Recipe

How to Set Up the File That Tells Your AI How You Work

A 20-minute setup that stops your AI guessing and makes it work the way you do — every session, without you retyping a thing.

Before you point AI at anything that matters, spend 20 minutes writing the file that tells it how you work. It is called a CLAUDE.md, an AGENTS.md, or just "project instructions," depending on your app — but it is the same thing: a page of plain English your AI reads first, every session. Have the AI interview you and draft it. Put in your vocabulary, your review rules ("check the source, do not guess"), and one example of a good result. Now it works the way you do without being retrained every morning.

Total time

20 minutes

Difficulty

Easy

Makes

A one-page instructions file your AI reads every session

Tool

Claude or ChatGPT

Ingredients

Twenty quiet minutes
A few facts you would tell a new hire on day one (store names, what your key numbers mean, the rules you would never want broken)

The prompt

I want to set up an instructions file so you know how I work. Interview me one question at a time — about my business, the words and codes I use, the numbers I care about, and the checks you should run before you show me anything. When you have got enough, write the file for me in plain language.

The method

By now you have seen both versions of AI. Dialed in, it is the sharpest assistant you have ever had. Cold, it invents a number or calls one of your locations something nobody has ever called it — because every new chat starts from zero, knowing nothing about your business. The fix is a one-page file the AI reads before it does anything: your instructions file (a CLAUDE.md, an AGENTS.md, or just "project instructions," depending on where you work). Here is how to set one up in twenty minutes.

Find where it lives. In a chat app (Claude or ChatGPT), make a Project and use its instructions box. In Cowork with a connected folder, or a coding tool, save a plain-text CLAUDE.md (or AGENTS.md) in that folder. Same idea, three homes — do not get hung up on the ".md," it just means plain text.
Let the AI draft it. Paste the prompt above and answer its questions like you are briefing a sharp new hire.
Make sure three things are in it: your vocabulary (codes to real names, what your numbers mean), your rules (check the source, show your work, ask when unsure), and one example of a result you would be happy with.
Save it, then test. Open a brand-new chat and ask, "What do you know about how I work?" If it plays your vocabulary and rules back, it loaded. If it is generic, you saved it in the wrong place.

A few rules for writing the rules. A file that actually changes behavior has testable rules ("never give me a number without saying which report and row it came from"), tells the AI what to do when it is unsure ("stop and ask one question instead of guessing"), stays about a page with the non-negotiables first, and leaves secrets out. Best of all, grow it from real mistakes — add a line every time the AI does something you did not want, and six weeks in the file is a record of exactly what trips up AI in your world.

This is the first of three on the files that make your AI actually yours — next, stop it making things up and keep long sessions from getting dumb and expensive. The natural prequel is setting up a folder your AI works from. More on the Learn page.

On the plate

The good draft

A few lines from a good instructions file: “We run 8 taco shops. When I say a store by number (Store 12), use its real name (Mission & 3rd). ‘Labor’ means loaded labor including benefits, not just wages. Never give me a number without saying which report and row it came from. If you are not sure, stop and ask me one question instead of guessing. Lead with the answer, then the reasoning — plain sentences, no walls of bullets.”

The bad draft

The version that does nothing: “Be accurate. Be thorough. Use good judgment. Do a great job and be helpful.” Every line is untestable — the AI already thinks it is doing all of them. Rules it cannot check are rules it cannot follow.

Notes from the kitchen

CLAUDE.md, AGENTS.md, project instructions — which one do I need?

It depends only on where you work. In a chat app (Claude or ChatGPT), skip the file and use a Project's instructions box. Pointing Claude at a folder, or using a coding tool? Save a plain-text CLAUDE.md in that folder. AGENTS.md is the same thing under a different name, used by coding tools like Codex, Cursor, and Copilot. Match your app; the contents are identical.

Do I need to know how to code to make one?

No. The .md just means plain text — the same words you would type in an email. If a file feels like too much, use a Project's instructions box and skip the file entirely. Either way you are only writing sentences.

How is this different from pasting my instructions every time?

The file loads automatically at the start of every session, so you write it once instead of retyping it — and forgetting half of it — every time. It is the difference between training a new hire once and re-explaining the whole job every morning.

How long should the file be?

About a page to start. Begin with your vocabulary and your top three rules, then add a line each time the AI makes a mistake you do not want to see again. The best instructions files grow one caught mistake at a time.

Should everything go in one file, or one per project?

Think in two layers. Keep one short standing file for how you always want the AI to work — that loads everywhere. Put project-specific details in that project's own instructions, or carry them in a handoff note. Cramming one job's specifics into your permanent file is the fastest way to make it long, stale, and ignored.

Find where it lives. In a chat app (Claude or ChatGPT), make a Project and use its instructions box. In Cowork with a connected folder, or a coding tool, save a plain-text CLAUDE.md (or AGENTS.md) in that folder. Same idea, three homes — do not get hung up on the ".md," it just means plain text.
Let the AI draft it. Paste the prompt above and answer its questions like you are briefing a sharp new hire.
Make sure three things are in it: your vocabulary (codes to real names, what your numbers mean), your rules (check the source, show your work, ask when unsure), and one example of a result you would be happy with.
Save it, then test. Open a brand-new chat and ask, "What do you know about how I work?" If it plays your vocabulary and rules back, it loaded. If it is generic, you saved it in the wrong place.

About the author

By the Expo Kitchen. The Expo Kitchen is where Expo tests AI recipes for restaurant operators — plain-English, do-this-on-Saturday guides, written by operators, for operators. Browse more recipes and guides.