Skillify
Portable orchestrator/worker harness kit for AI-assisted development
npx -y skills add hashorva/purrfect-harness --skill skillifyAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
3 things to look at
- 17 days oldThe repository was created 17 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Graduate a battle-tested procedure into a new SKILL.md in .agents/skills/. MUST be used at mission end (after the last feature passes), whenever the code-reviewer's flywheel notes flag the same failure class or procedure twice, or when the user says "skillify this", "make this a skill", or "we keep doing this". Orchestrator-only. Produces skills from EVIDENCE (PROGRESS.md, review findings, git log) — never from speculation.
SKILL.md
2.7 KB, as published. Nobody here has run it
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/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) +descriptionthat 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 for project specifics" pointers — never hardcode one repo's paths into a studio-wide skill.
Step 4 — Human gate (mandatory)
Present: the evidence (which PROGRESS entries / findings triggered this), the
full draft SKILL.md, and whether it's a NEW skill or a PATCH. Only after
approval: write to .agents/skills/<name>/SKILL.md (propagates via symlinks),
commit as docs(skills): graduate <name> from mission <goal>. If the skill is
studio-wide, remind the human to copy it to the agent-harness kit and other
repos.
Never
- Never create a skill from a single occurrence or a hypothetical.
- Never duplicate an existing skill's territory — patch it.
- Never skip the human gate; skills are law for every future worker.