering, which matches their actual jerseys. Bosnia gets navy. A small detail that makes the thing feel considered rather than slapped together.

Iteration 3 — Image transparency

I uploaded the GPT Imagen 2 illustrations as PNG files with transparent backgrounds, designed to float naturally on the dark app background. When Claude processed them, it converted the PNGs to JPEG, which strips transparency and adds a white background.

One prompt: keep the files as PNG. Transparent backgrounds restored.

Iteration 4 — Font colours

All UI text was muted grey on near-black. Hard to read. One prompt: make all text white. Single exception: the jersey name bar, which has a white background and coloured text for contrast. Two small fixes in the same prompt: change image labels from “HRVATSKA” and “BOSNA” to “Croatia” and “BiH.”

Subscribe now

6 - From Claude chat to Claude Code

Once I was happy with the prototype in chat, I saved the artifact file to a folder on my desktop.

I opened that folder in VS Code. If you haven’t set up VS Code yet, I have a step-by-step guide here:

With Claude Code installed, it read the CLAUDE.md and immediately understood the project. It knew what had been built, what the conventions were, what still needed doing.

That handoff took about 30 seconds.

7 - GitHub: version control in one step

Before deploying, I connected the project to GitHub.

GitHub saves every version of your code so you can roll back if something breaks. It also connects to Vercel so deployments happen automatically when you push changes. I have a setup guide here:

Claude Code handles the connection once it’s in place.

8 - Deployment and what broke after

I had Claude Code deploy the app to Vercel directly from VS Code. No terminal needed.

If you set up Vercel first using this guide, Claude Code can manage all future deployments for you the same way. You can also deploy manually from the terminal with vercel --prod if you prefer that route, but if the integration is in place, there’s no reason to.

The app was live. Then I actually tested it properly.

Iteration 5 — The share button (Claude Code)

This is the iteration you only discover after deployment, because you need a real URL to test it.

The share button was supposed to let users copy a link and share their generated fan image. Copying the text worked fine. Sharing the actual generated avatar image did not. The image lived inside the app as an SVG element, and getting that into a shareable format - a downloadable image, a proper Open Graph preview, something a user could drop into a WhatsApp chat - requires a different approach than just copying a URL.

I worked through this in Claude Code, iterating on the share logic. This is the kind of problem that only surfaces in a real environment, with a real URL, on a real device. It’s also the kind of problem that makes you realise why “2 minutes to a finished app” is the wrong framing. You don’t know what’s broken until real users can touch it.

9 - Now go find out which fan you are

Bosnia lost. But they were there for the first time in history, and that doesn’t change because of a scoreline. Croatia are still in it. The app covers both.

If you’re a football fan who needs a new team and a new identity, this is where you start 🤗.

Try the app: https://balkan-alter-ego.vercel.app/

Then come back here and tell me which fan type you got. Reply with your -ić name and your archetype. Are you a Screamer? A Weeper? The Pessimist who already knew how it was going to go?

The app took 30 minutes to build. Not because I’m fast or technically skilled. Because I had a spec, I used Claude chat before Claude Code, and I fixed one or two things at a time.

That sequence is the entire skill.

PS - Let me know how it goes. I would to see your -ić name and fan-type in the comments, and have a laugh!

Leave a comment

Subscribers get exclusive access to all my resources and chat anytime.

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

Not followers. Not page views. Citations. As in: AI tools are citing brisk.vision as a source when generating answers to people’s questions.

I have the screenshot. Most websites have zero.

I launched brisk.vision in April 2026 - and built it with one rule: be findable by AI from day one, not as an afterthought. Two months later, Microsoft Copilots and Partners have cited it 392 times. That number only covers Microsoft’s ecosystem - ChatGPT, Claude, Perplexity, and Gemini aren’t tracked here, so the real total is higher. I also have a slightly older website, re-built the same way, past 600 citations.

I’ll show you exactly how I did it. The method is at the bottom. But first, I need to explain what’s actually happening to search right now - because without that context, the method won’t stick.

What this covers

• Why your website is probably invisible to AI right now

• SEO, GEO, AEO: what each one actually means in plain English

• What’s happening to Google search in 2026

• The AI slop trap and why it always backfires

• Why the first-mover window is still open (for now)

• Exactly how I got 392 citations on a brand-new website in 2 months

• How to check your own AI citations for free, starting tonight

This one runs about 3,000 words. Worth it.

The game changed. Nobody sent you the memo.

For 25 years, the internet ran on one rule.

Write for Google’s algorithm. Get on the list. People click.

That rule still works. It’s just no longer the whole game.

AI Overviews. ChatGPT Search. Perplexity. Microsoft Copilot. These tools don’t show people a list of links. They give them an answer. Synthesized. Confident. Citing three or four sources.

If you’re not one of those sources, you don’t exist in that search.

Here’s the number that shifts everything: only 12% of URLs cited by AI tools like ChatGPT and Perplexity rank in Google’s top 10. That’s from Ahrefs research in August 2025. Traditional SEO does not guarantee AI visibility. You can rank first on Google and be completely invisible to the AI tools your clients are using every day.

Anyone building a professional presence online operates on reputation. Clients, employers, partners - they research you. That research increasingly happens through AI tools, not through a Google search. If AI doesn’t know you exist, you don’t exist in that search.

SEO, GEO, AEO: three acronyms in 60 seconds

This is the part where non-technical people usually check out. Stay with me for 60 seconds.

SEO (Search Engine Optimization) has been around since the late 90s. Write content that satisfies Google’s algorithm. Include keywords. Rank higher. Get clicked. You’ve heard of it.

GEO (Generative Engine Optimization) is new. It means writing so that generative AI engines - ChatGPT, Perplexity, Gemini, Copilot - cite your content when generating answers. Most non-technical professionals have never heard this term. Most never will until it’s too late.

AEO (Answer Engine Optimization) means structuring your content as direct, extractable answers. FAQ blocks. Clear definitions. Specific questions answered confidently. The goal is to become the source AI pulls when someone asks a question in your field.

The only distinction that matters: SEO gets you on the list. GEO and AEO get you into the answer.

These aren’t competing strategies. They’re layers. You build them together. In 2026, most websites are still only playing the first one.

Share

What’s actually happening to Google

Google didn’t get replaced. It evolved in a direction that hurts unprepared content creators.

AI Overviews now appear at the top of most Google searches. Not a list of links. A synthesized paragraph with two or three citations. Users read it. Then they leave. They never scroll to the organic results below.

The click-through rate on traditional organic search has dropped sharply since AI Overviews became the default. People get the answer in the overview. They don’t need to click.

The other shift: a growing number of users skip Google entirely and go straight to Perplexity or ChatGPT.

For anyone with a website, this matters for one specific reason. The people deciding whether to hire you, read your content, or follow you are doing research this way. Not in a few years. Right now.

The AI slop trap (and why it always backfires)

Here’s what a lot of website owners started doing in 2024 and 2025.

Flood the blog with AI-generated posts. Fast. Cheap. Fifty articles a month. Publish everything. Watch the traffic spike.

Then watch it collapse.

Google’s E-E-A-T system - Experience, Expertise, Authoritativeness, Trustworthiness - was built to detect exactly this. Google’s core updates throughout 2024 and 2025 specifically targeted content farms: websites flooding the index with thin, generated content with no human judgment behind it. The traffic spike lasts weeks. The penalty lasts months. Sometimes longer.

There’s a second problem nobody talks about.

AI slop doesn’t get cited by AI.

Generative AI engines pull from sources they’ve identified as trustworthy, authoritative, and clear. Content that reads as thin, repetitive, or hedged gets skipped. The consultants flooding their blogs with AI-generated articles are building websites that both Google and AI tools will systematically ignore.

For coaches, freelancers, consultants, solopreneurs - anyone whose name is their business - the reputational risk is separate from the SEO risk entirely. Content that reads like a machine wrote it without a human thinking behind it costs you credibility with the readers who actually matter.

Subscribe now

The first-mover window (and why it’s closing)

Most non-technical professionals have never heard of GEO or AEO.

The tools for measuring AI citations are still in beta. Best practices are still forming. The published research on what AI engines prioritize is 6 to 12 months old at most.

This is what early SEO looked like in 1999.

Chaotic. Experimental. Full of first-mover advantages for anyone willing to do the work while everyone else was still calling websites “home pages.” The businesses and consultants who built proper websites in 2000 and 2001 didn’t stop benefiting from that advantage until well into the 2010s.

By the time GEO and AEO become common knowledge - by the time there are courses and LinkedIn posts about “10 GEO hacks for managers” - the sites that built properly will already own the citations. Every AI tool recommending content on a topic builds a pool of trusted sources over time. Getting added early is significantly easier than competing to displace a source that’s been cited thousands of times.

It will take years for AI search to find its equilibrium. Which means the investment you make now compounds in ways that will be much harder to replicate in two years.

This is the LinkedIn newsletter moment of 2015. The professionals who started then own that space now. The ones who waited are still trying to break through.

How I got 392 citations in 2 months (the actual method)

Here’s what I actually did.

brisk.vision launched in April 2026. It’s my brand hub: where I publish an optimized version of everything I write here on Substack, plus a growing resource directory for non-technical professionals who want to use AI without learning to code.

I didn’t build it and then think “maybe I should optimize it for AI.” I built it optimized. From the first page.

A. An extensive FAQ section on the services page.

brisk.vision/services has a structured FAQ that covers the exact questions prospects and AI engines ask about the business. “What does brisk. do?” “What AI tools does brisk. use?” “What industries does it work in?” Each question gets a direct, complete, confident answer.

