SAVE THIS

I Burned 100 Billion Tokens. I'd Keep Three Things.

Six months and 100 billion tokens taught me three rules. Keep one setup. Build context instead of prompts. Run AI as a department with one person on top. Each rule is on this page with its first step and a prompt you can copy.

You tried a new AI tool last month. And the month before that. Each one looked faster than the last.

Your results did not get faster. Every switch reset something: your instructions, your saved chats, the way you had finally learned to ask.

I made that mistake at scale. Six months and 100 billion tokens, a lot of them spent learning three rules the slow way. If I started from zero today I would follow only these: keep one setup, feed it context instead of prompts, and run it like a department.

Rule One: Stop Switching Tools

A new AI tool ships every week. Most are worth knowing about. Almost none are worth moving into.

Think about what compounds in a business: a product you keep improving, a relationship you keep showing up for. Change your business idea every quarter and you never get past the first hard part. Change the people around you every month and nobody ever gets to know you.

AI works the same way. Every week inside one setup adds something, like an instruction you refined or a workflow that now runs without you. Move to a new tool and that work stays behind. You are back at week one with a nicer interface.

Chart: switching tools resets progress, one setup compoundsWhat your AI can do for youMonths in0123456new toolnew toolOne setupSwitching
Switching resets the curve. Staying lets it compound. An illustration of the idea, not measured data.

The launches are distractions for a simple reason. You can always plug the new thing into what you have, later. You cannot get back the months you spent rebuilding.

What I Do Instead

My setup is a folder of plain text files on my own machine. It holds my instructions, my client notes and the steps of every workflow. The AI reads the folder.

I run Claude most days and Codex on others. When a better model ships, I point it at the same folder and everything I built keeps working.

When a new tool launchesYour work lives inside the toolYour work lives in your own files
Your instructionsRewritten from memoryAlready there
What it learned about youLeft behindSame folder
Your workflowsRebuilt one by oneThey run on the new engine
Trying it costs youA restartA test

Your First Real Run

List every AI tool you paid for or tried in the last 90 days. Next to each one, write a thing you built there that still runs today.

Most lines will be empty. The tool with the most lines filled is your setup. Park the rest until one of them can work from your files.

Done. You have a home base.

The Mistake That Makes It Fail

Moving house because the new place has better furniture.

The model inside the new tool might be better this month. Models overtake each other all the time. If your work sits in files you own, you take the better model and keep everything. If it sits inside a product, the better model costs you a restart.

Ask one thing before you touch anything new: can it work from my existing files as they are? If yes, add it. If no, it waits.

Rule Two: Stop Writing Prompts. Build Context.

Someone I coach asked me a question last week. Did his AI get this good because of the brain we had built, or because the model is just that smart?

It is 80/20, and the 80 is the structure. The intelligence is the same for everyone. What changes is how much it knows about you and how fast it can find it.

Picture two librarians. The first has thousands of books and no idea which shelf anything is on. Ask him about World War II and you wait while he digs. The second has 50 books and knows exactly where each one sits. He answers first every time.

A long prompt turns your AI into the first librarian. You paste 2,000 words of background, rules and examples. The model digs through all of it on every message, and you pay for every word, every time.

With context, the background lives in files. The AI reads a short index, opens the two files this job needs, and starts. Three words can beat a 2,000-word prompt, but only if your AI knows where everything is.

Diagram: a long prompt versus structured contextA long promptContextall of it,on every messagethe indexthe index,then 2 filesyou pay for all of ityou pay for what it reads
A prompt sends everything on every message. Context lets the AI read a short index and open only what the job needs.

This is also where the token bill falls, because less gets sent. One example from my own setup: the index my AI reads at the start of every session went from 181 lines to 37. I deleted nothing. I moved the detail into topic files one step away, so it is read only when a task needs it.

How To Build It: 5 Steps

  1. Make one folder on your machine and fill it with plain Markdown files. Obsidian is a good window onto it, and any agent that reads files can work in it. If you have never built one, this guide sets the folder up for you.
  2. Put one manual at the top. Who you are, what the company sells, how you want work done, what the AI must never do. It reads this first, every session.
  3. Add one index. One line per topic, pointing to the file that holds the detail. This is the librarian's catalogue.
  4. Feed it the raw material: call transcripts, your offer, your pricing, past proposals, the emails you wish you could clone.
  5. Close every session the same way. Have the AI write down what changed today: decisions, new facts, corrections. Tomorrow starts ahead of today.

