Ask five people in your company what a standard job costs. Then ask the AI each of them uses.
You will not get one answer. Each person has their own chat, their own saved prompts and their own picture of how the company works. Some of those pictures are wrong. You cannot see which, because none of it lives anywhere you can open.
The fix is to stop treating AI as a tool people use and start running it as a department. This guide builds that in one folder on your own machine, in six steps.
What You Are Building
An AI department is one shared brain, agents that do each team's repeatable work, and one person accountable for all of it.
Think of the IT person who used to make sure every employee's computer worked. Or the operations lead who keeps each team's process up to date. The head of AI does that job for agents. They place an agent inside each department, where it works for that team's manager. They keep the agents on one shared memory, and they decide which part of the brain each department can open.
Under the picture there are four parts, and you build them from the bottom up.
I wrote the reasoning behind this in the three rules. That page is the why. This one is the build.
Step One: Make The Brain
The brain is a folder of plain text files on a computer you own. It is not a chat history and it is not a feature inside one vendor's product. Any AI that reads files can work in it, so it survives every change of tool.
Mine has more folders than this. These are the ones that matter on day one.
Copy this.
company-brain/ AGENTS.md the manual. The AI reads this first, every session. index.md one line per topic, pointing to the file with the detail. notes/ your offer, pricing, decisions, call transcripts. clients/ one folder per client. os/ skills and routines (steps four and five). runs/ one line for every job an agent finishes. archive/ old things, out of the way but never deleted.
The manual is the file that turns a general AI into your AI. Start with this and edit it as you go.
Copy this.
# How to work in this company ## Who we are [COMPANY NAME] sells [WHAT YOU SELL, e.g. "bookkeeping for restaurants"] to [WHO BUYS IT]. The owner of this system is [YOUR NAME]. ## How I want work done - Read index.md before you answer anything about the company. - Open only the files a task needs. - If something you need is missing, say so. Do not guess. - Plan first. Wait for my yes. Then act. ## What you must never do - Send anything to a client, or spend money, without my explicit yes. - Print or copy passwords and keys. - Delete files. Move them to archive/ instead. ## Where things are See index.md.
Your First Real Run
Make the folder. Save the manual. Open the folder with the AI you already use. Claude Code and Codex both work. Then type: "read the manual and the index, then tell me in five lines what you know about this company."
It will know almost nothing. That is your starting point.
If you would rather have this installed for you, the Starter Kit does it. It is free and arrives by email. Its first skill checks your machine and its second builds the brain by interviewing you.
The Mistake That Makes It Fail
Designing forty folders before you have forty files. Start with the seven above and let the structure grow when something has no home.
Step Two: Fill It With What Only You Know
An empty brain knows nothing about you. It needs the things that today live only in your head and in your calls.
Start with four kinds of material: your offer and prices, transcripts of real client calls, past proposals that won, and the decisions you keep re-explaining. Record your calls from now on. A transcript dropped into notes/ teaches the brain more than an hour of typing.
Then let the AI tell you what is missing. You know what you forgot to tell it. You do not know what it does not know about you, so make it ask. The gap prompt in the three rules does exactly this. It lists the ten things it would need to know, asks you one at a time, and saves each answer as a file.
End every working session the same way. Have the AI write down what changed today: decisions, new facts, corrections. That one habit is what makes the brain better every week instead of older every week.
Your First Real Run
Drop your last three client call transcripts and your offer document into notes/. Run the gap prompt. Answer all ten questions.
The Mistake That Makes It Fail
Uploading 500 unlabelled files and calling it a knowledge base. A thousand books in a pile lose to fifty on labelled shelves. Every file gets a clear name and one line in the index, or it stays out.
Step Three: Run Every Conversation In 5 Moves
Most people's AI conversations are loops. You ask. It builds something 70 percent right. You correct it, and it changes three other things. Nobody ever says "finished".
I run mine in five moves, and they are written in the manual so the AI follows them without being asked.
- Plan. It tells me what it is going to do, as a short list, before it touches anything. For anything visual it shows me a rough version first.
- Approve. I say yes, no, or what to change. It does not start on a maybe.
- Act. It does what I approved and nothing else. If the plan did not cover something, it stops and asks one question.
- Feedback loop. It shows me what changed. I react to that part only, it fixes that part only, and we repeat until it is right.
- Done. We agreed at the start what finished looks like. When that is met, it says so and stops.
The fifth move is the one people skip. Without an agreed end, the AI keeps offering improvements and you keep accepting them.
Copy this.