This is AEO at its most fundamental. AI engines can extract a clean answer and cite the source without needing to read the surrounding context. The services page FAQ is responsible for a significant share of those 392 citations.

B. Every Substack article gets rewritten before it goes on the blog.

Not reformatted. Not copy-pasted with new headers. Rewritten.

The original Substack articles I publish (you are reading one right now) are written for humans - for the way readers follow a narrative. The blog version on my website is optimized for how AI reads. The key insight: research from Growth Memo (February 2026) found that 44.2% of all LLM citations come from the first 30% of an article. Bury the conclusion at the bottom, and AI ignores it. Lead with the conclusion, and AI cites it.

So key conclusions get moved to the top. Headers get restructured for clarity. Hedging gets removed entirely. Every article on brisk.vision opens with its most important point.

I run my aeo-geo-seo-optimizer skill on every piece before it goes live. It does a full GEO and AEO audit: checks structure, rewrites hedged language, surfaces the key answers that should lead each section. That skill is the reason the citations are consistent and growing, not occasional and random.

Want the skill? Comment below and I’ll share it with you directly.

Leave a comment

C. I write definitively.

No hedging. No “I think” or “perhaps” or “it could be argued.” AI engines skip uncertain content. Every claim on brisk.vision is stated with confidence because I only publish things I’ve tested and know to be true.

The result after 2 months: 392 citations from Microsoft Copilots and Partners alone. That’s one citation source - the one tracked by Bing Webmaster Tools. Perplexity, ChatGPT, and Gemini aren’t included in that number, which means the actual total is higher.

I also have an older website, I rebuilt it the same way before I formalized the method into a skill. That site went from 0 to past 600 citations.

Same method. More time.

How to check your own AI citations for free

Start here: Bing Webmaster Tools.

It’s free. Go to bing.com/webmasters, add your website, and verify ownership. Then look for the “AI Performance” tab in the left sidebar - the one you can see in my screenshot at the top of this article. It’s currently in beta but it works. You’ll see exactly how many times Microsoft Copilots and Partners have cited your content.

For the other major AI engines, the manual check is the most direct method. Open Perplexity, ChatGPT with browsing on, and Gemini. Ask a question your ideal client would ask about your field. Check whether your name, your website, or your content appears in the answer or in the cited sources.

Run this check with 3 to 5 different phrasings of the same question. AI tools can source differently based on how a question is worded.

If you want ongoing monitoring without checking manually: Otterly.ai tracks AI mentions across platforms and has a free tier. LLMRefs scans multiple engines at once.

The window is open

brisk.vision launched in April. By June it had 392 AI citations.

I built it to be found. Not just by Google. By whatever comes after Google.

Professionals with a website are not wrong to focus on their clients and their work. But AI is already recommending professionals in your field. Every day you operate without an optimized presence is a day a competitor shows up in that recommendation instead of you.

Start with Bing Webmaster Tools tonight. Check if you’re visible. If you’re not - comment below and I’ll share the skill.

PS - Let me know how it goes. I encourage you to share ….

Leave a comment

Subscribers get all my installation guides for free 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

So I stopped writing them.

I built a tiny newsroom of 8 AI agents to do it instead. Last Wednesday it drafted a full week of notes before I woke up.

It runs on its own every Wednesday at 6pm, the night before I publish my article. Eight agents, one job each. It costs about $6 a month.

And nothing goes out without me. I read and edit every note before it publishes. The system does the heavy lifting. I keep the voice.

What this covers

• Why 35 notes a week broke me (and why a scheduler alone didn’t fix it)

• The newsroom: what each of the 8 agents actually does

• The Writer, taken apart: how one note gets built from nothing

• The Retro Analyst (the coach), taken apart: how the system learns to sound like me

• What I automated, and what stays mine on purpose

• What it costs and what it runs on

• How to get the scheduler and build your own

This one runs long. It’s the full behind-the-scenes, including two of the agents pulled apart piece by piece. Grab a coffee.

1. Why 35 notes a week broke me

Every growth guide says the same thing. “Post Substack notes consistently.” True. And useless. Because consistency is the entire problem.

The bottleneck was never ideas. It was never writing one note. It was the volume, the rhythm, and the timing. Five notes a day, seven days a week. That is 35 notes, on top of the actual newsletter, client work, and a life.

I did the obvious thing first. I solved the scheduling.

Substack already has a scheduler. The catch: you feed it one note at a time. Paste the text, save a draft, pick a time, then do it all again.

For 35 notes a week, that is soul-destroying. I was too lazy to do it 35 times, so I built a small tool that bulk-schedules from a CSV.

One upload, and it creates all 35 drafts and their posting times at once, using Substack’s own scheduler underneath. That solved the busywork. It did nothing about what to write, or whether what I wrote was any good.

Scheduling an empty calendar is still an empty calendar. I needed something that filled it, intelligently, in my voice. That is why the engine exists. The scheduler is the hands. The engine is the brain.

2. The growth engine: 8 agents across 3 phases

Here is the simplest way I can explain it.

If I had hired humans to run my Substack notes, I would have needed a small team. Researchers, a strategist, a writer, a coach. I did not hire people. I wrote eight prompts instead, gave each one a single job, and called the whole thing the growth engine.

It works in three phases.

Phase 1: Parallel research (five agents at once)

Five agents act like a newsroom, all working at the same time, each pulling intelligence from a different place.

• The Past-Performance Analyst reads last week’s actual numbers. Likes, restacks, replies. It says what worked and what flopped, with real figures, not vibes.

• The Article Analyst reads my new article for the week and pulls out 10 to 14 angles worth turning into notes.

• The Evergreen Analyst mines everything I have ever published for timeless ideas, so the week is not all reactions to current news.

• The Competitor Analyst studies the top creators in my niche and spots the patterns worth adapting. Adapt, never copy.

• The Audience Analyst reads my engagement heatmaps to find the best times to post.

Phase 2: Strategy and writing (two agents, in order)

Now the research hands off to two agents who work one after the other.

• The Synthesizer (the strategist) turns all those notes into one weekly game plan, built on a fixed content mix: 40% from the new article, 30% evergreen, 20% competitor-inspired, 10% pure brand voice.

• The Writer takes that plan and drafts all 35 notes in my voice, each one scheduled for the right date.

Phase 3: Continuous improvement (one agent, on a loop)

• The Retro Analyst (the coach) runs after the week is over. It reviews how every note performed and writes the instructions for next week’s team: fix the misses, double down on the wins.

That last one is the difference between a tool and a system. The first seven do the work. The eighth makes next week’s work better.

Two of these are worth taking apart, because they are where the real work happens. The Writer, which builds the notes. And the Retro Analyst, the coach, which makes next week’s notes better than this week’s. Let me open them up.

3. The Writer, taken apart

The Writer is the agent people assume is the whole system. It is not. But it is the one that turns a plan into 35 finished notes, so let me show you exactly how it does that.

What it reads before it writes a single word

The Writer does not start from a blank page either. Before it writes, it is handed a stack of context. The week’s brief from the Synthesizer. The exact posting times from the Audience Analyst. The angles pulled from this week’s article. The evergreen angles from my archive. The competitor patterns to adapt. My full voice file. My “about me” with my real stories in it.

That last part matters. The writer is not allowed to invent my life. If it writes a personal story, it has to pull from real details I have written down. Real numbers, real places, real moments.

The seven shapes a note can take

A note is not just “a short post.” For months I have tracked the notes that actually land for creators in my niche, and the winners almost always fall into one of seven shapes. The writer works from those seven fixed formats, and every note has to be one of them.

• Imperative Command. Do this. A direct instruction.

• Bold Claim plus Reality Check. A strong statement, then the twist.

• Definitional Framework. Name a thing, define it in a fresh way.

• The Confession. A vulnerable admission that earns trust.

• Quick Tip. One small, usable thing.

• Ultra-Short Truth. Under 50 words. A single clean observation.

• Data Truth. A number that changes how you see something.

The point of fixed formats is rhythm. A feed of 35 notes that are all “bold claims” reads like a robot shouting. A mix reads like a person.

The house rules it cannot break

Here is the part I like most. The writer is forced to live by the same rules I do.

• No two notes in a row use the same format.

• No two notes open with the same first line.

• The first sentence is always the hook. No warm-up.

• Specific numbers always. “884 subscribers,” never “almost 900.”

• No emojis in the notes. No em dashes. No closing call to action.

• A set mix of lengths every week. At least a few under 50 words. A stack of short ones. A handful of longer ones.

These are not suggestions in the prompt. They are hard constraints. If the writer drifts, the rules pull it back.

Watch one note get built

Say the week’s plan calls for a personal story note tied to courage.

The writer opens my “about me,” finds a real moment, and builds the note in four beats. The setup. The moment. What it taught me. Then the line that hands the lesson to the reader.

So it might pull the time I landed in Costa Rica at 19, alone, unable to count to three in Spanish. Setup. The moment I realized no one was coming to translate for me. The lesson, that being lost in public is the price of learning anything fast. Then the turn to the reader: if you are four months into something new and crying, that is the part where it works.

It writes that as one tight note, around 150 words, in the Confession format, and stamps it with a posting time. Then it does that 34 more times, in different shapes, on different angles, never repeating an opener

.That is the Writer. A very fast, very disciplined drafter that never gets tired and never breaks the house style. What it cannot do is know whether any of it actually worked. That is the next agent’s job.

4. The Retro Analyst, taken apart (the coach)

This is the agent I did not expect to love. The Retro Analyst, the coach, is what turns a clever drafting tool into a system that gets better every week.

