AI CEO LabCLAUDE.md File: Turn Your Experience Into One AI File
Everyone's agent runs on the same models. The only thing that makes yours answer like you is a file that holds your business in four sentences, the numbers you check first, the rules you do not break, and how you actually talk. Mine is the first thing every agent reads. Here is how to write yours.
My agents did not get useful when the models got better. They got useful the day I wrote one file they read before doing anything. Before that, every session started from zero and I re-explained the business to a machine that had forgotten it by morning.
The file has two halves. The first is the company operating file: the business in four sentences, the north star number, the money ladder, who does what, the standing priorities, and a map of where every other fact lives so no fact is copied twice. The second is the judgment file: the number I check first, the rules I do not break, the mistakes that changed how I work, and three real sentences the way I actually say them.
The word I use for it is operating system, which I took from the way Alex Hormozi lays out a business as connected layers that all feed one number. The file is that: every layer, on one page, feeding the number, written so a stranger or an agent could step in tomorrow.
This guide is how to write it, how to keep it true, where to load it, and the one test that tells you whether it is doing anything. The judgment half on its own, with the gate that lets an agent decide like you without acting like you, has its own guide, Give Your Agent Your Judgment, and the interview behind it is a free skill I published, Give Your Agent Your Judgment. This guide is the whole file, both halves on one page, and how it runs the business.
The three levels
The business lives in your head and in old emails. Every AI session starts with you explaining it again, and every answer sounds like a LinkedIn post.
One file: the business in four sentences, the number, the rules, the mistakes, how you talk, and pointers to where everything else lives. Loaded before every conversation. Answers start from your judgment.
Every agent reads the file on start. Facts changed during a session get written back to the file, dated. A monthly review removes any rule that stopped being true. You can hand the file to a new hire or a new agent and they run the business the same way.
The mental model
Give the agent a biography and a mission statement and hope it picks up the rest from context.
Give it the business in four sentences, the number, the rules with the time each got tested, how you talk, and a pointer to every other fact. Judgment, not credentials.
| Role | Talks to you | Job |
|---|---|---|
| The operating half | Top of the file | Four sentences, north star, offers, roles, priorities, where truth lives |
| The judgment half | Below it | Numbers you check first, rules, mistakes, contrarian calls, how you decide, how you talk |
| The map | Bottom | One home per fact. Pointers, never copies. |
| The date | Top line | Last full audit. Stale and marked beats wrong. |
Write it in one sitting
1. The business in four sentences
Who you are and who you serve. What you sell today and what the money ladder looks like from free to the top offer. The one number the business is aimed at. The long term bet. If a stranger read only these four sentences, they should be able to tell whether a decision moves the business toward the number or away from it. Mine took longer to write than the rest of the file.
2. The map, not the copies
List where each kind of fact lives: offers and pricing, clients and how they are doing, how a recurring job is done, what every output should look like, project status, why things were built the way they were, secrets. One home per fact. The file points; it does not duplicate. The day I found the same price in two places with two values was the day this rule got written.
3. The judgment interview
Seven questions, one at a time, pushing on vague answers. The number you check first and its threshold. A rule you would never break and what happened the one time it broke. A mistake that actually taught you something, with what it cost. Something most people in your field get wrong. A time you decided with incomplete information and what you ignored on purpose. The question you ask before saying yes. Three real sentences you have said to a client. I will not repeat the whole interview here; the free skill runs it for you and the judgment guide covers the gate that goes above it. What this file adds is the operating half around it, so the judgment has a business to act on.
4. Compress, do not summarize
For each answer keep the rule or number stated plainly, the one real example that proves it in two sentences, and when it applies and when it does not. A rule with no exception is a rule nobody should trust. Cut anything that is credentials or backstory with no decision attached. A fact about you is not judgment unless it changes what someone should do next.
5. Standing priorities and the date
Three to five priorities, ranked, re-ranked at a monthly review. The date of the last full audit on the first line. Every status claim inside carries a date. When reality changes, change the file the same day, and write the decision and the reason in the decisions section so the next agent does not relitigate it.
6. Load it everywhere, then test it
A Claude Project's instructions and files. The config file Claude Code reads on start. Your chief of staff's instructions if you run the org chart from the Grok Bot guide. A skill folder any agent loads. Then the test: paste a real scenario from last week into a fresh session with the file and without. If the two answers sound the same, the file is still too generic. Go back for more specific stories, not more principles.
Starter prompts
Paste these as written. They are short on purpose, because the long ones drift.
