Agent skills are reusable named procedures an agent loads on demand

A skill is a named procedure (run this eval, do a code review, ingest a source, refine this prompt) packaged once so an agent can pull it in only when the task at hand matches. It is the on-demand counterpart to always-on context such as an AGENTS.md file: both are soft controls inside the harness (see: an agent harness is everything that constrains and shapes an agent), but a skill applies when triggered, not all the time.

In practice, teams build dozens of them. The Show Us Your Agent Skills series collected 51 skills and 79 workflows from 20 builders, each one a hard-won method lifted out of a one-off run into something any future run can call. An “explain” skill makes a fleet of running agents narrate each change like a colleague handing off; an eval-driven skill turns every failed case into a library feature so the system cannot regress on it.

Like workflows, skills encode domain knowledge better than re-deriving the steps from a fresh prompt each time (see: workflows encode domain knowledge better than agents). They are a form of on-demand context (see: context management is the dominant variable in agent quality and cost), and a skill that captures corrections compounds across sessions (see: persistent memory lets corrections compound).