It runs after the week is over, once the notes have collected real engagement. And it reads three things most people never think to feed back in.

One: the numbers, scored the right way

The Coach pulls the actual likes, restacks, and replies for every note. But it does not just add them up. It scores each kind of note on the thing that note was trying to do.

A story note is judged on likes and replies, because its job is to make people feel something. An identity note, the kind that says “this group deserves more visibility,” is judged on restacks, because its job is to be amplified. A question note is judged on replies, because its job is to start a conversation.

So the Coach writes a scorecard. Best note on likes. Best on restacks. Best on replies. Often three different notes. Then a verdict per format: pulled its weight, or did not, in plain numbers.

Two: my edits, which are the real gold

Here is the part that surprised me.

The Coach reads the edits I made by hand. Not the notes I approved as-is. The ones I rewrote before publishing.

It treats every edit as a correction. It compares what the writer drafted to what I actually published, and it looks for the pattern. “She cut the throat-clearing opener from three notes.” “She removed the closing call to action every time.” “She shortened a two-line hook to one line.”

Those patterns get written into a running voice file. A flat list of rules. Next week, the Writer reads that file before it drafts anything.

The first few weeks, I rewrote a lot. The notes were fine, but they were not mine. They had that faint smell of AI trying hard. Now I rewrite less. Not because of magic. Because of a loop that compounds, one corrected note at a time.

The voice file is capped, on purpose. It holds the strongest rules, not every passing whim. If a new correction contradicts an old one, my latest behavior wins and the old rule is dropped. It is a living style guide, written by my own red pen.

Three: a plan for next week

The Coach does not just grade the past. It writes hypotheses for the future, and it states them like bets.

“If we run two identity notes next week instead of one, expect weekly restacks to grow about 80 percent, based on this week’s numbers.” “If we move the identity note to Tuesday morning, expect more restacks, based on the Tuesday peak.” Then it hands the next Synthesizer a short list of changes to make.

And because I asked it to be a truth machine, not a hype machine, the Coach is blunt. If a format flopped, it says so in numbers and tells next week to do less of it. It is the colleague who actually remembers your feedback, and is not afraid to tell you the thing did not work.

5. What I automated, and what stays mine

I stole my golden rule from a creator I read, from AI blew my mind. Automate the work you dread, not the work that is you.

Here is where I drew the line.

Automated: the research, the timing, the first draft of all 35 notes, and the weekly retro. The heavy, repetitive, soul-draining volume. The stuff I would never get to on a busy week.

Stays mine: the final edit, the approve button, and the voice. Nothing publishes without me reading it. The newsletter article itself is never touched by this system. That is the part I love, so I keep it.

And because I am a truth machine, here is where it still fails.

• It over-rates its own work. The Coach’s grades are a starting list, not a verdict. I still decide what was actually good.

• It drifts toward generic when the week’s article is thin. Weak input, weak notes. Garbage in, garbage out is not solved by AI.

• The timing is a guess. Substack does not tell you when your readers are actually online. So the Audience Analyst uses my own past engagement as a proxy. Useful, but not the same as truth.

None of these are dealbreakers. They are just the honest shape of the thing. Anyone who tells you their AI system runs itself perfectly is selling you something.

6. What it costs and what it runs on

You do not need to be technical to follow this part. Promise.

The whole thing runs in the cloud on a schedule. It fires itself every Wednesday at 6pm, drafts the week, and emails me when the drafts are ready. My laptop can be closed. I can be asleep, or out, or on a train.

It is built on tools that are free or nearly free. A small web app on Vercel’s free hobby plan for the dashboard. A free Supabase database to remember everything. Claude doing all the thinking. Add it up and it costs about $6 a month. That is roughly $1.50 a run, once a week.

When I open the dashboard, I read the brief, edit the notes I want to change, and click one button. Approve and push to scheduler.

The engine hands the finished list to my bulk-scheduler as a CSV. The tool creates all 35 drafts in Substack at their scheduled times in one go. No more pasting a note, saving a draft, and setting a time, 35 separate times.

7. Build your own

Here is the honest path, in the order I would do it again.

Start with the scheduler. It is the simplest piece and it solves the most annoying problem first, the timing. Get comfortable handing it a list of notes and watching them go out on their own.

Then add a writer. One prompt, one good voice file, and you have a first draft instead of a blank page. That alone changes your week.

Then layer the rest of the team as you learn what you actually want. The analyst. The scout. The coach. You do not build all eight on day one. I did not either.

The three files that do 80 percent of the work are not code. They are your voice profile, your “about me,” and your list of writing habits to avoid. The agents are only ever as good as those three files. Spend your time there.

The deal. I am giving the scheduler template away. Recommend the brisk. newsletter, and the scheduler repo is yours. The code, the setup steps, all of it.

I am keeping the full engine private, because it is wired so tightly to my voice and my data that it would not help you as-is. But if you want to build your own version of the eight-agent engine, reply to this email and tell me. If enough of you are interested, I will write up exactly how I set mine up.

And if you would rather watch me build it, I recorded a full walkthrough with where I share my screen and go through the whole setup, step by step. Watch it here:

One month ago, the notes were the part I dreaded most. Now they are the part I think about least.

You can keep writing 35 a week by hand. Or you can spend one weekend teaching a system to draft them, and get your Wednesdays back.

Recommend brisk., grab the scheduler, and go build your newsroom.

PS - Let me know how it goes. I encourage you to share ….

Leave a comment

If you are interested in seeing my growth engine repo on GitHub, please reach out to me in the comments. If enough people do so, I will make it a public template.

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

TL;DR Claude doesn’t remember you between chats. Skills fix that - teach it once, it knows every time.

Three ways to build one: from scratch, chat-then-package using skill-creator, or screen-record your workflow and let Claude analyze it (works for whole teams too).

Warning: Don’t download skills from strangers. One exception: official tools from verified publishers like Anthropic. Curated list at brisk.vision/skills.

Last week: the spectrum. Prompt → skill → scheduled workflow. Skills are the lowest-friction entry point. No Pro subscription, no desktop app, just Claude chat.

This week: what to actually do with that.

1 - Claude doesn’t know you.

Not a complaint. A fact worth sitting with.

Every time you open a new chat, Claude starts from scratch. It doesn’t know your writing style. It doesn’t know your brand colors, your format preferences, your constraints, or how you’ve told it to structure things a hundred times before. You either re-explain all of that, or you accept output that doesn’t quite fit.

Most people compensate without realizing it. They keep a notes document with their “Claude context” that they paste at the start of each session. They retype the same instructions in slightly different words every time. They get inconsistent output and blame the tool, when the real issue is they’re asking Claude to remember something it has no mechanism to remember.

This is the problem skills fix. You teach Claude who you are once. It knows from then on.

2 - What a skill is. And what it isn’t.

A skill is a saved instruction set that lives inside Claude. When you activate one, Claude already knows the job before you type a word: the format, the context, the constraints, everything you’d otherwise have to explain. You stop reconstructing. You just start.

Skills live in Claude chat. No Pro subscription. No desktop app.

A skill is not a scheduled workflow. That’s Cowork territory, and it requires different setup. A skill doesn’t run automatically. You activate it when you need it, and it knows what to do.

When to build one: → You do the same type of task more than once a week → You’re re-explaining context that never changes → You want consistent output every time, not output that varies based on how well you remembered to explain things that day

When not to bother: → It’s a one-off task → A single line prompt already handles it → You need it to run on a schedule

On downloading skills from the internet: Don’t. Not from strangers. A skill built by someone else was built for someone else’s workflow. It encodes their context, not yours. Generic output is the best you’ll get from it.

There is one exception: official skills from verified sources. Anthropic publishes skills for specific technical tasks - creating Word documents, PowerPoint decks, Excel files, PDFs. These are technical reference tools that tell Claude how to use a specific library correctly. The people who built Claude wrote them. Same goes for Microsoft’s Azure skills, Supabase’s database skills, shadcn’s component library. When the skill is about the tool and the tool’s team built it, that’s worth installing.

I’ve curated the ones worth knowing at brisk.vision/skills.

- 51 skills across 8 categories, filtered for quality and source.

3 - How to build one.

Two official methods exist. I use one of them. And will teach you a third.

Method 1: From scratch. Open Claude chat. Find the Skills section in the sidebar. Click Add new skill. Write the instruction set directly: what Claude should know, what it should always do, what format to use, what to never do. Name it. Save it.

This works when you already know exactly what the skill needs to say. If you’re not sure (and most of the time you won’t be), use Method 2.

Method 2: Chat → get it right → package it.

This is the method I use. The idea: prove the skill works before you save it.

Before you start, check that skill-creator is active. Anthropic ships it pre-installed in Claude Desktop and Cowork. It’s already there when you open the app. In Claude chat on the web, check under Customize > Skills. Skill-creator is what turns your conversation into a properly structured, installable skill file. Without it, you’re formatting that file manually.

The process:

1. Open Claude. Do the task you want to systematize, normally.

2. Iterate until the output is genuinely good. Not fine. Good enough to use.

3. Tell Claude: “Use skill-creator to package everything we’ve worked out as a reusable skill.”

4. Skill-creator generates a complete skill: trigger description, instruction set, example outputs, correct structure throughout.

5. Ask Claude to save the skill directly to your account. This skips downloading a file, renaming it, zipping it, and uploading it manually.

6. Review the saved skill. Adjust anything that’s off.

The main mistake: packaging before the output is actually good. The conversation is the draft. The skill is the published version. Don’t publish a draft.

