Skip to content

Self-curating skills

v0.5.0 adds a lightweight skill-draft lifecycle. Completed turns that use a reusable sequence of tools can produce a draft SKILL.md under the draft skill root. Drafts do not become active automatically. They appear in the catalog as pending material with source-session metadata and a detection rationale.

  • Detect: after a successful non-command turn, Kanade reviews the tool trace.
  • Draft: the daemon writes a candidate skill with frontmatter, source session id, and rationale.
  • Review: the catalog shows draft cards separately from active skills.
  • Promote: the CLI or catalog moves a draft into the active skills root.
  • Prune: stale active skills can be reported as candidates for removal.

The CLI mirrors the UI:

Terminal window
kanade skill draft list
kanade skill draft promote <id>
kanade skill prune --since 30d

A skill should capture a repeatable procedure: release verification, a CT deploy pattern, a browser QA ritual, or a project-specific setup checklist. It should not be a changelog, a transcript summary, or a secret store. Promotions remain operator-controlled so a single lucky turn cannot rewrite the working method.