Act as the operator who will run my business from one file. Interview me one question at a time and push back on vague answers. Part one, the operating half: (1) who I am and who I serve; (2) what I sell today and the money ladder from free to the top offer, with prices only if I give them; (3) the one number the business is aimed at; (4) the long term bet; (5) who does what, by name and role; (6) the three to five standing priorities, ranked; (7) for each kind of fact, offers, clients, how recurring jobs are done, what outputs look like, project status, decisions, secrets, where it lives. Part two, the judgment half: (8) the number I check first and its threshold; (9) a rule I never break and the time it broke; (10) a mistake that taught me something and what it cost; (11) something most people in my field get wrong; (12) how I decided with incomplete information, and what I ignored on purpose; (13) the question I ask before saying yes; (14) three real sentences I have said. Then write the file: the business in four sentences first, a last audit date on the first line, every status dated, each rule with its one example and its exception, a map of pointers with no copied facts and no secret values, and a rule that says write back before the session ends. Leave any answer I did not give blank. Do not invent anything.
Here is a real scenario from my business this week: [paste]. Answer it twice. First, with only general best practice. Second, with my one file loaded: [paste file]. Then list every place the two answers differ and which of my rules or numbers caused the difference. If they do not differ, tell me which question in the file needs a more specific story.
During this session the following changed: [what]. Update my one file: change the fact in its one home, date it [date], add a one line decision with the reason if a choice was made, and re-rank the standing priorities only if this changes them. Show me the diff, not the whole file. Do not touch any other line.
Pasted at the top of a chat, then read on start
Layer 1 is the file pasted at the top of a conversation. It works in any chat app and it is enough to change the answers. Layer 2 is the file living where every agent reads it without being asked, and being written back to when a fact changes. These are the settings.
One canonical copy, synced
A vault or a repo that every machine and agent can read. Copies drift; a synced file does not.
Read on start, by name
Claude Code's project config file points at it. A Project attaches it. The chief of staff's instructions open with read this file before acting.
Write back is a rule in the file
Learned something durable in a session, a decision, a client fact, a broken assumption? Write it to the file before the session ends. Dated.
Secrets are never in it
The file points at where secrets live. Values are never copied into it, never printed, never committed.
Monthly review removes, it does not defend
A rule that stopped working gets deleted. The file is only worth loading while it is true.
Give it to your agent, three ways
Same skill, three worlds. Pick the one you actually use. The skill file at the bottom of this page is the instructions in every case.
An agent with connectors (Claude Desktop, Claude Code, Grok Bot)
No connector is needed to write the file. What matters is where it lives so Claude reads it on start. Projects attach it; Claude Code reads a config file that points at it; a skill folder loads it by name.
- Paste the skill at the bottom of this page as project instructions, say the trigger, and answer both interviews with specifics.
- Save the returned file in your synced vault or repo. Attach it to your Claude Project.
- In Claude Code, add the line below to the project config so every session reads it first.
- Run the with and without test on a real scenario before you trust it.
# CLAUDE.md, read on start cat >> CLAUDE.md <<'EOF' Read ~/vault/Company OS/00 START HERE.md before answering any question about the business. If a fact changed this session, write it back there, dated, before you finish. EOF
ChatGPT (a project or a custom GPT)
A ChatGPT Project holds the file the same way. Custom GPTs can carry it for a single role, like the chief of staff.
- Create a Project, paste the skill as instructions, run both interviews.
- Upload the returned file to the Project's files and add read this file before answering to the instructions.
- For a role specific agent, make a custom GPT with the file in its knowledge and the role in its instructions.
- Re-upload after every change. There is no write back in ChatGPT; you are the write back.
Anything with an API (a token and a curl call)
If you run agents from code, the file goes in the system prompt or is fetched from the vault at start. The write back is a commit.
- Fetch the file from the synced repo at session start and prepend it to the system prompt, with the last audit date visible.
- Keep secrets out: the file contains pointers, and the code resolves them from the environment.
- When the agent learns a durable fact, append a dated line to the file and commit with the reason.
- A monthly job lists every dated status older than the review interval so the human re-ranks.
# read the one file into a system prompt, secrets resolved from env, never from the file
SYSTEM="$(cat "$VAULT/Company OS/00 START HERE.md")"
# statuses older than 30 days, for the monthly review
grep -Eo '20[0-9]{2}-[0-9]{2}-[0-9]{2}' "$VAULT/Company OS/00 START HERE.md" | sort -u | awk -v cutoff="$(date -v-30d +%F)" '$1 < cutoff'Failure modes
Every one of these has happened to me or to someone I set this up for.
