Claude Md

CLAUDE.md Skill

Create effective CLAUDE.md files that transform Claude Code from a general-purpose assistant into a specialized tool for specific codebases.

Core Philosophy

CLAUDE.md is Claude Code's "constitution" - it provides persistent context automatically loaded at conversation start. Every token counts because CLAUDE.md competes for context window space with conversation history, file contents, and tool results.

Golden Rule: Document what Claude gets wrong, not everything Claude might need. Start small, expand based on friction.

CLAUDE.md Hierarchy

Claude Code loads CLAUDE.md files from multiple locations (all are merged):

Creation Workflow

1. Analyze the Codebase

Before writing, understand what Claude will struggle with:

# Identify key files and structure
find . -maxdepth 3 -type f -name "*.md" | head -20
ls -la package.json pyproject.toml Makefile 2>/dev/null

# Check for existing documentation
cat README.md 2>/dev/null | head -50

# Identify build/test commands
grep -E "scripts|test|build|dev" package.json 2>/dev/null

2. Structure Template

Use this proven structure (adapt sections as needed):

# CLAUDE.md

[One-line project description]

## Quick Start

```bash
[Package manager] install
[Dev command]
[Test command]

Architecture

[2-3 sentences on structure, key patterns, main technologies]

Key Directories

• src/ - [purpose]
• tests/ - [purpose]

Code Standards

• [Style rule with example if non-obvious]
• [Critical pattern to follow]
• [Common mistake to avoid → correct approach]

Critical Rules

MUST:

• [Non-negotiable requirement]

MUST NOT:

• [Common mistake that breaks things]

Common Commands

### 3. Content Priorities

Include (in order of importance):

1. **Build/test/lint commands** - What Claude types most often
2. **Critical gotchas** - Things that break silently or waste time
3. **Architecture overview** - How components connect
4. **Code patterns** - Non-obvious conventions
5. **Workflow rules** - Branch naming, commit format, PR process

Exclude:

- Information Claude already knows (language syntax, common libraries)
- Verbose explanations (prefer examples)
- Comprehensive API docs (link to them instead)
- Anything unchanged from project defaults

## Writing Guidelines

### Conciseness Patterns

**Bad (verbose):**
```markdown
When you are writing TypeScript code in this project, you should always 
make sure to use strict mode and avoid using the 'any' type because it 
defeats the purpose of TypeScript's type system.

Good (concise):

- TypeScript strict mode - no `any` types

Negative Rules Need Alternatives

Bad:

- Never use npm

Good:

- Use pnpm (not npm) - monorepo requires pnpm workspaces

Reference External Docs (Don't Embed)

Bad:

[500 lines of API documentation]

Good:

For complex widget usage or FooBarError, see docs/widgets.md

Emphasis for Critical Items

**IMPORTANT:** RLS policies auto-filter by user_id - don't add redundant .eq('user_id', userId)

Monorepo Strategy

For large codebases, use hierarchical CLAUDE.md files:

monorepo/
├── CLAUDE.md              # Shared: package manager, CI, overall arch
├── apps/
│   ├── web/CLAUDE.md      # Frontend-specific context
│   └── api/CLAUDE.md      # Backend-specific context
├── packages/
│   └── shared/CLAUDE.md   # Shared library context

Root CLAUDE.md should contain:

• Monorepo tooling (pnpm/nx/turborepo commands)
• Cross-package patterns
• CI/CD workflows
• Shared conventions

Subdirectory CLAUDE.md should contain:

• Package-specific commands
• Local testing patterns
• Package-specific gotchas

Size targets:

• Root: <500 lines, <10k words
• Subdirectories: <200 lines each
• Total loaded context: <15k tokens

Quality Audit Checklist

Run this audit on existing CLAUDE.md files:

• Actionable: Every item helps Claude do something concrete
• Non-obvious: Excludes things Claude already knows
• Current: Reflects actual project state (not aspirational)
• Tested: Commands actually work when copied
• Minimal: Each line earns its token cost
• Alternatives: Every "don't" has a "do instead"
• Scannable: Uses tables/bullets, not paragraphs
• Examples: Critical patterns show code, not just describe

Anti-Patterns to Fix

Iteration Workflow

CLAUDE.md improves through use:

1. During coding: Press # in Claude Code to add memories
2. After sessions: Review what Claude struggled with
3. Add sparingly: Only document repeated friction points
4. Prune regularly: Remove outdated or unused content

Post-Task Maintenance

After completing significant tasks, update CLAUDE.md:

What to Add

What to Update

• New dependencies - Add install commands
• New scripts - Document in commands table
• Architecture changes - Update directory descriptions
• Deprecated patterns - Remove or mark as legacy

Maintenance Checklist

After each significant task:

• Did Claude struggle with anything? → Document it
• Did I discover a new pattern? → Add to Code Standards
• Did commands change? → Update Quick Start
• Is anything now outdated? → Remove it

Learn from Mistakes

When Claude makes errors:

## Critical Rules

**MUST NOT:**
- Use `npm install` (breaks pnpm lockfile)  ← Added after Claude broke deps
- Import from `@/old-path` (migrated to `@/new-path`)  ← Added after refactor

Golden Rule: If you corrected Claude, future Claude should know too.

Advanced Patterns

Machine-Readable State

For complex projects, add parseable metadata:

## Project State
<!-- machine-readable -->
Version: 1.2.0
Status: development
Last migration: 0016_search_analytics

Conditional Workflows

## Workflow

**Creating new feature?**
1. Create branch from main
2. Implement with tests
3. Run full suite before PR

**Fixing bug?**
1. Write failing test first
2. Implement fix
3. Verify only that test changes

Performance Targets

## Metrics

| Metric | Target | Notes |
|--------|--------|-------|
| Cold start | <200ms | Measure with `time cmd` |
| API p95 | <500ms | Check after changes |

Output Requirements

When creating CLAUDE.md:

1. Analyze codebase first (don't guess structure)
2. Keep total length under 500 lines
3. Test all commands before including
4. Use markdown tables for reference data
5. Save to /mnt/user-data/outputs/CLAUDE.md

When auditing CLAUDE.md:

1. Report issues with specific line references
2. Provide concrete fixes, not vague suggestions
3. Prioritize by impact (token waste vs. broken workflows)