Method 3: Record → analyze → package.

Most people hit a wall before they even start: they don’t know how to explain the skill. They do the task by feel, not by steps they could put into words.

Here’s what to do instead. Record your screen while you do the task. Save it. Upload the recording to Claude and ask: “Watch this and describe exactly what I’m doing, step by step.” Claude produces a summary of your own workflow. You proofread it, adjust anything that’s off, then use that as the instruction set - and package it as a skill.

This works especially well in a company setting. Instead of trying to write down a team workflow in a document nobody reads, record someone doing it. One screen recording. Claude turns it into a replicable instruction set. You package it as a skill. Now the whole team has access to a process that used to live in one person’s head.

This is how you scale what one person knows to an entire organization. Without a training session. Without a manual. Without anyone having to articulate from scratch what they can’t easily put into words.

4 - What I taught Claude once.

The brisk Instagram carousel skill is the clearest example I have of what this looks like in practice.

Before it existed: every carousel started with me pasting roughly 300 words of context. Brand colors, hex codes, font rules, logo variants per background type, the seven slide roles, overlay opacity logic for dark and light and teal backgrounds, image matching rules, caption format, the non-negotiables checklist. All of it. Every chat. Every carousel.

The skill was built because of one specific moment: the first carousel that came out good enough to post under my name. Not fine. Actually good, by my own standard. That moment made it clear: this process needed to be repeatable. I couldn’t re-explain 300 words every time.

I built v1 using Method 2. How long did it take? About 30 minutes, not 15. The reason: I gave Claude the brand brief (colors, fonts) but not a design example. Without something visual to reference, Claude approximated the aesthetic and I iterated more to get the look right. If I had started with an example slide alongside the brief, it would have been faster.

That’s the thing about Method 2 that matters: time scales with how clearly you can picture the output before the conversation starts. Writing tasks, analysis tasks, structured formats - 15 minutes is realistic. Anything where you’re discovering what “right” looks like as you go: add time for that discovery, because it’s going to happen regardless.

The skill is now on v3. The gap between v1 and v3 was real carousels, specific friction, specific fixes. Text on bright backgrounds was unreadable. The v3 solution was auto dark/light text selection based on composite brightness. Now the skill handles it without manual input. Version 1 worked. Version 3 works better. The iterations come from using it, not from planning it.

If you want to learn how to create your own Carousel skill, I have a whole article on that:

That’s what you’re building. Not the finished version. The first version that’s good enough to use. Then you improve it from there.

Claude doesn’t know you yet.

Teach it once. Open Claude chat. Think of the task you’ve re-explained more than three times. Start the conversation. Get the output right. Ask skill-creator to package it.

If you can’t easily explain the steps, record your screen first. Let Claude watch and describe. Then package what it gives you.

That’s the whole method. The skill does the rest every time you use it.

PS - Let me know how it goes. I encourage you to share ….

Leave a comment

Subscribers get all my special resources for free in the subscriber chat.

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

That’s prompting. It works. I did it for months. I still do it for the fast stuff.

But there’s a ceiling to it, and at some point you hit it. Not with a crash. More like a slow friction. You realize you’ve explained your writing style to Claude eleven times. You realize the output was great last Tuesday and mediocre today and you can’t tell if it’s the model, or you, or both. You realize the thing you need most isn’t a better prompt. It’s a system that doesn’t depend on you remembering to do it right.

That’s the difference between prompting and a workflow. And once you see it, you can’t unsee it.

1 - The ceiling you’re about to hit.

Let me describe it more specifically, because I think a lot of people are sitting exactly here and haven’t named what’s wrong.

Prompting is reconstructive. Every time you need an output, you rebuild the context from scratch. You paste the instructions you saved in a note somewhere. You add the background the tool needs. You adjust for whatever mood you’re in, whatever you can remember from the last time it worked well. The output quality isn’t just a function of the prompt. It’s a function of how well you reconstructed the prompt today.

And here’s the part nobody talks about: the model is a variable too. Claude updates. ChatGPT updates. Anthropic and OpenAI run experiments in the background. The tool you used on Monday is not quite the same tool you’re using on Friday. You feel it in your outputs before you can articulate what changed.

This applies equally across tools. ChatGPT, Gemini, Claude. Same ceiling at the prompting level. The moment you need consistency, repeatability, or anything that touches a tool outside the chat window, you’ve outgrown what prompting alone can do.

The good news: there’s a next level. Most people don’t know it’s there because nobody explained that the map had more to it.

Share

2 - There’s a spectrum. Prompting is just the start.

The spectrum looks like this:

Prompt → Skill → Scheduled workflow

A prompt is a single request. Every time you need the output, you make the request again. This is where most people live.

A skill is a saved instruction set that lives inside Claude. When you use one, Claude already knows the job before you type anything: the format, the context, the constraints, the things you always want it to do and the things you never want it to do. You don’t reconstruct the instructions. You just start. Skills only need Claude chat. No Pro subscription, no desktop app.

I’ll cover both methods for building skills next week, because they deserve their own article. The short version: you can either build one from scratch using the Add new skill button, or you can have a conversation with Claude, get the output where you want it, and then ask Claude to package everything you’ve worked out into a reusable skill. That second method is what I use.

A scheduled workflow is different again. It runs on a timer or trigger, moves data across tools, and produces an output without you initiating it. Two types matter here, and understanding the difference saves you a headache:

Cowork routines run in the cloud. Laptop closed, machine off, they still run. This is where connectors do the real work. A routine can reach into your Gmail, pull data from Notion, push outputs wherever they need to go, completely on its own. You set it up in Claude Code.

Cowork scheduled tasks run locally on your machine. The desktop app needs to be open and the laptop needs to be on. More flexibility for working with local files, but you’re tied to your hardware. This is an important distinction: if you want something running at 3am while you sleep, it’s a routine, not a scheduled task.

Both types require a Pro subscription and the Claude desktop app to set up.

Now, the part that changes what’s possible: connectors.

Subscribe now

Claude can connect to Gmail, Notion, Slack, and other tools you already use. Pulling data in, pushing outputs out, triggering actions across your actual stack. You can check what’s available under the Connectors section in Claude.

If a tool you need isn’t in the directory yet, you’re not stuck. Find the tool’s MCP server URL in its documentation or developer settings, paste it under Add custom connector, and you’re in. The standard that makes this possible is called MCP, Model Context Protocol, and the URL is all you need to know about it.

For prompting, ChatGPT, Gemini, and Claude are broadly comparable. For workflows, Claude is currently the strongest non-technical option because of Cowork. If you want to go further and build something more robust that isn’t tied to any single AI platform, Claude Code or OpenAI Codex is where that goes. I have a full setup guide here if you’re ready for that.

But here’s the thing I want to name before we get into the examples. The shift from prompting to workflows isn’t a technical shift. It’s an organizational one. You’re not learning to code. You’re learning to think about your work as a repeatable system before you open the tool. That’s it. Everything else is just setup.

3 - Three examples from my own stack.

A. The Substack content engine.

I subscribe to too much. I write too much. And for a long time I had no real answer to “what’s actually working?” on Substack.

I had feelings about it. I had hunches. I noticed that some notes got more engagement than others, but I couldn’t tell you why with any confidence.

So I built a system that answers the question from data instead of intuition. It scans 125 of my past Substack Notes, scores them against engagement metrics, and ranks them by format. The finding that came back: personal_story format is 3.8x overrepresented in the top-performing quartile. Not a hunch. A pattern in the data.

From that, 21 draft posts were generated and pushed automatically to my Notion database. Ready for the week. Formatted. Prioritized. Done.

Any database works here, not just Notion. That’s just what I use.

The difference from prompting: I never sit down and ask “what should I write next?” The system answers that question on a cadence. It doesn’t depend on my energy or memory or how good I am at self-assessing that day. The decision moved out of my head and into a process.

B. The LinkedIn analytics engine.

Same logic, different problem.

I was posting consistently but making strategy decisions based on gut feel. The hook felt strong. The timing seemed right. But was it?

I built a system that runs hook-lift and timing analysis on my actual post data. The output lives in a live dashboard with two tabs I check every week.

The strategy tab surfaces patterns from past performance: admitted weakness framing produces a 2.88x engagement lift. AI news without personal stake scores 0.28x. That second number changed what I write about more than any piece of advice I’ve read on LinkedIn growth. It also generates a recommended posting cadence: 9 posts a week, front-loaded Monday and Tuesday with the strongest hook types, Sunday personal story to prime the audience for Monday.

The best times tab breaks my audience down by timezone, then maps optimal posting slots against day-of-week quality. Peak, good, or low. Based on what actually happened in my own data over time, not someone else’s “best time to post on LinkedIn” article.

I’m not asking “was this post good?” I’m running the same analysis framework against fresh data on a cadence and reading the results. One is a question. The other is a system.

Subscribe now

C. The newsletter scanner.

This one I want to describe differently, because it illustrates something the first two examples don’t.

It scans and summarizes the newsletters landing in my inbox on a schedule, so I stay informed without spending time on triage. It’s a Cowork scheduled task, which means the laptop needs to be on when it fires. You can also set it up as a Cowork routine if you want it running in the cloud regardless of whether your machine is open.

But here’s what makes this example matter beyond the mechanics: I didn’t know I needed it.

I didn’t sit down one day and think “I should automate my newsletter reading.” The need wasn’t visible to me, because the workflow couldn’t exist until connectors existed. You can’t tell Claude “scan my inbox” through a plain prompt. It has no access. The moment Gmail became connectable, a whole category of possible workflows appeared that simply didn’t exist the week before.

