Desktop Step 5: Write a CLAUDE.md File So Your AI Remembers Your Project
Table of Contents
👋 Before you start: this post follows Step 3, Claude Code and Step 4, automatic deployment with GitHub. If you are still on the free Antigravity route, you can do this too: the file name
CLAUDE.mdis specific to Claude Code, but the idea works everywhere. Say each of the actions below in the Antigravity management panel instead, and at the end tell it "save these rules as a rules file you read automatically every time you start work." It will do the same thing with its own mechanism.
After using Claude Code for a while, you will run into this: in yesterday's conversation it clearly understood your project, and today you open a new conversation and it is a stranger again. You have to explain all over again what the project is for, that you cannot code, and that it should not randomly change the buttons.
The problem is not that the AI is stupid. Its memory only lives inside a single conversation. Close the conversation and the memory resets to zero.
The fix is almost absurdly simple: put a text file called CLAUDE.md in your project folder. Every time Claude Code starts work in that folder, it reads this file first, automatically. Think of it as an employee handbook: a new employee flips through it before every workday, so you do not have to interview them again every morning.
What goes in CLAUDE.md
Four kinds of things, in order of importance:
1. What this project is and who it is for
This is my services website, for clients looking to hire me for graphic design. The key thing is that it looks good on phones, because 90% of my clients come in from Instagram.
2. Who I am and how to talk to me
I cannot code. Explain everything in plain language, no jargon. Before you change anything, tell me in one sentence what you are about to change.
3. Hard rules (things it must never do)
Do not change any prices. Only I change prices. Do not delete any image files. Before uploading to GitHub, check for anything that should not be public, such as API keys.
4. How we usually do things
"Save and upload" means commit and push. Every time you change the layout, show it to me once at phone size.
That is it. It does not need to be long; ten lines is plenty to start. Vague adjectives ("be professional," "make it look good") do nothing. Concrete rules and facts are what work.
How to create it: have it write its own (10 minutes)
You do not even have to do it yourself. In the project folder, type claude, then say:
Create a CLAUDE.md for this project. First look through the project's files, then interview me with questions: who the project is for, my technical level, what must never be touched, and what I often ask you to do. Write my answers up as CLAUDE.md, in English.
It will look through the project while interviewing you and write a first version. Read it through, and if something does not sound like you, tell it to change it.
You can also take the lazy route with the built-in command: type /init directly in Claude Code and it scans the project and generates one automatically. Once it is done, remember to add categories 2 and 3 above (who you are, hard rules). Only you know those two.
Check that it works: the new-conversation test
Once it is written, test whether it took effect. "Open a new conversation" concretely means: close the terminal window completely, open a terminal in the project folder again, and type claude. Then just ask:
What is this project for? What are my rules?
If it can answer, you are done. From today on, every new conversation starts with the backstory already loaded.
When to update it: the second-correction rule
CLAUDE.md is not something you write once and leave. There is only one maintenance rule:
The second time you correct it on the same thing, write that thing into CLAUDE.md.
The first correction is communication. The second correction is a pattern. Tell it: "Add this to CLAUDE.md: from now on, when changing fonts, only change headings, leave body text alone." It will add it itself. That way the handbook grows along with your project, and every rule in it comes from a mistake you actually hit.
The next step up: from using it yourself to teaching others
At this point you have the core kit for the desktop route: an AI workshop (Antigravity), a real piece of work, a terminal (Claude Code), automatic publishing (GitHub + Vercel), and project memory (CLAUDE.md). You are now a small studio of one person plus a team of AI.
If you start wanting to teach this setup to others (colleagues, students, family), the next post moves into the Work Systems series: The right way to teach AI tools: do 90% of it for your students, on how to package your environment into a starter kit so the next person can get going in 10 minutes.
FAQ
What is a CLAUDE.md file?
CLAUDE.md is a text file you put in your project folder. Every time Claude Code starts work in that folder, it reads this file first, automatically, so you do not have to re-explain your project in every new conversation. Think of it as an employee handbook for your AI.
What should I put in CLAUDE.md?
Four kinds of things, in order of importance: what the project is and who it is for; who you are and how to talk to you; hard rules, meaning things it must never do; and how you usually do things. Ten lines is plenty to start. Write concrete rules and facts, because vague adjectives like 'be professional' or 'make it look good' do nothing.
How do I create a CLAUDE.md? Can the AI write it for me?
Yes. Type 'claude' in the project folder and ask it to look through the project's files first, then interview you: who the project is for, your technical level, what must never be touched, and what you often ask it to do. It writes your answers up as CLAUDE.md. You can also type '/init' to have it scan the project and generate one automatically; once it is done, add who you are and your hard rules, since only you know those two.
How do I check that CLAUDE.md is working?
Run the new-conversation test: close the terminal window completely, open a terminal in the project folder again, type 'claude', and ask 'What is this project for? What are my rules?' If it can answer, you are done.
When should I update CLAUDE.md?
Use the second-correction rule: the second time you correct it on the same thing, write that thing into CLAUDE.md. Tell it 'Add this to CLAUDE.md' and it will add the rule itself.