How we work together, every time: 1. PLAN. Before you change, send or build anything, give me your plan as a short list. Include what "done" will look like. 2. APPROVE. Wait for my answer. I will say yes, no, or what to change. Do not start on a maybe. 3. ACT. Do only what I approved. If you hit something the plan did not cover, stop and ask me one question. 4. FEEDBACK. Show me what you did. When I give a note, fix that note and show me only what changed. Do not rework what I did not mention. 5. DONE. When the result matches what we agreed, say "done" and stop. Do not offer extra improvements.
Your First Real Run
Paste the block into your manual. Give the AI one real task today and do not let it act until you have read the plan.
The Mistake That Makes It Fail
Giving feedback on everything at once. Ten notes in one message and it rewrites the parts that were fine. Send one note, look at the fix, then send the next.
Step Four: Turn A Job Into A Skill
A skill is one job written down once, so an agent does it the same way every time. It is a plain text file. Nothing about it belongs to one AI vendor.
A task earns a skill when you have done it three times and can describe it without thinking.
Copy this.
# Skill: [NAME, e.g. "weekly client report"] ## When to use [The trigger, e.g. "every Friday, for each active client"] ## What it needs [The files and facts it reads, e.g. "clients/<name>/, this week's lines in runs/"] ## Steps 1. [First step] 2. [Second step] 3. [Third step] ## Done when [What finished looks like, e.g. "a one-page report saved in the client's folder"] ## Never [The limits, e.g. "never email the report. Save it and tell me."]
Save it as os/skills/[name]/SKILL.md. To run it, tell the AI: "read and follow the weekly client report skill."
You do not have to write the first ones yourself. Every skill I have published is in the skills section, free. Take one that fits and change the details. When you want to write your own properly, this guide goes deeper on how.
Your First Real Run
Pick the task you did three times last week. Say to the AI: "interview me about how I do this, ask about the edge cases, then write the skill." Run the skill once and correct it.
The Mistake That Makes It Fail
Writing a skill for something you have done once. You do not know the job yet, so you write down a guess, and the agent repeats your guess forever.
Step Five: Put It On A Schedule
A routine is a skill with a clock. The daily brief runs at seven. The lead list refreshes on Monday. Nobody has to remember.
Every routine gets three things before it is switched on:
- A schedule, written in plain words.
- Two limits: the most money one run may spend and the longest it may take. An agent with no cap is how surprise bills happen.
- A line in
runs/every time it finishes, saying what it did and whether it worked. This is how you check a department you are not watching.
The Starter Kit has a small control panel for this. It lists your skills, lets you pick a time and a day, and has one switch to run a skill automatically.
Your First Real Run
Take the skill from step four. Run it by hand three times. If all three are clean, give it a schedule and its two limits.
The Mistake That Makes It Fail
Scheduling anything that talks to a client before a person approves the message. Agents draft. People send, until you have seen enough drafts to trust the skill.
Step Six: Name The Owner
A department without a head is five people keeping their own numbers again. One person owns the brain, the skills and the routines.
From zero to your first million in revenue, that person is you. You are still experimenting, the context is in your head, and you cannot judge a job you have never done. Past a million you need your time back for partnerships and growth, and someone else takes the department. Pick someone who knows the company, not someone who knows the code. Train a person from inside yourself, or bring in an outside expert who trains them and leaves.
- The brain. One source of truth, and everyone reads and writes the same place.
- What goes to agents. Which work is handed over, which stays with people, and in what order.
- Quality. They answer for what an agent produces before it reaches a client.
- Cost, access and risk. The token bill, the tools, and which part of the brain each department can open. The sales agent has no reason to read payroll.
- Integration. This is the hard one. Companies decide they want AI and then meet friction from employees and managers. The head of AI makes the agents fit into every department. The agents stay current with what each team is doing, bring it back into the brain, and report out in a way people accept.
Your First Real Run
Add one line to the manual: "The owner of this system is [NAME]. Changes to skills and routines go through them." Then tell the team.
The Mistake That Makes It Fail
"Everyone owns it." Then nobody prunes the brain and nobody checks the runs. In six months you are back to the first picture on this page.
Honest: What This Costs You
The first weeks are slow. You are writing things down, and nothing looks automated yet.
Local means your machine. If the laptop is closed, the routines are off. A computer that stays on fixes that, and it is one more thing to own.
It needs the closing habit. A brain nobody writes back to is out of date in a month.
Some companies do need an engineer near this. That means the ones whose product is software, the ones holding regulated data, and the ones running on old custom systems. In those cases the engineer works for the head of AI.
The Order I Would Do It In
Steps one to three this week. Step four when a task has repeated three times. Step five when that skill has run clean three times by hand. Step six on the day a second person touches the system.
If you only do one thing today, make the folder and save the manual. It takes one sitting, and every later step stands on it.