Documentation

What are Claude Agent Skills?

A Claude Agent Skill is a SKILL.md file that teaches Claude how to do something β€” domain expertise, conventions, workflows, best practices β€” bundled in plain text.

How Skills Work

  1. 1 You place a SKILL.md file in your project (or your ~/.claude/skills/ directory).
  2. 2 Claude reads it at the start of every session β€” absorbing the skill's knowledge without repeated prompting.
  3. 3 Skills use lazy loading: Claude first sees just the name and description, then reads the full file only when relevant β€” saving tokens.

Examples

  • β–ΈA Laravel skill that knows your app's architecture, naming conventions, and common patterns.
  • β–ΈA security auditing skill with OWASP checklists and code review procedures.
  • β–ΈA brand guidelines skill so Claude always writes in your company's voice.

Lazy Loading

Skills are token-efficient. Claude first sees just the skill's name and one-line description. It only reads the full SKILL.md when the current task is relevant β€” so you can have many skills installed without bloating your context.

Skills vs MCP Tools

Skills and MCP (Model Context Protocol) tools are complementary β€” they solve different problems.

Skills

  • βœ“Provide knowledge, expertise, and workflows
  • βœ“Plain text β€” no code execution
  • βœ“Work on web, desktop, and CLI
  • β†’Best for: conventions, domain knowledge, writing style

MCP Tools

  • βœ“Connect to APIs, databases, and external tools
  • βœ“Execute code and make network calls
  • βœ“Require desktop or CLI (not web)
  • β†’Best for: real-time data, file operations, integrations