| Failure | Fix |
|---|---|
| The file reads like a LinkedIn bio | Cut everything with no decision attached. Add the mistake and what it cost |
| With and without the file, the answers match | More specific stories, not more principles |
| The same price in two places with two values | One home per fact. The file points; delete the copy |
| An agent relitigated a decision you made in March | The decision and its reason were never written down. Log them the same day |
| A rule in the file that nobody follows anymore | Monthly review removes it. The file is only worth loading while it is true |
| An API key in the file, synced to three machines | Pointers only. Rotate it now |
| The agent invented a rule you never said | Leave blanks blank. Never let it fill in your judgment |
The tools I use for this
| Tool | What it is for here | |
|---|---|---|
| Claude | Runs the interviews, compresses the file, and reads it on start in Projects and Claude Code. | Open |
| Obsidian | Where my file lives. Synced to every machine, read by every agent. | no link, just use it |
| Wispr Flow | How I answer the interview. Speaking gets the real stories out; typing gets the polished ones. | Get it |
| Give Your Agent Your Judgment | The free interview skill for the judgment half of the file. | Get the skill |
The free skill
It builds the file in two passes. First it interviews you for the operating half: the business in four sentences, the north star number, the offers, the roles, the priorities, and where each fact lives. Then it runs the judgment interview: the number you check first, the rules you do not break, the mistake that taught you something, the contrarian call, how you decide with incomplete information, the filter before yes, and three real sentences. It compresses, dates it, and refuses to invent anything you did not say.
---
name: one-file-judgment
description: Builds the one file an agent reads before acting on an owner's behalf. Two interviews: the operating half (the business in four sentences, the north star number, offers, roles, priorities, where every fact lives) and the judgment half (numbers checked first, rules never broken, mistakes that taught something, contrarian calls, how they decide, how they talk). Compresses, dates it, keeps secrets out, and refuses to invent anything the owner did not say. Trigger on "build my one file", "write my company operating file", "make my agent run the business like me", "turn my experience into a file".
---
# Turn Your Experience Into One AI File
Everyone's agent runs on the same models. The file is the only thing that makes this
one answer like the owner. It has two halves and a map. You interview, compress and
write. You never invent judgment on the owner's behalf: a question with no real answer
stays blank.
## Part one: the operating half (interview, one question at a time)
1. Who they are and who they serve.
2. What they sell today and the money ladder from free to the top offer. Prices only if
given.
3. The one number the business is aimed at.
4. The long term bet.
5. Who does what, by name and role.
6. The three to five standing priorities, ranked.
7. For each kind of fact, where it lives: offers and pricing, clients and their status,
how recurring jobs are done, what outputs should look like, project status, decisions
and why, secrets. One home per fact.
From 1 to 4, write **the business in four sentences**. Test: could a stranger read only
those four and say whether a decision moves the business toward the number or away
from it? If not, rewrite them before continuing.
## Part two: the judgment half (interview, push on vague answers)
Push back the way you would on a new hire who says "I just know it when I see it". Do
not continue until the answer has a specific number, name, date or example.
8. The number they check first, with its threshold.
9. A rule they never break, and what happened the one time it broke.
10. A mistake that actually taught them something: the decision, what it cost, what
they do differently now.
11. Something most people in their field get wrong.
12. A time they decided with incomplete information: what they weighed, what they
ignored on purpose.
13. The question they ask before saying yes to new work.
14. Three real sentences they have said to a client or on camera. Not a description of
tone.
## Compress, do not summarize
For each judgment answer keep only:
- The rule or number, stated plainly.
- The one real example that proves it, in one or two sentences.
- When it applies and when it does not. A rule with no exception is one nobody should
trust.
Cut anything that is credentials or backstory with no decision attached.
## Write the file
```
# [Business] Operating File
Last full audit: [date]
## The business in four sentences
## North star
## Standing priorities (ranked, re-ranked monthly)
## Who does what
## Where truth lives (pointers, never copies; secrets never here)
## The numbers I check first (with thresholds)
## Rules I do not break (each with the time it was tested, and its exception)
## What most people in my field get wrong
## Mistakes that taught me something
## How I decide under uncertainty
## How I talk (three to five real sentences)
## Decisions log (date, decision, reason)
## Maintenance: one home per fact; every status dated; write back before the session ends; monthly review removes what stopped being true
```
Every status claim carries a date. Secret values never appear; the file points at where
they live.
## Load and test
Tell the owner where it goes: a Project's instructions and files, the config file a
coding agent reads on start, the chief of staff's instructions, a synced vault. Then the
test: one real scenario from last week, answered in a fresh session with the file and
without. If the answers sound the same, the file is too generic. Go back for more
specific stories, not more principles.
## Write back
When a session changes a fact, update it in its one home, dated, and log the decision
with its reason the same day. Show the diff, not the whole file. Touch nothing else.
## Rules
- Never invent judgment. Blank beats guessed.
- Specific beats impressive. One real number beats three principles.
- No one else's private data, no benchmark that is not theirs, no competitor's numbers
presented as their own.
- The file is only worth loading while it is true. A rule that stops working gets
removed, not defended.
What done looks like at thirty days
- The file exists, dated, with the business in four sentences at the top
- Every agent you run reads it on start, and you have stopped re-explaining the business
- The with and without test produced different answers, and the difference was your rule
- At least three facts changed this month were written back, dated
- A decision was logged with its reason and an agent did not relitigate it
- The first monthly review removed one thing that stopped being true
Want the file written with you and loaded into your agents?
Inside the AI CEO Lab the first session is this file: your business in four sentences and your judgment, written and loaded so every agent you run starts from you.