CLAUDE.md templates that pull their weight.

Each is fill-in-the-brackets: copy the closest one to your repo root as CLAUDE.md, replace the <angle brackets>, delete what doesn’t apply. Ten minutes, honestly.

Most CLAUDE.md files fail the same way: they describe the project instead of directing the agent. Claude can read your file tree; it cannot read your mind. The templates are built on one test applied to every line: would this change what Claude does?

## What doesn't work          ## What works

"We use pnpm in this          "Package manager is pnpm.
project and prefer            Never use npm or yarn."
consistent tooling."
                              "Typecheck: pnpm typecheck.
"Please try to keep           Run it after any multi-file
the code type-safe."          change."

"The tests can be a           "Tests hit a real local
little tricky to run."        Postgres: docker compose up -d
                              first, or they hang for 60s."

The highest-value lines are the gotchas. Every “the dev server caches schema, restart after editing db/” saves a full wasted debugging loop, every session, forever. When Claude does something wrong twice, that is not a prompting problem; that is a missing line in CLAUDE.md.

Claude reads the nearest CLAUDE.md as well as the root one, which is the clean way to keep workspace rules and package rules separate:

repo/
├── CLAUDE.md            ← workspace rules (this pack's monorepo template)
├── apps/
│   └── web/
│       └── CLAUDE.md    ← web-app-specific rules (nextjs template)
└── packages/
    └── ui/
        └── CLAUDE.md    ← component-library rules

The monorepo template is written for exactly this split: cross-package rules at the root (workspace commands, import boundaries, codegen), app-specific rules next to the app.