That’s what connectors unlock, and it’s genuinely different from what I expected. Not just the automation of things you already do manually. Entirely new workflows you’d never have thought to build because they had no path to existing before.

I wrote a full article about this one. It’s called “read.” and it covers the setup, what it replaced, and the specific kind of mental clutter it removed.

4 - Where to start.

One question picks the right path.

Does it need to run without you, and does it touch external tools? Start with a Cowork routine. Runs in the cloud, works with the laptop off, connects to your actual stack. Requires Pro and the desktop app. This is the most powerful entry point, and also the most setup.

Does it need a schedule but only touches local files? Cowork scheduled task. Same subscription requirements, but it runs on your machine. The laptop needs to be on when it fires. Useful for file-based work, less reliable for anything you want fully unattended.

Is it a repeating task that doesn’t need a schedule? Start with a skill. Claude chat only, no Pro subscription required. You build the instruction set once, and it’s there every time you need that output. This is the lowest-friction entry point, and the right place for most people to begin.

Next week I’ll walk through both methods for building skills, and why even if you never touch Cowork, skills alone change the way you work.

Prompting got you here. A workflow gets you further.

Pick the task you use AI for most this week. One question: does it need to run without you, or does it just need to remember itself?

That answer tells you where to start. Next week, I’ll show you the faster of the two.

PS - Let me know how it goes. I encourage you to share ….

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

What you get in this article: the full step-by-step guide to setting up a VS Code project that works with Claude Code, Codex, and any future AI agent. A one-command trick to keep everything in sync automatically. And if you already have a Claude Cowork setup you want to migrate - the conversion prompt is in the subscriber chat (welcome email if you are new).

It was a Monday morning in Munich. I opened my laptop. 50% of my weekly Claude tokens were already gone.

I had just flown back from Vietnam. Different timezone. And while I was somewhere over the Gulf of Tonkin every scheduled task I had set up in Claude Cowork had been firing. On European peak hours. Burning through my token allocation in the exact window I’m supposed to be sleeping (not literally - Claude can’t run on a closed laptop without WiFi, I’m being dramatic. But you get the point.).

That was the moment I decided I was done.

1 - What I Had Built

For the past few months, Claude Cowork had been running my life. Scheduled tasks, live artifacts, weekly content workflows. Zero code. Serious automation.

It was genuinely useful. I had a content engine, a research tool, a weekly analytics summary - all running on autopilot while I worked my day job.

Then it started breaking.

Tasks failed silently. I fixed them. The fix broke something else. Fixing that cost more tokens. By the time I was done troubleshooting, I had spent more tokens debugging than the original task would have used.

And then Anthropic started quietly changing how much intelligence the models actually deliver. Not in a press release. Just... less. Results that were solid two weeks before started coming back wrong. Tasks I had not touched were suddenly producing garbage. More fixes. More tokens. More time I was not getting back.

It was a leak I could not close. And then I flew home to a half-empty token budget on a Monday morning.

2 - This Is Not a Claude Problem

The timezone was just the trigger. The real problem was that I had built everything inside one proprietary tool I do not control.

One subscription. One pricing tier. One company’s decisions about token limits, model updates, and how much intelligence to quietly dial down on any given week. I did not own my workflows. I rented them.

The math made it concrete. Upgrading to the next Claude tier costs €90 per month alone. Claude Code plus Codex together cost around €40 total. Same capability. Two agents. No single point of failure.

But the money is not even the main argument.

The main argument is this: if Claude Cowork changes its pricing tomorrow, deprecates a feature, or shifts the model again, every workflow I have built breaks. I start over. The longer I wait, the more there is to rebuild.

The most expensive thing in any AI setup is not the subscription. It is starting over.

3 - The Fix: Build in Files, Not in Tools

Here is what nobody tells you when you start automating with AI tools. The workflows are not the problem. The problem is where they live.

If your instructions, context, and project knowledge live inside a proprietary tool’s interface, you are locked in. If they live in plain text files you own, any agent can read them. You point a different agent at the folder and keep going.

This is what “tool-agnostic” means in practice. The agent is replaceable. The files are not.

My new setup:

VS Code as the project home. Claude Code as one agent. Codex (OpenAI’s coding agent) as another. The same setup works with Cursor, Windsurf, GitHub Copilot, Gemini CLI, or any future tool that does not exist yet. The moment something better ships, you switch. Your project knowledge moves with you.

I will be honest: I expected Codex to be the weaker option. It is not. I have been genuinely surprised by how proactive it is. It plans ahead, flags potential problems before touching anything, and explains its reasoning without being asked. Right now, for my workflows, it tops Claude Code.

My developer friend was testing the setup at the same time. His verdict on Claude arrived in four messages: “fucking dumb bitch,” “nah, im done with his dumb ass,” “we over,” “getting blocked.” Then: “unblock tomorrow.” Then: “codex credits coming back in 20min or so.”

I sent him my take in the same conversation: “claude keeps forgetting and correcting itself constantly. codex is like the eager new employee who goes 20 extra miles.”

He did unblock Claude the next day. We are both still using both.

Note: To be fair, when Claude is behaving normally, it is still outperfoming Codex on most things that I use it for.

Already built in Claude Cowork and want to migrate?

Most people reading this are not starting from scratch. They have existing projects, tasks, and workflows they want to keep.

I built a full conversion prompt for exactly this. You paste it into Codex, it inspects your existing Claude project, creates all the adapter files automatically, and shows you the full plan before touching a single file.

Already subscribed? I posted it in the community chat.

Not subscribed yet? You will get it in your welcome email when you join.

If you are starting a fresh project, keep reading. The setup guide below is for you.

Subscribe now

4 - How to Set It Up

A note before we start: this section is a bit more technical than what I usually cover. If you have never used VS Code, I wrote a full setup guide for non-technical professionals here:

Work through that first, then come back. It is worth learning now. The longer you wait, the more you build inside tools you do not own.

If you have a developer in your team, send them this article. They will have the whole thing running in an afternoon.

A. Start with a folder. Just a folder.

Here is the thing that makes this approachable: a “project” in this setup is not a special thing inside an app. It is literally a folder on your computer.

Create a folder anywhere. Desktop. Documents. Wherever you keep your files. Name it after your project. That is your project. Everything goes inside that folder - your instructions, your workflows, your config files. Plain text files in a plain folder.

Then open VS Code, go to File → Open Folder, and select that folder. VS Code will show everything inside it in the left panel. You edit files there. Your AI agents read from there.

That is the whole idea. No proprietary format. No sync. No cloud account required just to hold your files.

A note on creating files and folders: You can create every file and folder in this setup either directly on your computer (using Finder on Mac, File Explorer on Windows) or inside VS Code using the file panel on the left. Both work exactly the same way. Section 4D shows the full folder structure you are building towards - take a look there before you start so you know what the end result looks like.

B. Install the extensions - and no, you do not need the terminal

Both Claude Code and Codex have official extensions for VS Code with full chat interfaces. You type to them exactly the way you type in Claude.ai or ChatGPT. No terminal. No commands. Just a chat panel in your editor.

For Claude Code:

→ Click the Extensions icon in the VS Code left sidebar (it looks like four squares).

→ Search for “Claude Code.”

→ Install the one published by Anthropic.

→ A chat panel will appear. Sign in with your Anthropic account

For Codex:

→ Same process: Extensions icon

→ search “Codex”

→ install the one published by OpenAI.

→ A Codex icon will appear in the sidebar. Click it.

→ Click “Sign in with ChatGPT” and log in with your OpenAI account.

Both panels work like a chat window. You type what you want. The agent reads your project files and responds. Neither requires you to open a terminal for everyday use.

C. The file your agents actually read

Every AI coding agent needs a briefing document. A file it reads automatically when you open the project. Your instructions, your context, your goals - all in one place.

The open standard for this file is called AGENTS.md. It is not just OpenAI’s format. It is a cross-tool standard, now stewarded by the Linux Foundation, and used by Codex, Cursor, GitHub Copilot, Windsurf, Gemini CLI, and a growing list of others. Over 60,000 public repositories on GitHub already use it. The idea: write your project instructions once, and any agent can read them.

Claude Code uses a file called CLAUDE.md. Same idea, different name, Claude-specific format.

The solution: write everything in AGENTS.md. Then make CLAUDE.md point to it automatically. I will show you exactly how in the bonus section.

To create AGENTS.md in VS Code: right-click anywhere in the left file panel → New File → name it AGENTS.md. Then write your project instructions inside it.

Here is a simple starting template:

# Project Name

## What this project is
[Describe the project in 2-3 sentences]

## My goals
[What you are trying to accomplish with this project]

## How I want the agent to behave
[Your preferences: tone, approach, what to always do, what to never do]

## Important context
[Anything else the agent needs to know]

Do not overthink it. Three sentences is enough to start. You will update it as you go.

Share

D. The folder structure

Beyond AGENTS.md, each tool has its own config folder inside your project. You do not need to understand all of it. Here is what matters:

Your project folder/
│
├── AGENTS.md          ← write your instructions here (the open standard)
├── CLAUDE.md          ← we will link this to AGENTS.md automatically
│
├── .claude/           ← Claude Code's config (created automatically)
│
├── .codex/            ← Codex's config
│   └── config.toml
│
└── docs/              ← shared knowledge both agents can read

Folders starting with a dot are hidden by default on Mac and Linux. VS Code shows them normally in its file panel. On your desktop you will not see them unless you enable “show hidden files” in your system settings - this is fine, you do not need to go into them manually.

