AI library

Tooling

Agent skills

Last reviewed 4 October 2026

Overview

A skill is a folder of instructions, scripts, and reference files that teaches an agent how to do a specific job: fill in a PDF form, follow a company style guide, run a deployment. The agent sees only each skill's name and description until a task needs it, then loads the rest. Skills package know-how; MCP provides connections.

Key concepts

  • SKILL.md: the entry file, with frontmatter (name, description) and the instructions in Markdown.
  • Progressive disclosure: metadata always loaded, body loaded when relevant, extra files and scripts read or run only when needed. Many skills can be installed at little context cost.
  • Scripts: skills can include code the agent runs instead of generating it each time.
  • Where they run: Claude apps, Claude Code (project, user, and plugin skills), the Claude API (with code execution), and Managed Agents.
  • Plugins: bundles of skills, commands, hooks, agents, and MCP servers that can be shared through marketplaces (Claude Code plugins).

How to use it

  • Write the description carefully: it decides when the skill is used.
  • Keep SKILL.md focused; move long reference material into separate files.
  • Include scripts for steps that must be exact.
  • Test that the skill triggers on the right requests and not on the wrong ones.
  • Review third-party skills before installing: they can contain instructions and code.

Current state

  • 2026-10-04: The Skills API (/v1/skills) is generally available; the earlier beta header is no longer needed.

Open questions

  • Will a shared skill format be adopted across vendors, as MCP was?

Sources