Copy The Gap Prompt

You know what you forgot to tell it. You do not know what it does not know about you. So ask.

Copy this.

I want you to find the holes in what you know about my business.

First read everything you have access to about me and my company.
Then list the 10 most important things you do NOT know and would need
in order to do real work for me without asking.

For each one give me:
1. The question, in one plain sentence.
2. Why it matters: what you would get wrong without it.
3. Where the answer probably lives (my head, a call, a document, a tool).

Then ask me the questions one at a time. After each answer, tell me in
one line what you now understand, save it as its own short file with a
clear name in [YOUR NOTES FOLDER], and add one line to the index.

If one of my answers is vague, say so and ask again. Do not guess.

My business: [ONE SENTENCE, e.g. "we install and service solar panels for homes in Arizona"].

In a plain chat window it cannot save files. Say: "give me each answer as a note I can paste, with a file name."

How I Keep A Conversation From Looping

You know the loop. You ask for something. It builds a version that is 70 percent right. You correct it, and it changes three other things. An hour later you are further from done than when you started.

I run every conversation in five moves, and the AI knows them because they sit in the manual.

Diagram: plan, approve, act, feedback loop, done1PlanIt proposes. It touches nothing.AI2ApproveYes, no, or change this.You3ActIt does only what was approved.AI4Feedback loopYou react to what changed.Both5DoneThe agreed end is met. It stops.Bothrepeat
How I run a conversation with AI. It plans, I approve, it acts, we loop on what changed, and it stops when the agreed end is met.

Plan: it tells me what it is going to do, as a short list, before it touches anything. Approve: I say yes, no, or what to change. Act: it does what I approved and nothing else. Feedback loop: it shows me what changed, and I react to that part only. Done: we agreed at the start what finished looks like, and when it is met the conversation ends.

Paste this into your manual:

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

Run the gap prompt today, in the tool you already use. Answer all ten questions.

It feels like the first coffee with a new hire.

The Mistake That Makes It Fail

Dumping everything in.

A thousand books in a pile lose to fifty on labelled shelves. What matters is whether the AI can find the one file it needs. If it has to read everything to use anything, you built the first librarian again.

Rule Three: Run AI As A Department, Not A Tool

Most companies use AI like this. Marketing has its prompts. Sales has a different tool. Operations has one person who is "good with ChatGPT". Nobody shares anything.

Now run finance that way. No finance department, and every employee keeps their own numbers in their own spreadsheet. You would never know what the company earned.

Diagram: a company where every person uses a separate AI with its own notesThe founderWhat does our AI know?Salesits ownnotesMarketingits ownnotes!Opsits ownnotesFinanceits ownnotes!Supportits ownnotes5 people. 5 AIs. 5 versions of the company.
Most companies today. Everyone has their own AI, each one knows a different version of the company, and nobody can see which ones are wrong.

That is AI in most businesses today. Some of those private AIs have the company wrong, and you cannot see which, because nothing they know lives anywhere you can open. The fix is structural, and almost nobody builds it.

What Is An AI Department?

An AI department is one shared brain, agents that do each team's repeatable work, and one person accountable for all of it. The brain is the single source of truth: what the company knows, in one place, read and written by every agent. The agents are the staff. The person on top decides what they do and answers for the result.

Diagram: an AI department with one shared brain, agents in each team and one head of AIHead of AIthe founder, later a Chief AI OfficerTHE BRAINone source of truthits partagentSalesits partagentMarketingits partagentOpsits partagentFinanceits partagentSupportOne brain. An agent in every team. One owner.
The same company with an AI department. One brain, an agent inside every team, and one person who deploys and manages them.

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.

The Setup Behind It

The setup is local. It lives in files you own instead of inside a vendor's chat history, and it has four parts. The full guide builds each one step by step.

PartWhat it isThe business equivalent
The brainThe folder from Rule TwoThe company's memory
SkillsA job written down once, step by step, so an agent does it the same way every timeA trained employee's procedure
RoutinesA skill on a scheduleA recurring task in someone's calendar
The ownerOne person who decides and checksThe head of the department

Every skill I build ends up in the skills section of the guides, free. Take the ones that fit your business and skip the rest.

