Agent Skills are folders of instructions (and optional supporting files) that teach Claude Code how to perform a specific kind of task — like writing release notes, scaffolding an API route, or following a team's code style. Each skill is a directory containing a SKILL.md file whose YAML frontmatter gives the skill a name and description, followed by the procedure Claude should follow. The folder can also bundle scripts, templates, or reference docs so the agent does the job consistently.
Skills are progressively disclosed: Claude reads only each skill's name and short description until a task looks like a match, then loads the full SKILL.md (and any extra files) into context. This keeps the context window lean while still giving the agent deep, on-demand know-how. Because a skill can ship runnable helper scripts, it does more than prompt the model — it can run deterministic steps in a workflow without loading that code into context.
Skills live in a skills directory at the user level (under your Claude Code config directory) or the project level (checked into the repo so a whole team shares them), and they can also be distributed via plugins. Claude invokes a skill automatically when it's relevant, or you can trigger one directly like a slash command. They complement subagents for delegated work and MCP for external tools. For the exact directory layout and authoring format, see the official Claude Code docs.
Why it matters
Agent Skills turn one-off prompting into reusable, version-controlled expertise, so a developer or a whole team can encode how a task should be done once and have Claude Code apply it consistently across sessions — cutting repeated explanation and keeping the agent on-style without bloating context.
Deep dive: What Are Claude Code Skills? A Practical Guide
See Agent Skills in action across releases — browse the Claude Code changelog.
Get release alerts