Remember: every folder and file shown below can be created either on your computer the normal way (right-click on your desktop or in Finder/File Explorer → New Folder or New File) or inside VS Code using the left file panel. There is no difference. Use whichever feels more natural.

Create the docs/ folder by right-clicking in VS Code → New Folder → name it docs. This is where you put anything both agents should know about your project that does not belong in the briefing file itself: reference documents, templates, style guides. Both agents can read from here.

For the .codex/ folder: right-click → New Folder → name it .codex. Inside it, right-click → New File → name it config.toml. Paste this minimal starting config:

[model]
provider = "openai"

The .claude/ folder is created automatically by Claude Code the first time you run it. You do not need to create it manually.

E. The three-layer mental model

Before you start filling things in, this is the idea that makes the whole setup portable:

→ Layer 1: Shared knowledge. Anything both agents need to know. Lives in docs/. Write it once. Both agents read it. Never duplicate it inside a tool folder.

→ Layer 2: Skills (reusable workflows). Same format, two folders: .claude/skills/ for Claude Code, .agents/skills/ for Codex. Copy the same files into both.

→ Layer 3: Tool config. .claude/ and .codex/ are tool-specific. They configure different things and are not interchangeable.

The goal: your actual project knowledge lives in files any agent can read. The tool-specific folders are just adapters.

F. Four things that trip people up

✦ The .agents/skills/ folder is where Codex looks for reusable workflows - not inside .codex/. Create it at your project root: right-click → New Folder → .agents, then inside that, New Folder → skills.

✦ Agent definition files use different formats. Claude Code agent files end in .md. Codex agent files end in .toml. The instructions inside work the same way; the file format is different.

✦ The two config files (.claude/settings.json and .codex/config.toml) configure completely different things. Do not copy one into the other.

✦ Put shared knowledge in docs/. The moment you put project knowledge inside .claude/ or .codex/, only one agent has access to it. That defeats the purpose.

Subscribe now

BONUS - The Symlink Trick

The laziest possible way to keep AGENTS.md and CLAUDE.md in sync. Permanently.

The problem: AGENTS.md is your source of truth. But Claude Code reads CLAUDE.md. If you keep two separate files, they will drift apart. You will forget to update one. Your agents will work from different instructions.

One terminal command. Run it once when you set up the project. Never think about it again.

On Mac or Linux, from your project root:

ln -s AGENTS.md CLAUDE.md

On Windows - open Command Prompt as Administrator (or enable Developer Mode in Windows Settings first):

mklink CLAUDE.md AGENTS.md

CLAUDE.md is now a symbolic link. It is not a copy. When Claude Code reads CLAUDE.md, it is reading AGENTS.md directly. Every update you make to AGENTS.md is automatically what Claude Code sees. No sync step. No drift. No maintenance.

This is also why AGENTS.md is the right file to write in. It is the open standard. Every serious AI coding tool already reads it, and more are adding support every month. You write once. Every tool follows.

Don’t build workflows in tools. Build them in files.

Any agent can read a file. Not every agent can read a proprietary interface you don’t own.

Set it up , the sooner the better.

PS - If you are also comparing Claude Code and Codex, please let me know your experience. I am leaning towards Codex for everything other than writing at the moment.

Leave a comment

Subscribers get the full migration prompt for Codex for existing projects 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

I didn’t subscribe because I’m lazy. I subscribed because I genuinely didn’t want to miss anything. The person writing was good. The topic was relevant. The intention was real.

Then they all arrived on the same Tuesday. I didn’t have time. So I left them on unread - I’d obviously get to them later. Two weeks later, my inbox looked like a catastrophe and I’d read exactly none of them.

The problem wasn’t the newsletters

I cannot function with unread emails. Inbox zero isn’t a productivity philosophy for me. It’s a psychological requirement.

But I subscribed to 40+ newsletters because I love them. Every single one felt worth it the moment I hit subscribe. I write one myself, which means I know exactly how much work goes into each issue.

The unread pile wasn’t laziness. It was FOMO stacked on top of bad timing, stacked on top of the kind of low-grade mental load that drains you every time you open email and see that number creep upward instead of down. I’d miss something good, buried in the pile, and only discover it three weeks later when it was no longer useful. The newsletters weren’t the problem. The delivery format was broken.

What I actually built

Before automating anything, I needed to know what I was dealing with. So I asked Claude Cowork to scan my Gmail and produce a list of every newsletter I was subscribed to, with a proposed tier list based on what it could infer about my priorities from the content.

It came back with a ranked list. I edited it. Some newsletters I thought I loved turned out to be tier 3 material when I actually looked at them honestly. A few I’d forgotten about were quietly excellent.

Then the practical part. I created a Gmail filter. Every newsletter on my approved list now gets a label: “claude newsletter.” They land in my inbox with that label, clearly sorted. That took 5 minutes.

The scheduled task does the rest. Once a week, Claude scans everything labeled “claude newsletter,” clusters the content by theme, summarizes what’s relevant to me based on a brief I wrote, flags any topic appearing across multiple newsletters, and formats it as one curated email. Business ideas that surface in the digest get a separate note dropped into my Notion. Everything else stays in the email.

One honest limitation: Claude can’t send email directly to my inbox. It’s a current platform constraint. So the digest lands in my Drafts folder. I open it, hit send to myself, done. One click, thirty seconds.

Total setup time: 10 minutes with Claude to build the tier list and configure the task. 5 minutes to set up the Gmail filter.

2 Substackers actually made it to my Top Picks for the week, shoutout to and - great articles!

What I didn’t expect

The digest is one email per week. I check it Monday morning, usually while coffee is still hot.

After a few weeks I noticed something I hadn’t planned for. The same topics kept surfacing across different newsletters in the same 7-day window. Not because anyone was copying each other. Because those topics were actually moving.

Last week’s digest flagged it clearly: five newsletters mentioned AI agents in the context of non-technical users. Three flagged the same model update. Reading them individually, scattered across different days, I would have missed that pattern entirely. The compression made the signal visible in a way that 43 separate tabs never could.

For content creators, that’s worth something on its own. If a topic is trending across your corner of the internet in the same week, you have maybe 10 days before the market saturates. The digest became part of my content radar without me designing it that way.

The honest part

There’s a version of this that feels like cheating. Newsletters are someone’s distilled thinking, and I’m having Claude compress them further. Worth sitting with for a second.

But here’s the honest answer: I wasn’t reading them before. The choice was never between AI compression and deep, careful reading. It was between AI compression and a pile of unread guilt that meant I was engaging with zero ideas per week instead of forty.

The digest didn’t replace reading. It made reading intentional. When something in the summary is actually interesting, I click through. I read the full piece. That now happens by choice instead of obligation, and it happens more often than it used to.

How to build it yourself

What you need:

• Claude Pro (€17/month)

• Gmail connected in Claude Cowork

• Notion connected (optional - only if you want to capture business ideas separately)

• 15 minutes

Step 1: Ask Claude Cowork to scan your Gmail for newsletter subscriptions and produce a tier list. Edit it ruthlessly. Be honest about which ones you actually read vs. which ones just felt good to subscribe to.

Step 2: Set up a Gmail filter. Every newsletter on your approved list gets a label. This is the sorting mechanism that makes the whole thing work.

Step 3: Set up a weekly scheduled task in Claude Cowork with a brief like this one:

You are my weekly newsletter digest assistant.

Scan my Gmail for emails labeled "claude newsletter" from the past 7 days.
For each newsletter: extract ideas relevant to [your topic or audience — be specific].
Skip anything that isn't directly relevant.

Cluster the content by theme.
Flag any topic that appears across 3 or more newsletters this week.
For anything worth reading in full, include the original link.
If you find a business idea worth saving, note it separately at the top.

Format everything as one curated email.
Save the output as a Gmail draft.

The brief is everything. Vague brief gets you summaries you still won’t read. Specific brief gets you signal you’ll actually use. The more clearly you define what “relevant” means for your niche, the better the output.

This might not be for you

I’m sharing this because it works for me. Not as the answer. As an example of what’s possible.

If you love reading newsletters one by one with your morning coffee, this will feel wrong and cold and pointless. If you’re subscribed to five newsletters, you genuinely don’t need it. If inbox anxiety isn’t part of your life, skip everything above.

But if you’re subscribed to things you love and somehow never reading them, if the unread badge is quietly eating your headspace, if you’re missing good ideas not because you don’t care but because the format keeps defeating you... then 15 minutes of setup might be worth trying.

AI doesn’t have to be complicated to be useful. Sometimes a Gmail label, a brief, and one email per week is enough.

Open Claude Cowork. Start with step one.

PS - Let me know how it goes. I encourage you to share ….

Leave a comment

Subscribers get all my Installation Guides including my explanations for dummies in the Resource section.

You are not a subscriber? Change it now!

Subscribe now

FAQ

Can Claude read my Gmail newsletters automatically? Yes, with the Gmail connector in Claude Cowork. You set up a scheduled task that runs once a week, scans emails with a specific label, and outputs a curated digest. No coding required.

Do I need Claude Pro for this? Yes. Claude Cowork requires a Pro subscription (€17/month). The Gmail and Notion connectors are included.

What’s the difference between a newsletter summary and a newsletter digest? A summary captures what was said. A digest — as set up here — captures what’s relevant to you, filtered through a brief you write. The brief is the whole system. Without it, you get summaries. With it, you get signal.

How long does setup take? About 15 minutes total. 10 minutes with Claude to build the tier list and configure the task. 5 minutes to create the Gmail filter.

