- Tool
- Claude Code
- Price
- Included in a paid Claude plan
- Time
- 15 min
- Level
- Beginner
Every Claude Code session starts with no memory of the last one. CLAUDE.md is its written memory: Anthropic's documentation suggests writing down whatever you would otherwise have to re-explain. For a freelancer, that is not only code: your offer, how you name your quotes, the tone of your client emails, what needs your approval.
Where to put the file
CLAUDE.mdat the root of your working folder (for example your business folder): the rules for that folder.~/.claude/CLAUDE.md: your personal preferences, applied in every folder.CLAUDE.local.md: private notes for this folder only, kept out of sharing (add it to.gitignoreif the folder is on GitHub).
Create a first version
- Open Claude Code in the folderIn the Terminal, go to your business folder and start Claude Code.
cd ~/Documents/my-business claude - Generate a draft with /initClaude looks through the folder and suggests a starting CLAUDE.md. If one already exists, it suggests improvements instead of overwriting it.
/init - Add what Claude cannot guessOpen the file with the
/memorycommand and fill it in. Here is a template to adapt with your own details:# My business ## Who I am - [First name], freelance [trade], clients: [type of clients]. - Offer and rates: see @offer.md ## Tone - Short, plain answers, no jargon. - Client documents: formal tone, no emoji. ## Where things go - Quotes: quotes/ (name: QUO-YYYY-NNN_Client.md) - Invoices: invoices/ (an issued invoice is never edited) - One folder per client: clients/<client-name>/ ## Always ask before - Sending an email, publishing, invoicing, deleting a file. ## Never - Write a password, a key or bank details in a file. - Make up legal information: write [TO COMPLETE]. - Check that it is loadedThe
/contextcommand lists the memory files loaded in the session. If yours is not there, Claude cannot see it./context
The rules that make the difference
- Concrete and checkable. "Quotes in
quotes/, numbered QUO-YYYY-NNN" works better than "be organised". - Short. Anthropic recommends under 200 lines per file: beyond that, it takes more room and is followed less reliably.
- No contradictions. If two instructions conflict, Claude may pick one arbitrarily. Reread the file now and then.
- Import instead of copying. The
@path/file.mdsyntax loads another file (your offer, your rate card). It still counts towards the space used. - Grow it as you go. When Claude makes the same mistake twice, add a line.
Good to know. CLAUDE.md is guidance, not a lock: Claude usually follows it, with no guarantee. To truly block an action, Claude Code offers permission settings and "hooks" (scripts that run before an action). And never put a password or an API key in it: the file is readable text, often shared.
Sources
- Claude Code docs: How Claude remembers your project
- Claude Code docs: Quickstart (installation and plans)
Want to go further, with an agent that drafts your quotes or follow-ups from these rules? That is what I build as AI agents.