A small, opinionated CLAUDE.md starter for running a monorepo of many small projects built with an AI coding agent (Claude Code, or any agent that reads a project instructions file).
It's the operating system for "I keep starting new little projects and they turn into a mess." One repo, every project in its own folder, a type and a README each, predictable structure the agent maintains for you.
When you build a lot of small things with an agent, the bottleneck isn't code — it's entropy. Files land in the root, folders get named misc/ and phase-3/, two projects bleed into each other, and three weeks later you can't find anything. This file gives the agent a few hard rules so the repo stays navigable on its own.
The core ideas:
- One folder per project under
projects/, with a unique slug. - Every project has a type (
web/app/code/research) and aREADME.mdthat acts as its map. - Group by subject, not by phase —
competitors/, neverphase-2/. - Nothing in the repo root except
CLAUDE.mditself. - Git workflow by type — research goes straight to
main, code/apps go on feature branches. - Secrets never touch the chat and never get committed.
- Copy
CLAUDE.mdinto the root of your repo. - Edit the
<placeholders>at the top — your language, your skill level, your domain vocabulary. - Create
projects/anddocs/folders. - Start building. The agent picks up
CLAUDE.mdevery session and keeps the structure for you.
Adapt freely — delete the parts that don't fit, add your own. It's a starting point, not a law.
This is the portable subset. It leaves out anything tied to a specific personal setup — no server/VPS sync, no deploy remotes, no vendor-specific secret tooling. Just the project-keeping logic that works the same for everyone. If you later want server automation on top, that's a separate concern you can add yourself.
MIT — see LICENSE.