Why does the digest land in Drafts instead of my inbox? Current limitation of the platform. Claude can’t send email to your inbox directly. The draft lands in Gmail and you send it to yourself. One click.

What if Claude misses a newsletter? It can happen, especially if the email lands outside your labeled folder or arrives from a sender variant you haven’t whitelisted. Check the output Monday morning and adjust your filter over the first two weeks. It gets more accurate fast.

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

Social media automation sounds like a developer thing. It is not.

I’m non-technical. I don’t code. I’ve never built an app. And I now have automated posting set up across Instagram, Facebook, Threads, and LinkedIn. The accounts are live: @brisk.vision on Instagram and Threads, brisk.ai on Facebook, and brisk.ai on TikTok - all started few weeks ago to prove that this can be built and run without a developer.

And without paying for a 3rd party provider.

The setup sounds more complicated than it is. Stick with me.

0 - Start here: the Claude Code folder

Before you touch a single platform, build one folder.

If you’ve been reading brisk. for a while, you know I work inside Claude Code and try to rebuild everything for as little subscription money as possible.

Instead of starting from scratch every session, I have a project folder that Claude reads every time. It knows my brand, my platforms, my tone. I describe what I want. It runs.

Here’s the folder structure I recommend for a social media automation setup:

/social-automation/
│
├── CLAUDE.md                   ← tells Claude who you are and what to do
├── .env                        ← your API keys (more on this in a minute)
│
├── /content/
│   ├── /carousels/             ← images from the carousel workflow
│   ├── /images/                ← single-image posts
│   └── /captions/              ← text content, ready to post
│
├── /skills/
│   ├── carousel-generator.py   ← from my carousel article
│   └── image-generator.py      ← from my image stack article
│
├── /posting/
│   ├── threads-post.py
│   ├── instagram-post.py
│   ├── facebook-post.py
│   └── buffer-queue.py
│
└── /logs/
    └── post-log.json

Recap: what is a CLAUDE.md file?

It’s a plain text file that sits at the top of your project and briefs Claude on everything it needs to know. Your brand voice. Your platforms. Your posting schedule. What your skills do. Instead of re-explaining yourself every session, it just knows.

Think of it as the briefing document you’d give a new assistant. Except this one reads it in under a second.

→ I’m sharing my exact CLAUDE.md template as a resource for my subscribers.

Subscribe now

What is a .env file?

A .env file - pronounced “dot env” - is where you store your API keys. An API key is a password for a tool or platform. You fill this file in once and your scripts read from it automatically.

It looks like this:

META_ACCESS_TOKEN=example_abc123yourtokenhere
META_PAGE_ID=example_123456789
INSTAGRAM_ACCOUNT_ID=example_987654321
BUFFER_ACCESS_TOKEN=example_def456yourtokenhere

One rule: never share this file. Don’t upload it to GitHub. Don’t paste it anywhere. It’s your master key.

1 - The easy ones: Meta platforms

Facebook, Instagram, and Threads all live inside the same developer system: Meta’s Graph API. You create one developer account, one app, and you get access to all three.

This is the best deal in social media automation.

Step A: Create your Meta developer account

Go to developers.facebook.com and sign up. You need a personal Facebook account to do this.

Once you’re in, click “Create App.” Select “Business” as the app type.

Note: My Meta developer account is a few years old. When I set this up, there was no approval process for personal use. If your account is brand new, Meta now technically requires “app review” before your app can be used by others - but for posting to your own accounts, you can stay in development mode and skip the review entirely. Development mode is all you need.

Step B: Add the products you need

Inside your app, add these three products:

→ Instagram Graph API

→ Facebook Pages API

→ Threads API

Step C: Understand the three things you need

Every platform API works with three pieces of information. Here’s what they mean in plain language:

• App Secret - a password for your app. Never shared publicly.

• Access Token - a temporary key that proves you’re authorised to post. Think of it like a session badge. Standard tokens expire after 1 hour. Long-lived tokens last 60 days.

• Account ID / Page ID - the unique number for your account or page. Not your username. A number.

You’ll need all three for each platform.

Step D: Get your access token

Inside your Meta app, go to the Graph API Explorer (linked from your app dashboard). Connect your Instagram Business account and your Facebook Page. Request these permissions:

→ instagram_basic 
→ instagram_content_publish 
→ pages_manage_posts 
→ pages_read_engagement 
→ threads_basic 
→ threads_content_publish

Generate a User Access Token. Then exchange it for a Long-Lived Token - this one lasts 60 days instead of 1 hour.

Technical flag: Long-lived tokens still expire every 60 days. You’ll need to refresh them. This is where having a developer, a Claude Code setup, or a tool like Blotato helps - the refresh can be automated so you don’t have to remember.

Step E: Find your Account IDs

→ Your Facebook Page ID: go to your Facebook Page, click About, scroll to Page Transparency.

→ Your Instagram Business Account ID: visible in the Graph API Explorer after connecting your account.

→ Your Threads User ID: also pulled from the Meta Graph API - separate from your Instagram ID, even though the accounts are linked.

Step F: Fill in your .env file

META_APP_ID=your_app_id
META_APP_SECRET=your_app_secret_never_share_this
META_ACCESS_TOKEN=your_long_lived_token
META_PAGE_ID=your_facebook_page_id
INSTAGRAM_ACCOUNT_ID=your_instagram_account_id
THREADS_USER_ID=your_threads_user_id

That’s it. Your posting scripts read from this file. You don’t paste keys around.

→ Full screenshot walkthrough of this entire setup - every screen, every click - is available for subscribers.

Subscribe now

A. Threads - what to know

Threads uses the same Meta developer infrastructure as Instagram and Facebook. Same app, same token flow. The difference: it has its own permission scopes (threads_basic, threads_content_publish) and its own User ID - separate from your Instagram ID.

One limitation worth knowing: the Threads API only publishes immediately. There’s no native “post at 9am” option built in. If you want to schedule, your script needs to run at the right time - which Claude Code can handle with a scheduled task (just ask Claude Code to set this up in the chat interface).

Text posts up to 500 characters, single images, single videos, and carousels up to 20 items all work. GIFs and Stories are not supported via the API.

B. Facebook - the easiest of the three

The Pages API is the most straightforward of all the Meta products. Images, text, video, and scheduled posts all work out of the box.

One thing to know: API publishing only works for Facebook Pages, not personal profiles. If you haven’t already, create a Page for your brand. That’s what you connect.

C. Instagram - what works and what doesn’t

Images, carousels, and single videos: fully automatable.

Reels with trending or licensed music: not automatable via the standard API. Meta blocks music permissions at the API level. This is a known limitation.

My developer friend found an open-source solution that incorporates trending audio into the automated posting flow. We are currently testing it. If it holds up, I’ll share the full details and the GitHub link in a later article. Watch this space.

Bonus tip: Instagram to Facebook cross-posting

If Facebook is not a priority channel for you but you want presence there, you can configure your posting script to mirror Instagram posts to your linked Facebook Page automatically. One API call, two platforms. Worth setting up if you don’t want to create separate Facebook content.

2 - The medium one: LinkedIn

LinkedIn has an official API. It requires enterprise-level partner approval to access for posting. That’s not for individual creators.

The workaround: Buffer.

Buffer is a social media scheduling tool. It has its own open API, and it connects to LinkedIn. The flow is: your Claude Code script sends content to Buffer, Buffer posts it to LinkedIn on schedule. Buffer is the middleman - and in this case, the middleman does all the hard work.

Buffer’s free plan covers 3 connected social accounts and up to 10 scheduled posts. That’s enough to test the whole setup. Paid plans start at around $6/month per channel if you need more volume.

buffer.com

Step A: Create a Buffer account and connect LinkedIn

Sign up at buffer.com. Connect your LinkedIn profile inside Buffer’s dashboard. This is a standard login flow - no developer account needed.

Step B: Get your Buffer access token

Go to buffer.com/developers and create an app. You’ll get an Access Token. Copy it.

Step C: Get your LinkedIn Profile ID from Buffer

Buffer shows you all connected accounts and their IDs after you connect LinkedIn. Copy your profile ID.

Step D: Fill in your .env file

BUFFER_ACCESS_TOKEN=your_buffer_token
BUFFER_LINKEDIN_PROFILE_ID=your_linkedin_profile_id

Step E: Send posts via API

Your posting script sends content to Buffer’s API endpoint:

POST https://api.bufferapp.com/1/updates/create.json

The post appears in your LinkedIn queue. Buffer publishes it at the scheduled time.

What you can automate: post text, images, scheduling. What you still can’t automate: LinkedIn analytics. That’s still a manual CSV export for now.

3 - The partial one: Substack

Substack has an API. It’s useful - but not for posting.

You can use it to read your own analytics, pull subscriber data, and monitor other newsletters. Competitor research, performance tracking. Genuinely useful for content strategy.

What it won’t do: schedule Notes, automate publishing, queue anything.

Notes must be posted manually. A custom script is technically possible but requires a developer or solid Claude Code skills.

4 - The ones not worth it (for now)

TikTok and Pinterest

Both platforms have APIs. Both require you to submit a developer application and demonstrate a working use case in a sandbox environment before getting access. A sandbox is a testing environment - you essentially have to prove your integration works before they let you touch the real platform.

This is a genuine developer process. If you don’t have a developer or a solid Claude Code setup, don’t start here.

Claude can write all your TikTok scripts and Pinterest captions. The posting stays manual.

X / Twitter

As of February 2026, X moved to pay-per-use pricing for all new developer accounts. Posting a single link costs $0.20 per post. Old flat-rate plans still exist for accounts that subscribed before the change, but new accounts pay per call. At any meaningful volume, this doesn’t make financial sense.

