agentleFS
EastkapEastkap/autocomp/.claude/skills/initiate-qa/SKILL.md

initiate-qa

Skill4 starsChanged 5 months agoMITView on GitHub

Run exactly ONE bounded cycle of the persistent QA verification lane — sweep the qa queue + context stream, atomically claim one review-status card, verify it from INDEPENDENT evidence via a matched verifier persona, verdict PASS→done / FAIL→back to the acting lane, then record and heartbeat. QA is the only lane that moves review→done. Use when the user says "initiate QA", "run a qa cycle", "/initiate-qa", or the qa lane runner fires this as its headless cycle prompt.

The file

We split this skill into the rules it gives an agent

4 rules. About 0 are common, 2 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Read roles/qa.md in full and adopt it.

    line 14
  2. Very specific:

    If private/state/.lane-qa-pause exists: run tools/context.sh heartbeat qa paused, report "qa lane paused (sentinel present)" and STOP — a …

    line 18
  3. Distinctive:

    Never self-schedule — the runner owns cadence.

    line 77
  4. Distinctive:

    Bouncing a card a third time: instead of escalating to CEO.

    line 87

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

initiate-qa — one QA lane cycle

One bounded cycle; the lane runner loops it, you don't. roles/qa.md IS the contract — this skill adds only the mechanics of a cycle. CLAUDE.md hard rules bind throughout. Core law: an acting lane's status write is a CLAIM; only independently observed evidence (live URL, screenshot, server state, file content) is evidence.

1. Load

Read roles/qa.md in full and adopt it. Read CLAUDE.md hard rules; have tools/verifiers.md (persona registry) and tools/verify-goal.md (generic check) at hand.

2. Gate

If private/state/.lane-qa-pause exists: run tools/context.sh heartbeat qa paused, report "qa lane paused (sentinel present)" and STOP — a cheap no-op cycle, not an error.

3. Sweep (whole queue, never from memory)

  • tools/tasks.sh list-all --tag qa — ALL statuses.
  • tools/context.sh sweep qa — unread rows; the claimed card's linked rows give you the goal + success criterion — never the verdict.

4. Claim & verify (one card per cycle)

  • tools/tasks.sh claim qa — claims a review-status qa-tagged card (it STAYS in review, claimed_by=qa); empty result = empty queue, not an error.
  • Dispatch a matched verifier persona: match the card's tags via tools/verifiers.md and spawn it fresh-context (Agent tool) at the persona's declared model tier. Graceful fallback: while tools/verifiers.md does not exist yet (lands in U11), EVERY card gets the generic tools/verify-goal.md check — fresh subagent, handed ONLY the goal + the verifiable success criterion + how to check (URL/file/SQL), never the actor's narrative. Nothing skips review either way.
  • Verdict:
    • PASS → tools/tasks.sh update <id> done "<evidence one-liner>" + verdict row: tools/context.sh post qa result "PASS: <evidence observed>" --tags qa,verdict --refs '{"card":"<uuid>","evidence":"<url/path>"}'.
    • FAIL → verdict row first (same shape, "FAIL: <concrete gap>"), then restore the acting lane's tags MINUS qa (tools/tasks.sh tag <id> <acting-lane> — the verb replaces, which is what you want here) and tools/tasks.sh update <id> todo "QA FAIL: <gap> — see verdict context row <id>".
    • 2 consecutive FAILs on one card → create a ceo-tagged escalation card (tools/tasks.sh add "ESCALATION: <card> failed QA twice" 3 "<both gaps>" agent ceo); stop bouncing it.
    • Verifier crashed/timed out → card STAYS in review; post a blocker row. A dead verifier never silently passes work.
  • Queue empty → tools/registry.sh next, then ONE standing-audit step for that venture (live surfaces return 200 + headline, KPIs match raw tables, recent done cards carry verdict rows). Genuinely nothing viable → tools/context.sh post qa handoff "<escalation>" --tags ceo — never falsely idle, and never a spurious row when there is simply nothing to verify.
  • Measurement discipline: any number you repeat comes from its raw table this cycle; per-day columns are deltas; unwired source = n/a, never a guess.

5. Record

  • Card statuses + verdict/blocker rows as above (rows carry --slug + refs).
  • Venture slug worked → private/state/.current-cycle-slug-qa (one line, overwrite; company for whole-board audits).
  • ONE ledger digest line to private/state/ledger.md via tools/append.sh (pending U5 — until it exists, append with >> only, never editor-style read-modify-write).
  • Disposition invariant: never exit half-verified — unfinished verification → the card stays review with progress notes; never left in a claim you can't stand behind.
  • tools/context.sh heartbeat qa <ok|idle|blocked> "<one-liner>" — last, always.

6. Report

Final message = the cycle report: per card PASS|FAIL — evidence actually observed (quoted text, URL + status, screenshot path) and where it moved; escalations/blockers raised; metric corrections; what was NOT verified and why (Principle 11).

Headless deltas (lane runner / claude -p)

  • No AskUserQuestion.
  • Gated classes (money / owner's personal identity / mass-cold outbound / destructive — rare for QA, but e.g. a paid solve to reach evidence): PENDING appended to private/state/approvals.md + tools/context.sh post qa handoff "<gate>" --tags ceo (CEO owns the human card) + tools/notify.sh push. Never executed in-cycle.
  • Never self-schedule — the runner owns cadence.

Anti-patterns

  • Passing on the actor's say-so — reading the doer's notes/logs as evidence.
  • Handing the verifier the narrative — it gets goal + criterion + how-to-check only.
  • FAIL without the gap named — "failed" alone bounces work uselessly.
  • Fixing what you found — you never write the repo; the fix is a FAIL verdict or a ceo-tagged proposal.
  • Absence from one failed lookup — full UUIDs, correct filters, reproduce a second way; an HTTP error is "my query failed", not "it doesn't exist".
  • Bouncing a card a third time instead of escalating to CEO.

Cost note

Cheap tiers do cheap checks: the persona registry declares the model tier (Haiku for copy/link checks, Sonnet for browser QA) — dispatch at that tier, not the session default. One card per cycle; the verifier subagent is the spend, so never dispatch two for one card.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.