Claude Skills

Claude Skills are modular capabilities that extend what Claude Code, Clawdbot, and other Claude-based agents can do. Each skill is a self-contained package that teaches an agent how to perform specific tasks—like controlling a camera, querying a database, or interacting with an API.

Think of skills like apps for your AI agent:

• The base agent understands language and can reason
• Skills give it specialized abilities and knowledge
• You install only the skills your agent needs

Skills follow a standard structure, making them easy to discover, share, and install.

Skills are defined by a SKILL.md file that contains everything the agent needs:

skills/
└── github/
    ├── SKILL.md           # Instructions for the agent
    ├── scripts/           # Helper scripts (optional)
    └── examples/          # Usage examples (optional)

SKILL.md structure:

• Metadata block: Name, description, required tools/env vars
• Description and triggers: When to use this skill
• Instructions: Step-by-step guidance for the agent
• Examples: Concrete usage demonstrations

Tool skills - Wrap external tools and CLIs (GitHub, Docker, cloud providers)

API skills - Connect to web services (weather APIs, calendar integrations, CRMs)

System skills - Control local capabilities (file system, camera, screen capture)

Workflow skills - Define multi-step processes (code review, deployment, publishing)

Platform skills - Integrate with platforms (Slack, Notion, email)

1. Identify the capability What should this skill enable? Be specific: "Query weather forecasts from Open-Meteo API" not "Do weather stuff."

2. Define requirements What does the agent need? External tools, API keys, file system access.

3. Write clear instructions Tell the agent exactly how to accomplish tasks with step-by-step guidance.

4. Include examples Show input/output pairs so the agent understands expected behavior.

5. Handle edge cases Document what to do when things go wrong—error handling is crucial.

Both extend agent capabilities, but differently:

Many agents use both: MCP for structured tool calling, skills for flexible guidance.