AI skills are modular, markdown-based instruction sets that give AI applications step-by-step procedures for performing specific tasks. Using a progressive disclosure pattern, an AI app initially loads only the name and description of each skill (under 100 tokens each), then fetches the full body (up to 5,000 tokens) only when the LLM determines it needs that skill. Skills can reference MCP server tools, APIs, and CLI commands to carry out their steps. A practical demo shows a skill that fetches the 10 most recent YouTube videos, executed both via a custom Python harness calling a local LLM and via GitHub Copilot in VS Code. The agentskills.io specification defines the required file structure (skill.md inside a named subdirectory) and optional metadata fields like compatibility, license, and allowed tools.