With a skill in place, the agent follows a method you've already worked out instead of improvising a new one each time.

The closest everyday comparison is a recipe card pinned up in a kitchen: whoever picks it up follows the same steps, and the dish comes out the same.

How do agent skills work?

Each skill has a short name and description that the agent can always see, plus fuller instructions it reads only when a request matches. Anthropic's Agent Skills documentation describes this design: a skill is a folder containing a SKILL.md file of instructions, optionally with scripts and reference files, and the agent loads it automatically when it's relevant. Because only the descriptions sit in the context window until needed, an agent can have many skills available at little cost.

Some agents can also write skills themselves. When a method works, the agent saves it so it can repeat it later.

How is a skill different from a prompt?

A prompt belongs to one conversation, and you have to give it again next time. A skill is stored in the agent, which can pick it up without you remembering it exists. A saved template sits between the two: reusable, but only when you paste it in.

Why are skills useful?

  • Consistency. The tenth time a job is done follows the same method as the first, so quality doesn't drift.
  • Less repetition. You stop re-explaining the format, the rules and the exceptions every week.
  • Written-down know-how. A new colleague can read a skill and see how the job is done, which often makes it the procedure document your team never got round to writing.

An example

Every supplier quote that comes in should be checked the same way: compare the unit price with the last quote from that supplier, note the payment terms, flag anything over £5,000, and sum it up in five lines.

You set that out once as a skill, with one example of a good summary. From then on, whenever you paste a quote and say "check this", the agent runs the same checks in the same order and returns the five lines. You still read the result before acting on it, but you no longer have to explain the method.

What are the risks of agent skills?

A skill preserves whatever method is inside it, including a flawed one, and repeats it reliably. If the procedure was wrong when it was saved, every result that follows will be wrong in the same way.

Skills from unknown sources carry a different risk, because they can contain instructions or code you wouldn't want run. Anthropic's documentation advises using only skills you wrote yourself or got from a trusted source. Have a person review a new skill before the agent relies on it (see human in the loop), and revisit it whenever your process changes. Our guide to agent skills covers this in more depth.

Common questions

Can an agent create its own skills?

Some can. When a method works, the agent can save it for next time, which is useful but worth reviewing, because an agent can just as easily save a workaround that only made sense on one unusual day.

How many skills should an agent have?

Fewer than you might expect. A small set that someone keeps up to date works better than a long list nobody checks, since an outdated skill will keep producing outdated work with complete confidence.

Do skills need permissions?

Often, yes. A skill that reads files, queries another system or sends a message needs access to do so. Give it the narrowest access the task allows.