Claude Md Authoring

CLAUDE.md Authoring

Overview

CLAUDE.md is the most powerful lever for making Claude Code effective on your project. It's read at the start of every session — think of it as the briefing document Claude reads before doing any work. A well-written CLAUDE.md means Claude follows your conventions without being asked, and a missing or stale one means Claude reinvents conventions every session.

When to Use

• Starting a new project and want Claude Code to understand context from day 1
• Claude keeps asking the same questions or making the same mistakes
• Onboarding Claude Code to an existing codebase
• Wanting Claude to follow specific conventions consistently

Core Jobs

1. CLAUDE.md Hierarchy

Claude Code reads files in priority order:

1. ~/.claude/CLAUDE.md          (global — applies to all projects)
2. /project/CLAUDE.md           (project root — main project context)
3. /project/subdir/CLAUDE.md    (subdirectory — specific area rules)

Higher priority overrides lower. Use global for personal preferences, project for team conventions.

2. What to Put in CLAUDE.md

High-value content:

# Project: [Name]

## What This Is
[2-3 sentences: what the project does, who uses it, tech stack]

## Architecture
[Key architectural decisions Claude should know: monorepo, microservices, etc.]
- API: FastAPI on port 8000, PostgreSQL via SQLAlchemy
- Frontend: Next.js 14 App Router, Tailwind CSS, shadcn/ui
- Auth: Clerk (not custom auth — never implement auth yourself)

## Conventions
- Python: use `ruff` for formatting, type hints required on all functions
- Commits: conventional commits (feat:, fix:, chore:)
- Never use print() — always use the logger: `from app.core.logger import logger`

## Do Not
- Never modify migrations directly — always use `alembic revision --autogenerate`
- Never commit .env files
- Never use synchronous DB calls in async functions

## Current Focus
[Optional: what's being built right now, context for current sprint]

Low-value content (don't include):

• File-by-file documentation (Claude reads the code)
• Obvious conventions (PEP8, etc.)
• History of decisions (use ADRs for that)
• Todo lists (use task management tools)

3. Global CLAUDE.md (~/.claude/CLAUDE.md)

Personal preferences that apply across all projects:

## My Preferences
- I'm [role] working primarily in [languages]
- Communication style: concise, direct, minimal explanation unless asked
- When in doubt, ask before making large changes
- Always run tests before declaring done

## Tools I Use
- Package manager: uv (Python), pnpm (Node)
- Formatter: ruff (Python), prettier (JS/TS)
- Testing: pytest (Python), vitest (JS/TS)

## Workflow
- Read existing code patterns before suggesting new approaches
- Prefer editing existing files over creating new ones
- Commit frequently with conventional commit messages

4. Dynamic CLAUDE.md Updates

Update CLAUDE.md as the project evolves:

• After major architectural decisions: document the "why"
• After Claude makes a repeated mistake: add a "Do Not" rule
• After onboarding: add what was confusing (to help future Claude sessions)
• Before a sprint: add "Current Focus" section

Key Concepts

• CLAUDE.md — markdown file Claude Code reads at session start for project context
• Global CLAUDE.md — ~/.claude/CLAUDE.md — personal preferences across all projects
• Project CLAUDE.md — /project/CLAUDE.md — shared team conventions
• Subdirectory CLAUDE.md — /subdir/CLAUDE.md — area-specific rules (e.g., /frontend/CLAUDE.md)

Checklist

• Project CLAUDE.md exists in repo root?
• Tech stack and key dependencies documented?
• Critical "Do Not" rules added (the things Claude gets wrong)?
• Current focus/sprint context updated (not stale)?
• Global CLAUDE.md configured with personal preferences?
• Conventions documented with examples (not just rules)?

Key Outputs

• Project CLAUDE.md: architecture, conventions, do-nots, current focus
• Global CLAUDE.md: personal preferences, tools, workflow style

Output Format

• 🔴 Critical — no CLAUDE.md (Claude reinvents conventions each session), CLAUDE.md outdated by months
• 🟡 Warning — CLAUDE.md too generic (no project-specific content), missing "Do Not" rules for known pitfalls
• 🟢 Suggestion — add subdirectory CLAUDE.md for complex areas (frontend/, api/), update "Current Focus" weekly

Anti-Patterns

• Copying generic CLAUDE.md templates without project-specific content
• Documenting things Claude can read from the code (file structure, imports)
• "Current Focus" section that's 3 months old (stale context is worse than no context)
• No "Do Not" section (the most valuable part — captures hard-learned lessons)

Integration

• Use with claude-project-settings for complete Claude Code configuration
• Use with claude-hooks (document hook behaviors in CLAUDE.md so team understands them)
• Agent: @architect helps write the architecture section of CLAUDE.md