What is a Claude skill?

A Claude skill is a small, shareable package that teaches an AI coding agent a role, a toolset, and a set of guardrails for a specific job — "Android engineer", "SEO auditor", "pitch deck designer". It's closer to a prompt template with muscle memory than a plugin: no new inference endpoints, just a crisp slice of context the agent pulls in when you ask for help.

The format is SKILL.md — an open plain-text standard: one Markdown file in a folder named after the skill. You can read the full spec at claudeskil.com/docs.

Why they matter

Skills make Claude's best behaviour repeatable. Instead of re-describing your style every conversation, you install a skill once and every thread it touches starts with the right tone, the right checklist, and the right tools.

They also make AI behaviour auditable. Because a skill is a plain file you can keep in a Git repo, you can see every change, every version, and every issue report. You know exactly what you're running.

How to pick a good one

Look for three signals: an active changelog (the skill is maintained), a readable README (the author understands their users), and evals or examples (the skill has been tested). Star counts are a lagging indicator — install counts and recent commits are the real signal.

How to try one

Copy the one-line install command from the skill's page and run it in your terminal. It saves the skill's SKILL.md to ~/.claude/skills/<name>/, where Claude Code finds it in every project (use .claude/skills/<name>/ to keep it to one project). Start a new session, try it on a representative task, compare the output to what you'd get without it, and keep it only if the delta is obvious — removing it is just deleting the folder.

Browse about 200 skills at claudeskil.com/explore.