Commands Claude cannot infer safely
Install, test, lint, build, migration, and local service commands with required working directories.
Claude Code instructions
Give Claude durable, repository-specific context
A project CLAUDE.md is a Markdown instruction file loaded by Claude Code. Use it for commands, architecture, conventions, safety boundaries, and completion checks that apply across sessions.
A small project-level example
Place shared project instructions at ./CLAUDE.md or ./.claude/CLAUDE.md. Keep personal, cross-project preferences in the user-level file described by the official documentation.
# Project workflow
## Commands
- Install: `pnpm install`
- Develop: `pnpm dev`
- Test: `pnpm test`
- Lint: `pnpm lint`
## Architecture
- Routes live in `src/app/`.
- Shared UI lives in `src/components/ui/`.
- Database access goes through `src/data/`.
## Working agreements
- Preserve unrelated user changes.
- Validate external input at the boundary.
- Add a regression test when fixing a bug.
- Ask before changing a public API or data model.
## Completion
- Run the relevant tests and lint checks.
- Report changed files, verification, and known limitations.
What belongs in the file
Architecture that changes decisions
Ownership boundaries, source-of-truth modules, data flow, generated code, and intentional exceptions.
Rules with observable outcomes
Specific naming, validation, accessibility, security, review, and testing requirements.
References instead of duplication
Point to maintained documents and use imports for focused material rather than copying stale content.
Instructions shape behavior; they do not enforce policy
Keep the file concise and non-conflicting, then use permissions, hooks, protected branches, CI, and application-level controls for guarantees that must hold regardless of model behavior.