Propose safe improvements to a skill after a real task trace. Use when a skill caused confusion, missed a failure mode, repeated manual steps, used too much context, or succeeded in a reusable way. Produces human-reviewed proposals, not automatic live edits.
