name: claude-code
description: Run Anthropic's claude CLI for agentic coding. Supports single-shot commands (-p) and interactive sessions. Use this for complex refactoring, bug fixing, or when you need Claude 3.7 Sonnet's specialized coding agent capabilities.

Claude Code Skill 🤖

This skill integrates the official claude CLI tool into OpenClaw, allowing you to leverage Anthropic's specialized coding agent.

🛠 Prerequisites

• Install: npm install -g @anthropic-ai/claude-code
• Auth: Run claude auth login once (interactive) or ensure ANTHROPIC_API_KEY is set.

🚀 Usage Patterns

1. Single-Shot Task (Best for Agents)

Use the -p (prompt) flag to execute a task and exit. This is the preferred method for OpenClaw automation.

# Basic usage
claude -p "Refactor src/utils.js to use arrow functions"

# Auto-approve (YOLO mode) - WARNING: Risks applying changes without review
claude -p "Fix the lint errors in app.tsx" --dangerously-skip-permissions

Tips:
- Always be specific about file paths.
- Combine with --dangerously-skip-permissions if you want the agent to edit files without asking (use with caution).

2. Interactive Session

Start a persistent session to discuss code or run multiple steps.

// Start the session (must use pty=true)
exec({ command: "claude", pty: true });

// Send commands
process({ action: "write", data: "/bug Fix the crash in main.rs\n" });

⚙️ Configuration

• Compact Mode: Use --compact to reduce output noise.
• Verbose: Use --verbose for debugging.

📋 Example Workflow

1. Read Context: The agent reads files.
2. Delegate: Agent calls claude -p "Analyze these files and fix X".
3. Review: Agent checks the changes made by Claude Code.