Agent Skills

Reusable packages of procedural memory that guide an agent through a class of tasks. A skill may combine instructions, decision criteria, scripts, references, and assets and is loaded when its scope matches the current request. It turns recurring practice into a versioned, testable artifact.

Formats vary across platforms and do not constitute one universal standard. A broad activation description injects unnecessary context; a narrow description prevents use when it would help. Skills also need least-privilege permissions, provenance, tests, and review because reusable instructions can repeat the same error at scale.