About Agent Skills
Agent Skills are modular "knowledge packages" that teach AI agents—such as Claude, Codex, and GitHub Copilot—to perform specialized tasks reliably and repeatably, such as following specific team coding standards or executing complex workflows.
How Skills Work
Defined primarily through a SKILL.md file containing
Markdown instructions, metadata, and optional scripts or resources, skills allow users to codify domain
expertise for the agent to reuse.
To maintain efficiency, skills utilize progressive disclosure: the agent initially loads only the skill's name and description, dynamically pulling the full instructions and context into the conversation only when the user's request matches the skill's purpose.
Skills vs MCP
Unlike MCP (Model Context Protocol), which connects agents to external tools and data, skills provide the specific "know-how" or best practices on how to use those tools effectively.
The Kitchen Analogy
You can think of an AI Agent as a master chef in a kitchen:
- 🔧 MCP (Tools) is the equipment: the ovens, blenders, and knives that allow the chef to physically manipulate ingredients.
- 📋 Agent Skills are the recipe cards: specific, step-by-step instructions that tell the chef exactly how to use that equipment to prepare a specific dish (like "Grandma's Lasagna") exactly the way you want it, every single time.
Supported Platforms
Claude Code
Anthropic's official CLI with native skill support via .claude/skills
OpenAI Codex
Supports skills via CLI and IDE extensions with .codex/skills
GitHub Copilot
Added skills support Dec 2025, reads from .claude/skills repos
OpenCode
Compatible with .opencode/skill and .claude/skills directories
Even More Platforms with Skilz
With the Skilz installer, you can use Agent Skills with 14+ different platforms, including:
Gemini CLI, Aider, Cursor, Windsurf, Qwen Code, KIMI Code, Zed AI, Plandex, and more.