My CRM is a text file.

Last week I told you I had cancelled Webflow, Pipedrive, Zapier, n8n, Gamma and Adobe.

The obvious question is how.

So this week, no big take. I'm going to show you exactly how I work, what's in the folder, and what I learned the hard way. Steal whatever is useful.

Every morning starts the same way. I open Claude, point it at a folder on my laptop, and type:

"Check my CRM. What do I need to do today?"

And I get my day: who's overdue, which lead went quiet, what a client asked for last week and what I promised them. I never open a dashboard to get there.

My CRM is a text file.

Four years in one paragraph

To understand why this works now, and didn't two years ago, you need the short history.

In 2022 we had chat windows. You typed "you are the best marketer in the world" and hoped. Then came custom GPTs and assistants: you gave the AI your documents and instructions once, and stopped repeating yourself. Then agents: the AI could use tools, send the email, update the record. So everyone built one agent per task, then tried to orchestrate them. (I built a lot of those for clients, and I still teach them in my bootcamps, because the first principles matter. And remember RAG? It was high tech at the time.)

Then, at the end of 2025, came the harness. It started with Claude Code. One very capable agent that already knows how to read and write files, search, run code, check its own work and keep going. You don't build it. You open it. When you choose the work mode in Claude or ChatGPT today, you're using a harness.

And that changed everything, because it means building agents is no longer the bottleneck. I say that as someone who builds agents for companies for a living.

The bottleneck is now what the agent knows about you. Your context.

AI engineering has become context engineering.

What's actually in the folder

My AI OS (AI operating system) is just a folder. Here's the anatomy, simplified:

Cockpit. What's in progress, what's done, what matters this week. The agent reads it first.

Clients. One file per client or lead: who they are, what we agreed, what happened, what's next. That's the CRM.

SOPs. Standard operating procedures. How each recurring piece of work gets done, step by step, written the way you'd brief a new employee.

Prompts. Forget the prompt engineering tricks of 2023. For modern models, a prompt is just the combination of words that reliably produces a certain result. When I find one that works, it goes here.

Knowledge. Everything stable about the business: brand guide, logo, offers, writing rules, past work, what I never say in an email.

Automations. The few things that run without me, documented.

That's it. No software to install. Any harness can read it. And when you point the agent at this folder, you have an agent that knows everything about your business.

What it looks like on a real day

A client emails: "Could you add a second dashboard for the regional team?"

Two years ago, I'd have dug out the scope of work, checked whether it was included, maybe discussed it with the team, then written the reply.

Now I type: "Run the scope check SOP on the latest email from this client."

The agent opens the contract terms in the client file, follows the steps in the SOP, and comes back with a verdict: out of scope, here's why, here's the clause, and here's the one item in their email that's genuinely ambiguous.

I read it, agree, and say: "Go with your suggestion. Draft the reply." The draft sounds like me, because my writing rules are in the folder. Then: "Prepare a one-page change request, two days of work, as a PDF." It comes back in my brand, with my logo and my standards, because those are in the folder too.

Producing a PDF with AI is no longer impressive. Everyone can do that. What matters is that it produced my PDF, for my client, the way I work. I never re-explained anything.

Why one folder changes everything

My folder doesn't stop at clients. My website is in there. So are my books, every newsletter I've written, my courses, my playbooks and my research.

That's where it gets interesting. When I draft a client proposal, the agent can pull from a playbook I wrote two months ago. When I write a letter, it knows what I've already said. When I build slides, it reuses the structure of decks that worked.

Every file I add makes every other file more useful. The more I add, and the more I use it, the better it gets.

The compounding effect is EXCEPTIONAL.

Everything that makes AI work well for me lives in that one folder.

Five things that make it work

This is the part I wish someone had told me earlier.

1. SOPs beat prompts. A business needs reliable work. If the agent decides how to do the task every time, you get a different result every time. An SOP fixes the method, and it also says where the agent must stop and ask me before going further. Then, when a session finds a better way, I say: "Update the SOP with what we just did." From then on, every future session does it the better way.

2. One conversation is one employee. I don't build separate agents anymore. A conversation about my website, started with the right prompt from my folder, becomes my web developer. Another one becomes my calendar assistant. Same folder, different focus. The vocabulary of the conversation shapes the specialist.

