CLAUDE.md generator

Answer eight questions and get a lean starter CLAUDE.md your coding agent will actually follow — built on the rules that survive contact with a real agent, not a wish list it learns to skim past.

✓ Runs entirely in your browser. Nothing you type is uploaded, stored or seen by us.

Project name

Package manager / toolchain

What is it, in one line?e.g. "REST API for order management, TypeScript + Fastify + Postgres"

Test commandThe single command that proves nothing broke

Lint / typecheck command

Build command

Dev / run command

Key directories, one per lineFormat: path — what lives there. Only ones the agent genuinely needs; skip the obvious.

Non-obvious conventions, one per lineOnly rules an outsider would get wrong. "Write clean code" is skimmed; "errors go through AppError, never raw throw" is followed.

Never touch, one per lineFiles, directories or actions that are off-limits

—≈ tokens

—lines

—rules

Next: a CLAUDE.md decays as the project moves. When yours has grown past a screen, run it through the free CLAUDE.md audit for token cost, contradictions and dead rules. And if you're budgeting context across a whole agent setup — memory files, handoffs, subagents — that's the agent context engineering skill ($9, by us).

Why generated, not copied

Most CLAUDE.md templates fail the same way: they're pasted from someone else's project, so they describe someone else's build commands and someone else's conventions. The agent reads confidently wrong instructions, follows them, and the human concludes the file format doesn't work.

The file only earns its context window when everything in it is true, specific and checkable. That's why this generator asks for your actual commands and your actual conventions, and why it leaves out anything you didn't provide rather than padding with plausible defaults.

What makes the cut

Commands first. The highest-value lines in any CLAUDE.md are the exact commands for test, lint, build and run. The agent uses them dozens of times a session; getting them from the file beats rediscovering them from package.json every time.

Only non-obvious rules. An instruction the agent would have followed anyway is pure token cost. "Use meaningful variable names" survives in no serious file; "table names are singular, migrations are append-only" earns its place because it's the kind of thing an outsider gets wrong.

Hard boundaries stated as boundaries. "Be careful with migrations" is a mood. "Never edit shipped migrations — create a new one" is a rule an agent can actually obey.

Short. Every line is read at the start of every session, forever. A 2,000-token CLAUDE.md costs its weight on every single prompt. Start minimal and add rules only when the agent gets something wrong twice.