Agent Skills

A Skill is a folder. Inside it are instructions, and optionally scripts and resources. Claude scans the skills available to it, and loads the ones relevant to what you have asked for.

That last part is the design insight: skills are loaded on demand, not always in context. Ten skills do not cost you ten skills’ worth of tokens on every request. Only the relevant one gets read.

Why this is a category and not a Claude Code feature

Skills work in four places, which is unusual:

Where	How
Claude apps	Available to Pro, Max, Team and Enterprise. Claude invokes relevant skills automatically.
Claude Code	Installed from plugin marketplaces or manually into ~/.claude/skills.
Claude API	Managed through a /v1/skills endpoint and used from the Messages API, with the code execution tool providing the sandbox.
Agent SDK	The same skill support, for custom agents.

They have also been published as an open standard for cross-platform portability, documented at agentskills.io. A skill is meant to be a portable artefact, not a Claude-specific config file — which is why this reference treats Skills as its own entity rather than a sub-page of Claude Code.

What it does not do

• It does not grant access. A Skill is instructions. If the task needs to reach an external system, that is MCP or a connector. See Skills vs MCP.
• It is not a plugin. A plugin is a distribution container that may contain skills, along with commands, hooks and MCP servers. See Plugins vs Skills.
• It is not CLAUDE.md. CLAUDE.md is always-on project context. A Skill is conditional, task-triggered procedure.
• Not available on the Free plan in the Claude apps.

Authoring

There is a skill-creator that walks you through setup interactively and generates the folder structure and files, so writing the format by hand is optional. Anthropic maintains an anthropic/skills marketplace as the reference source of prebuilt ones.

The mental model that makes this click

CLAUDE.md is what Claude always knows. A Skill is what Claude knows how to do when it comes up. MCP is what Claude can reach. Permissions are what Claude is allowed to touch.

Four different questions. Four different mechanisms. Most confusion in this part of the ecosystem is someone trying to use one to solve another’s problem.