/init  # Claude analyzes the codebase and generates a draft

# Project: [Project Name]

## Tech Stack
- **Frontend**: React + TypeScript
- **Backend**: Node.js + Express
- **Database**: PostgreSQL
- **ORM**: Drizzle

## Coding Standards
- Use TypeScript strict mode
- Prefer server components over client components
- Use `async/await` instead of `.then()`
- Always validate user input with Zod

## DO NOT
- Never commit `.env` files
- Never use `any` type in TypeScript
- Never bypass authentication checks
- Never expose API keys in client code

## Common Commands
- `npm run dev`: Start development server
- `npm run build`: Build for production
- `npm run test`: Run tests

Our authentication system is built using NextAuth.js, which is a
complete authentication solution for Next.js applications...
(5+ lines of explanation)

## Authentication
- NextAuth.js with Credentials provider
- JWT session strategy
- **DO NOT**: Bypass auth checks, expose session secrets

// ~/.claude/settings.json
{
  "hooks": {
    "PreToolUse": [
      {
        "pattern": "rm -rf /",
        "action": "block",
        "message": "Block root directory deletion"
      },
      {
        "pattern": "rm -rf /*",
        "action": "block",
        "message": "Block dangerous deletion command"
      },
      {
        "pattern": "sudo rm",
        "action": "warn",
        "message": "Caution: sudo delete command"
      },
      {
        "pattern": "curl * | sh",
        "action": "block",
        "message": "Block piped script execution"
      },
      {
        "pattern": "chmod 777",
        "action": "warn",
        "message": "Caution: excessive permission setting"
      }
    ]
  }
}

Rules that always apply → Project Description File
Knowledge needed only for specific tasks → Skills (token efficient)
Frequently used commands → Slash Commands
External service integration → Plugins / MCP

# Create skill directory
mkdir -p ~/.claude/skills/my-skill

# Write SKILL.md
cat > ~/.claude/skills/my-skill/SKILL.md << 'EOF'
---
name: my-skill
description: My custom skill
platforms: [Claude, Gemini, ChatGPT]
---

# My Skill

## When to use
- When needed for specific tasks

## Instructions
1. First step
2. Second step
EOF

# Check for dangerous commands with cc-safe tool
npx cc-safe .
npx cc-safe ~/projects

# Detection targets:
# - sudo, rm -rf, chmod 777
# - curl | sh, wget | bash
# - git reset --hard, git push --force
# - npm publish, docker run --privileged

# Auto-approve only safe commands
/sandbox "npm test"
/sandbox "npm run lint"
/sandbox "git status"
/sandbox "git diff"

# Pattern approval
/sandbox "git *"       # All git commands
/sandbox "npm test *"  # npm test related

# MCP tool patterns
/sandbox "mcp__server__*"

project/
├── .claude/                    # Claude Code settings
│   ├── team-settings.json
│   ├── hooks/
│   └── skills/
├── .agent-skills/              # Universal skills
│   ├── backend/
│   ├── frontend/
│   └── ...
├── CLAUDE.md                   # Project description for Claude
├── .cursorrules               # Cursor settings
└── ...

{
  "permissions": {
    "allow": [
      "Read(src/)",
      "Write(src/)",
      "Bash(npm test)",
      "Bash(npm run lint)"
    ],
    "deny": [
      "Bash(rm -rf /)",
      "Bash(sudo *)"
    ]
  },
  "hooks": {
    "PreToolUse": {
      "command": "bash",
      "args": ["-c", "echo 'Team hook: validating...'"]
    }
  },
  "mcpServers": {
    "company-db": {
      "command": "npx",
      "args": ["@company/db-mcp"]
    }
  }
}

Commit .claude/ folder → Team members Clone → Same settings automatically applied → Team standards maintained

.agent-skills/
├── backend/
├── frontend/
├── code-quality/
├── infrastructure/
├── documentation/
├── project-management/
├── search-analysis/
└── utilities/

~/.claude/settings.json     # Global settings
~/.claude/skills/           # Global skills
.claude/settings.json       # Project settings
.claude/skills/             # Project skills
.agent-skills/              # Universal skills
CLAUDE.md                   # Project AI manual

1. Block dangerous commands with Hooks
2. Auto-approve only safe commands with /sandbox
3. Regular audit with cc-safe
4. Experiment mode in containers only

Project Description File: Always loaded (keep concise)
Skills: Load on demand (token efficient)
.toon mode: 95% token savings