agent-configuration（Agent 配置）

AI Agent Configuration Policy (Configuration & Security)

When to use this skill

• Build AI agent environment for new projects
• Write and optimize project description files
• Configure Hooks/Skills/Plugins
• Establish security policies
• Share team configurations

1. Project Description File Writing Policy

Overview

Project description files (CLAUDE.md, README, etc.) are project manuals for AI. AI agents reference these files with top priority.

Auto-generate (Claude Code)

/init  # Claude analyzes the codebase and generates a draft

Required Section Structure

# 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

Writing Principles: The Art of Conciseness

Bad (verbose):

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

Good (concise):

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

Incremental Addition Principle

"Start without a project description file. Add content when you find yourself repeating the same things."

2. Hooks Configuration Policy (Claude Code)

Overview

Hooks are shell commands that run automatically on specific events. They act as guardrails for AI.

Hook Event Types

Security Hooks Configuration

// ~/.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"
      }
    ]
  }
}

3. Skills Configuration Policy

Skills vs Other Settings Comparison

Selection Guide

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

Custom Skill Creation

# 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

4. Security Policy

Prohibited Actions (DO NOT)

Absolutely Forbidden

• Using unrestricted permission mode on host systems
• Auto-approving root directory deletion commands
• Committing secret files like .env, credentials.json
• Hardcoding API keys

Requires Caution

• Indiscriminate approval of sudo commands
• Running scripts in curl | sh format
• Setting excessive permissions with chmod 777
• Connecting to unknown MCP servers

Approved Command Audit

# 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

Safe Auto-approval (Claude Code)

# 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__*"

5. Team Configuration Sharing

Project Configuration Structure

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
└── ...

team-settings.json Example

{
  "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"]
    }
  }
}

Team Sharing Workflow

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

6. Multi-Agent Configuration

Per-Agent Configuration Files

Shared Skills Directory

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

7. Environment Configuration Checklist

Initial Setup

• Create project description file (/init or manual)
• Set up terminal aliases (c, cc, g, cx)
• Configure external editor (export EDITOR=vim)
• Connect MCP servers (if needed)

Security Setup

• Configure Hooks for dangerous commands
• Review approved command list (cc-safe)
• Verify.env file in.gitignore
• Prepare container environment (for experimentation)

Team Setup

• Commit.claude/ folder to Git
• Write team-settings.json
• Team standard project description file template

Quick Reference

Configuration File Locations

~/.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

Security Priority

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

Token Efficiency

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