A skill is a folder containing a `SKILL.md` file (plus optional scripts and references) that an agent loads when the task matches. The model reads the instructions and follows them. There is no API, no runtime, and no install step beyond copying files into place.
What that buys you
Consistency. A skill is the difference between an agent that reviews code however it feels that day and one that applies the same checklist every time. It is the cheapest way to encode a team practice.
What it does not buy you
Capability. A skill cannot make the model able to run code, read a database, or see a browser — that is what MCP servers are for. Skills change how the model behaves, not what it can reach.
The common confusion: people write a skill hoping it will add a feature. It will not. If the agent cannot do the thing, the fix is a tool, not a longer instruction file.
What a good skill contains
Four things, in this order: **when it applies**, **what to do**, **what not to do**, and **how to verify the result**.
The fourth is the one people skip and the one that separates a skill from a wish. "Use this when reviewing a diff" plus "check that tests pass" is a skill. "Be a careful reviewer" is a mood.
Reading a skill before installing one
The ecosystem now has thousands of skills, many of them a paragraph of adjectives. A useful test: does the skill name a situation where it does *not* apply? Skills that never say no tend to consume context and change tone without improving output.
Browse the skill directory — every entry there carries a `bestFor` and a `notFor` for exactly this reason.