​
Definition

A Skill is a reusable operating procedure that an Agent can consult for a class of work. Skills can come from Rudder’s bundled set, an organization library, an Agent home, a user-level source, or an adapter-native directory. Four states matter: These states are not interchangeable. A loaded Skill is not proof of use.

​
One illustrative case

A release team repeats the same checks for notes, packages, and rollback details. After several reviewed releases, it turns the stable procedure into a release-check Skill. A release Agent can then use the same instructions on the next release, while each Run still records whether evidence shows that the Skill was actually used. Skills library

​
When it is useful

Create a Skill when a procedure has repeated, has a clear trigger, and should be reused across future work. Keep task-specific facts in the Chat, Issue, project resources, or Library. Keep durable role expectations in Agent instructions. Rudder includes a small always-available bundled baseline for product guidance, memory files, Skill creation, and visualization. Availability lets a runtime discover a Skill when relevant; it does not instruct the model to use every Skill on every turn.

​
Operating boundaries

Rudder resolves the Skill set for each invocation. Provider-native, stale, project, or user-home Skills do not become Rudder-loaded Agent Skills merely because a runtime can see their directories. Adapters must remove, isolate, or ignore entries outside the selected set. The highest-risk boundary is silent behavior change. Feedback does not rewrite a Skill automatically. Promote a lesson only after its evidence, scope, review, and rollback or evaluation path are clear. Read Reviews, feedback, and learning for that distinction.