Skip it unless you’re already heavily invested.

This is where Blotato comes in.

built Blotato because this patchwork of API applications, sandbox demos, and developer approvals is exactly what keeps non-technical creators stuck. She grew to 2.2M followers and 500M views without a team - and Blotato is the system she used and productized. Also she is educating basically the entire world on how to use AI for FREE.

(Big fan here 🙋‍♀️)

Blotato handles TikTok, Pinterest, Instagram, Facebook, LinkedIn, Threads, X, YouTube, Bluesky, and more. No app review. No sandbox demo. No developer. You connect your accounts through a normal login flow and it (she) handles the rest. AI writing is included. So are AI image generation, AI video creation, viral templates, and a scheduling calendar.

Plans start at $29/month (Starter). Creator is $97/month. There’s a 7-day free trial on every plan, no credit card required: blotato.com

The honest summary

Claude handles the content. The APIs handle the distribution. The table tells you exactly where one ends and the other begins.

The three subscriber resources for this article:

→ The .env setup sheet: every variable named and explained, with a checklist of where to find each value.

→ Full screenshot walkthrough of the Meta API setup - every screen, every click.

→ The social-automation starter folder: CLAUDE.md pre-filled, .env template, skeleton posting scripts with comments. Download, fill in your keys, start posting.

Subscribe now

Start with Meta. Get Facebook, Instagram, and Threads running off one setup. Then add Buffer for LinkedIn. That’s already four platforms posting automatically while you do something else.

The rest can wait.

PS - Let me know how it goes. I encourage you to share and ask questions if something wasn’t clear enough.

Also don’t be shy and follow me on my new channels. Let’s grow together:

Leave a comment

Subscribers get three resources for this article:

→ The .env setup sheet: every variable named and explained, with a checklist of where to find each value.

→ Full screenshot walkthrough of the Meta API setup - every screen, every click.

→ The social-automation starter folder: CLAUDE.md pre-filled, .env template, skeleton posting scripts with comments. Download, fill in your keys, start posting.

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 the same question.

I went looking for the answer. What I found was a hundred guides that all funneled into the same place: a subscription site selling generations.

I don’t like paid subscriptions.

Then it clicked. Higgsfield, Magic Hour, ImagineArt, that paid Canva plugin. They all wrap the same models I can access free. GPT Image 2 in ChatGPT. Nano Banana 2 in Google AI Studio. The ones underneath, no middleman.

I stopped reading tutorials. I started screenshotting examples I wanted to recreate. The challenge: build them all on a Claude Pro subscription, nothing else.

$20 a month for Claude Pro. A brand kit. The free models underneath. That’s the whole stack.

Let me show you, post by post.

Before the walkthroughs, the gate.

Everything below assumes you have a brand kit. Palette, fonts, voice, layouts, reference images.

Without it, AI gives you generic. With it, AI gives you you.

I covered the full brand setup in my carousel article. Subscribers get my exact template.

1. The single branded image

Claude Cowork with a brand skill produces these in one prompt. No Canva.

I doubted this for months. I was wrong. The brand skill is the whole game.

Here’s the brief I drop into Cowork:

Use my brisk. brand skill. Single-image Instagram post on this topic: “You don’t need a paid AI subscription yet” Format 1080x1350 PNG. Headline short with one italic accent word. Body: 4 numbered patterns, one line each. Brand colors only. Output as artifact.

Cowork builds an editable artifact. I tweak the headline if needed, download the PNG, post.

Under 3 minutes. Costs nothing on top of Claude Pro. The outcome (click on the image to enlarge):

The thing nobody tells you: text in image is solved now as well.

Nano Banana 2 (Gemini 3.1 Flash Image) and the new GPT image model both render typography accurately. The reason you used to leave the AI tool and finish in Canva no longer exists.

You go to Canva if you want extra kerning control or to drop in a hand-drawn element. Otherwise stay in Cowork.

Subscribe now

2. The branded carousel

I covered the set up of this one step by step in this article. I personally designed a range of different carousel pages in 2 design variations that I use for my brand:

3. The viral portrait

Polaroid trends. 3D figurines. You as a vinyl toy. All free. All produced with one of two tools.

Option A

ChatGPT free tier. GPT Image 2 launched April 21, 2026. You get around 3 generations a day, no watermark, commercial use OK. Best text rendering on the market (Status May 7, 2026)

Option B

Google AI Studio. Nano Banana 2 (Gemini 3.1 Flash Image). Around 50 generations a day, no watermark, commercial use OK. Higher volume.

I use both. ChatGPT when I want one polished output and bulletproof text. AI Studio when I want to iterate fast.

The setup is the same in either: upload one good headshot, reference it in the prompt, run.

Six I made of myself

✦ My facial analysis & personal styling report

✦ Me sitting on my own head

✦ Me on the cover of brisk. magazine, issue 47

✦ Me as a vinyl collectible figure in a brisk. branded box

✦ My Instagram page painted by a child in crayon

✦ Me & my brand as a comic character

Here’s the full prompt for the vinyl figure. Swap “brisk.” for your brand and run it tonight:

Oversized round-headed vinyl collectible figure of the person in the reference image. Large solid black dot eyes, minimal facial features, glossy plastic finish. Place the figure inside a collector’s display box with a transparent front window. Box exterior in deep teal (#00C9A7) with bold orange (#E8622A) accents. Top of box reads “brisk.” in clean sans-serif. Beneath, smaller text: “AI for people who don’t speak tech.” Inside the box beside the figure: a tiny laptop, a coffee cup, a small Substack icon. Studio product photography lighting, clean white background outside the box. 4:5 portrait aspect ratio. Maintain exact facial likeness to reference image.

Option C - and my favorite secret

The reverse-engineering trick is the part nobody publishes.

Find a viral image. Drop it into Claude or Gemini. Type: “Write me a prompt that would reproduce this image with a different subject.”

You get the prompt back. Swap the subject. Run it.

The other three prompts plus the meta-prompt and seven more are in the vault for subscribers.

Subscribe now

4. The Infographic

Three paths. I use all three depending on the look.

Path 1: Claude writes the HTML, I screenshot it.

The actual prompt I drop into Claude:

Create a self-contained HTML file with inline CSS for an Instagram infographic, 1080x1350px. Make sure the size is screenshotable. Topic: 3 skills I install on every Claude project. Brand: brisk. Colors: teal #00C9A7 accent, orange #E8622A secondary, off-white #F5F3EF background, near-black #0E0E0E text. Font: [your font]. Layout: bold title at top, accent word in cursive, 5 cards in a grid, footer with handle @brisk.vision. Render as artifact.

Claude renders it. I screenshot at 1080x1350. I post.

Good for: clean grids, comparisons, anything structured.

Can’t do: hand-drawn elements, sketchy aesthetics, custom illustrations. HTML renders HTML. For doodle-style, use Path 2.

Path 2: GPT Image 2 or Nano Banana 2 directly.

Best for hand-drawn or illustrated infographics. Free in either ChatGPT or AI Studio. Same headshot-as-reference workflow as the viral portraits.

Path 3: NotebookLM.

The option nobody mentions. Upload your sources, ask for an infographic, it builds one from your material (in your brand colors). Full playbook here.

Share

5. The 8-second video

Google made Veo 3.1 free for everyone in April. Nobody told you.

Where: Google Flow. 50 daily credits, around 12 videos a day. 720p, 8 seconds, watermarked. No card. No trial. Free.

The setup: upload a still image (your branded post, your headshot, anything). Add a motion prompt.

A prompt I used last week to animate a static picture of me (you can see the static image on the lower right side):

Subject in the reference image stands still while the camera slowly dollies in toward the face. Soft natural light from a window on the left. Subtle hair movement. Cinematic shallow depth of field. 8 seconds, no audio.

Veo returns a clip. Trim, post.

For animated posts that aren’t camera-on-photo (text reveals, infographic motion, branded explainers), I’m testing Blender and Cavalry. Cavalry just went fully free in April after Canva acquired it, no trial games. Full breakdown in a separate article.

The other AI video tiers, briefly:

✦ Kling: 66 daily credits, 720p, watermarked. Catch: free tier forbids commercial use.

✦ Hailuo (MiniMax): trial credits plus a small daily bonus, watermarked, also no commercial use free.

If you want commercial-safe and free for video, Veo 3.1 in Google Flow is the answer.

The rest of the stack

The list nobody publishes because there’s no affiliate revenue in it.

Every image on my feed comes out of the first list. Every one.

The vault has the rest:

7+ trending one-shot prompts ready to run, the reverse-engineering meta-prompt, my brand kit template, infographic HTML scaffolds, an aspect ratio cheat sheet for every platform.

For subscribers only.

Subscribe now

A small admission, and a new rule.

Last week I told you the next article would be a different one. Then this image question kept landing in my inbox, on Notes, in DMs. The signal was loud. I followed it.

From now on you decide. Every article ends with a poll. Whatever wins lands next.

Option 1:

✦ How to generate animated images: Blender, Cavalry, Veo 3.1 in practice

Option 2:

✦ How to set up automated posting via Claude for free: no extra SaaS, where the Meta API quietly limits you

Option 3:

✦ Live artifacts: How to set up competitor tracking and content prep automation in one dashboard

Now close this tab.

Open ChatGPT.

Make something for free.

PS - Let me know how it goes. I encourage you to share. Would love to see lots of different images in the comments.

Leave a comment

Subscribers get the full vault incl. detailed prompt and the brand kit from 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

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