3. The folder outlives the model. I use Claude today. If another model is better tomorrow, I point it at the same folder and keep working. Everything that makes my work good lives in my files, not in someone's product. (That's the portability I keep talking about since Letter 100.)

4. The expensive model teaches the cheap one. (I call it distillation, even if it's not exactly that.) For a complex process, I run it once with the most capable model. When it works, I ask it to document exactly what it did, step by step, in the SOP. After that, cheaper models run the same process with the same quality, because the hard thinking is now written down.

5. Connect, don't migrate. The folder is connected to my email, so the agent can read today's messages and update the client files. It's how a text file stays a live CRM. For the few things that must run on their own, the agent wrote small automations and deployed them for me. That's what replaced Zapier and n8n.

And that's how the six apps went. My website is run by one of those conversation-employees. The CRM is text plus email. The automations were written by the agent. My slides come out of the folder, with my brand, including the ones I make for clients.

When it's more than just you

I don't run Cohorte alone. We're a small team, and the folder is shared, at least the parts that should be. When a teammate runs an SOP from the shared folder, they produce the same quality of work I would, in the same time. And everything they improve flows back into the system for everyone.

Two practical lessons from doing this with my team and with clients:

Split only when you have to. Keep one shared folder until people start stepping on each other's work. In larger companies, what works is one setup per line of business: finance, marketing, and so on.

Back it up and version it. Once the folder holds real value, losing your laptop means losing your business brain. I keep mine versioned on GitHub. A synced cloud drive works too. Either way, you can always go back when something goes wrong.

The part that took me months of learning

Now the honest part, because I don't want to sell you a dream. Building the folder takes an afternoon. Making it something you can run a business on is a different story, and this is where most people get burned.

The compounding works both ways. The more good work goes into the folder, the better every future output gets. But if junk goes in, unchecked drafts and lazy SOPs, the system starts confidently producing junk, and it gets worse with use.

Your mind has to stay in the loop. If the agent drafts the email and you just send it, what happens when the client calls and asks about it? You don't really know what you said. (This is Letters 86 and 102 again: offload the thinking, and you lose it.) The SOPs have to put you at the moments that matter.

A company can't run on vibes. You're trusting this system with your clients, your reputation and your money. So you need to test it before you trust it: define what good looks like, run it on hard cases, write down what failed. Almost nobody does this part.

And underneath all three, there's something more basic. Most people use AI without understanding what it is: a prediction engine that drifts to the average unless your context tells it where to go (Letter 99). Without that, you'll build a beautifully organized folder and still get average work out of it.

That's why I keep saying the hard part is understanding the tools.

What to do this weekend

You can build version zero in an hour. Here's how I'd start:

Create the folder. Six sub-folders: Cockpit, Clients (or Projects), SOPs, Prompts, Knowledge, Automations. In Knowledge, write one page about you and your work: what you do, for whom, how you write, what you never do. Then keep adding: work you're proud of, your playbooks, your notes. Every file makes the others more useful.

Write one SOP. Pick a task you do every week. Write the steps as if briefing a smart new hire: what to read, what to produce, what good looks like, and where it must stop and ask you.

Run it, then improve it. Point Claude or ChatGPT in work mode at the folder and say "run this SOP." Fix what's wrong in the conversation. When the result is right, say "update the SOP with what we learned." Do it three times, and watch the difference.

That's the first brick. It compounds from there.

If you want the whole system

(Yes, this is the part where I tell you about something I sell. I'd rather say it plainly.)

I've spent the last few years building this with clients, teaching it in bootcamps to hundreds of people, and refining what actually holds up. I've put all of it into the AI OS Course, and it's available now.

It takes you from the first folder to specialists, automations and an agent working on your own files, under your rules. And it ends with the part almost no course teaches: how to test what you built and write an honest Trust Report on what it can and can't be trusted to do. It also includes a complete demo folder, so you can practise on a fictional business before touching your own.

It's self-paced, with lifetime access. The launch price is €397 until Friday 13 November, then €497.

Start building your AI OS →

Back to Monday morning

"Check my CRM. What do I need to do today?"

It still feels a little absurd to me that the answer comes from a folder of text files. But it does, every morning, and it gets a bit better every week, because the folder does.

Models expire. Your system compounds.

AI is only as good as the human operating it.

Have a great weekend.

Stay sharp.

— Charafeddine (CM)