Zero To A Million: The Department Head Is You

From zero to your first million in revenue, the head of AI is the founder. I see four reasons.

  1. You are still experimenting. The offer, the process and the tools change month to month. The person who decides has to be the person who builds, or every change waits for a handover.
  2. Only you hold the context. The business still lives in your head, and nobody else can teach the brain how the company really works.
  3. You cannot judge what you do not understand. Delegate AI before you have done it yourself and you cannot tell a good hire from a bad one. You cannot tell a good result from a lucky one either.
  4. The math does not justify a hire. At this stage you do one thing and scale it. With the brain in place, running the agents is one person's job.

This is the setup I would use to go from zero to a million without a payroll. The repeatable work runs on skills, and I stay on the part of the business I like.

Past A Million: A Chief AI Officer

Past a million the math flips. Your hours are worth more on partnerships, collaborations and the next opportunity than on tuning skills. You need your time back, so someone else takes the department.

That person should not be an engineer. The job is to know the company, and the code comes second. Two versions work:

For most companies the second is the better route, because you pay for the training once and keep the knowledge inside.

StageWho runs AIWhy
Zero to $1MThe founderStill experimenting, and the context is in one head
Past $1MA Chief AI Officer from inside the businessThe founder's time moves to partnerships and growth
Either stageNot a developer on their ownKnowing the company beats knowing the code

The 5 Jobs Of The Person On Top

  1. The brain. One source of truth, and everyone reads and writes the same place.
  2. What goes to agents. Which work is handed over, which stays with people, and in what order.
  3. Quality. They answer for what an agent produces before it reaches a client.
  4. Cost, access and risk. The token bill, the tools, and who or what can touch which data.
  5. 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.

Copy The Department Map Prompt

Copy this.

Act as my head of AI for one hour. Interview me to map my company as departments.

Ask me, one question at a time:
1. Which functions exist in my business today (sales, delivery, support,
   finance, hiring, marketing), even if I do all of them myself.
2. For each function: the three tasks that repeat every week.
3. For each task: what someone needs to know to do it well, and where
   that knowledge lives right now.

Then give me one table with these columns:
Function | Repeating task | What it needs to know | Where that lives today | Hand to an agent: YES / NOT YET / NEVER | Why

Rules:
- NEVER means a person must decide (pricing exceptions, letting someone go, anything legal).
- NOT YET means the knowledge is still only in my head. Tell me exactly what to write down first.
- Rank the YES rows by hours saved per week and tell me which ONE to build first.

My business: [ONE SENTENCE]. Team size: [NUMBER, e.g. "just me" or "6 people"].

Your First Real Run

Run the map prompt and build the one row it puts first.

If you would rather have the starting version installed for you, that is what the Starter Kit is for. It is free and arrives by email: five install skills, a control panel and six short guides. The first skill checks your machine. The second installs the brain by interviewing you. It runs on Claude Code or Codex.

The Mistake That Makes It Fail

Buying a tool for every team and calling it an AI strategy. Or hiring a developer to "do AI" before there is a brain for the agents to read.

Both leave you with five people keeping their own numbers.

Honest: Where This Does Not Fit

It is slower at the start. The first weeks go into 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.

A folder does not fix a company that refuses to write anything down. If the knowledge stays in people's heads, the brain stays empty.

The rule against engineers has exceptions. If your product itself is software or AI, you already need technical leadership, and that is a different job from running the AI department. If you handle regulated data, such as health records or client money, the head of AI will need an engineer. The same goes for a company that runs on old custom systems. In both cases the engineer works for the operator.

And the number in the title is a bill. The three rules are what I would keep, and the spend is what I would skip.

If You Only Do One Thing This Week

Run the gap prompt. Do it in the tool you already pay for, and answer all ten questions. Save the answers as files.

You will have the start of a brain, and you will know what your AI was missing. If it changes nothing, you lost one evening.

Come install these with me.
The community is free.

Operations Heroes is the free community where I install these systems live every Thursday, on real businesses. Three quick questions to join, and I call every new member.

Join the free community →
Take this with you Download as PDF ↓

Prefer to browse with company? The free community has the full skill library.

I write one system like this per week. Get the next one by email:

Free. Unsubscribe anytime with one click.