The moment to reach for this is recognisable: you are pasting the same paragraph of instructions into Claude for the third time.
A is a folder with a SKILL.md in it. The official template is six lines long. skill-creator writes the first draft, then — the part people skip — helps you measure whether it works.
The part that decides whether your skill is any good
The description field. It is the only part Claude always has loaded, and it is what decides whether the skill fires at all. A vague description produces a skill that either never triggers or triggers constantly. Write what it does, when to use it, and — just as important — when not to.
The document skills on this page are a good : read the description on xlsx and notice how much of it is spent on when not to trigger.
Before you write one, check it shouldn’t be something simpler
- A standing rule for one → put it in
CLAUDE.md(see the worked example). - A you rerun by hand → a slash command.
- Something that must happen every time, without Claude choosing to → a hook.
A skill is right when Claude should decide for itself that now is the moment to follow your procedure.
Worth knowing
- Decide user-invoked vs model-invoked deliberately. A skill Claude may fire on its own needs a much more carefully written description than one you always type yourself.
- Test it.
claude plugin evalruns eval cases against a and scores the result. Unmeasured skills tend to be worse than their authors think. - Keep it short. The body loads only when the skill triggers, but it still spends context. Write the gotchas, not a manual.
Access
Public — no access request needed. It ships in ’s official marketplace, so there is no marketplace add step.