skillifylisted

# Skillify — evidence-based skill graduation ## Step 1 — Gather evidence (disk, not memory) Read: PROGRESS.md (whole mission), the mission's review findings / flywheel notes, `git log --oneline` for the mission's commits, and the task files in docs/missions/<slug>/tasks/. Session memory is unreliable (compaction); the loop's artifacts are the record. ## Step 2 — Graduation criteria (ALL must hold, else stop) - **Recurrence:** the procedure or correction appeared ≥2 times across tasks or missions. Once = anecdote, not a skill. - **Determinism:** the steps were the same each time and can be written as an ordered sequence with no unstated judgment calls. - **Verifiability:** success is checkable (command output, file state, binary checklist) — "looks right" does not graduate. - **Not already covered:** grep existing .agents/skills/ descriptions; if an existing skill covers it, propose a PATCH to that skill instead (smaller is better — one new line beats one new file). ## Step 3 — Draft (agentskills.io format, house style) - Frontmatter: `name` (slug) + `description` that front-loads TRIGGERS ("MUST be used when...") — the description is the only thing loaded at startup, it does the routing. - Body: the failure this prevents (one paragraph) → numbered procedure with exact commands → binary self-verification checklist → "Never do these". - Keep it lean; details go in a references/ subfolder only if truly needed. - Agnostic core + "check AGENTS.md / docs/*.md