agentleFS
qengine13qengine13/forge-skills/skills/completion-hygiene/SKILL.md

completion-hygiene

Skill1 starsChanged 8 days agoApache-2.0View on GitHub

Run before saying done, complete, ready, merge-ready, shipped, or milestone reached for a feature, phase, sprint, release, or project. Audit the plan against real artifacts, production invocation paths, and verification evidence; emit a literal completion verdict and supported wording, then narrow or block unsupported claims. Do not use for routine status reports, exploratory work, or small fixes with no completion claim.

The file

Show its rules and text

We split this skill into the rules it gives an agent

14 rules. About 0 are common, 12 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈12

    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. Distinctive:

    Read model and tool selection before selecting models or delegating.

    line 16
  2. Distinctive:

    Do not collapse the audit into safe-sounding prose.

    line 38
  3. Distinctive:

    If the work had no written plan, recover the promises from the user's request and authoritative project files.

    line 56
  4. Distinctive:

    Run the gates in order. Keep the evidence concise, but do not compress missing proof into a conversational assurance.

    line 60
  5. Distinctive:

    Read the governing plan or recover a bounded promise list from the current request and canonical files.

    line 64
  6. Distinctive:

    Verify each promised artifact in the real files.

    line 76
  7. Distinctive:

    If a required artifact is missing, either implement it within the authorized scope or narrow the claim.

    line 78
  8. Very specific:

    Write the repository's required completion artifact.

    line 99
  9. Distinctive:

    When test or release evidence is unavailable, say "not verified" or "no result supplied".

    line 108
  10. Distinctive:

    Verify the live tree and current external state.

    line 115
  11. Distinctive:

    Do not bypass hooks or rename a claim to evade project doctrine.

    line 117
  12. Distinctive:

    Do not write persistent memory unless the user explicitly authorizes it.

    line 119
  13. Distinctive:

    If the task only asks for a status report, audit the evidence needed for that report without mutating files or external state.

    line 121
  14. Very specific:

    Use references/origin-paper-tigers.md when training or explaining why existence is not invocation.

    line 133

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 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
<!-- Public edition: adapted for portable installation and subscription-aware model selection. See collection NOTICE. -->

Completion Hygiene

Prevent paper-tiger completion claims: code or documentation that exists but is not reachable, exercised, or proven through the promised real path.

Repository doctrine overrides this generic workflow. Treat passing tests, local commits, remote publication, independent review, hosted checks, deployment, and user-visible verification as separate claims unless the project explicitly combines them.

Runtime compatibility

Read model and tool selection before selecting models or delegating. The host and user control tools, permissions, subscriptions, and costs.

Mandatory result on every firing

Even when the user asks only for announcement, commit, or handoff wording, emit these fields before any explanatory prose:

VERDICT: CLAIM_SUPPORTED / CLAIM_NARROWED / CLAIM_BLOCKED / PLAN_AMENDMENT_REQUIRED
SUPPORTED WORDING: <exact wording the current evidence permits, or NONE>

EVIDENCE BOUNDARIES:
- Local tree: <checked and result / not checked>
- Remote state: <checked and result / not checked>
- Hosted checks: <checked and result / not checked>
- Runtime/browser: <checked and result / not checked>

PROMISE MATRIX:
| Promise | Source | Artifact evidence | Invocation evidence | Status |

Do not collapse the audit into safe-sounding prose. Repository reporting doctrine may require an additional status prefix, but it does not remove this literal verdict contract.

Matrix status must be exactly PROVEN, EXISTS_UNPROVEN, MISSING, or NOT_APPLICABLE. User-supplied artifact existence without proven production reachability is EXISTS_UNPROVEN, never PROVEN. Use MISSING only after inspection affirmatively establishes that the promised artifact or path does not exist. Use EXISTS_UNPROVEN when evidence is absent, unavailable, or not checked.

Inputs

Collect before auditing:

  • the exact completion claim;
  • the originating plan, pre-flight, specification, or recovered promise list;
  • current repository doctrine and source-of-truth files;
  • the live diff and relevant source files;
  • available test, review, continuous-integration, deployment, and runtime evidence.

If the work had no written plan, recover the promises from the user's request and authoritative project files. Do not invent a broader contract or retroactively label unplanned work as a completed phase.

Four-gate protocol

Run the gates in order. Keep the evidence concise, but do not compress missing proof into a conversational assurance.

Gate A — Recover the promises

Read the governing plan or recover a bounded promise list from the current request and canonical files. Extract every promised:

  • artifact;
  • behaviour and trigger;
  • integration or invocation path;
  • acceptance or completion gate;
  • approval, review, publication, or deployment boundary.

Cite each promise to its source. If no defensible promise list can be recovered, stop the broad completion claim and report what narrower statement is supported.

Gate B — Inventory the artifacts

Verify each promised artifact in the real files. Record the path and relevant line or exact external evidence. An enum, schema, module, test double, comment, or document is existence evidence only.

If a required artifact is missing, either implement it within the authorized scope or narrow the claim. Do not continue toward a green verdict on the original claim.

Gate C — Prove invocation and behaviour

For each behavioural promise, trace a real path from the user, command, request, scheduled job, hook, or other production entry point to the implementation.

Classify each promise:

  • PROVEN — artifact exists, a real path reaches it, and appropriate behaviour evidence passes;
  • EXISTS_UNPROVEN — artifact exists but reachability or behaviour is not proven;
  • MISSING — promised artifact or path does not exist;
  • NOT_APPLICABLE — the plan explicitly makes invocation unnecessary, with the reason recorded.

Tests support the path but do not replace it. A unit test that calls a function directly does not prove production wiring. For user-interface claims, use visible browser evidence when available. For concurrency, tenancy, money, security, migration, or external-system claims, require evidence that exercises that exact boundary.

Use a second reviewer or subagent when the project requires independent review or the change is large enough to benefit from it. Route by capability and pass the raw promise list and files, not a suggested diagnosis. A self-check is acceptable only when project doctrine permits it; label it.

Any core EXISTS_UNPROVEN or MISSING promise blocks the original completion claim.

Gate D — Write the verdict

Write the repository's required completion artifact. If no local convention exists, use docs/completion-audit/YYYY-MM-DD-<scope>.md and the template in references/audit-template.md.

Choose exactly one verdict:

  • CLAIM_SUPPORTED — every promise required by the claim is proven;
  • CLAIM_NARROWED — evidence supports any meaningful smaller claim, which must be rewritten exactly;
  • CLAIM_BLOCKED — no meaningful claim in the requested family is currently supported;
  • PLAN_AMENDMENT_REQUIRED — the plan must change before the claim can be made.

When test or release evidence is unavailable, say "not verified" or "no result supplied"; do not assert the tests never ran or the release never happened.

Record absent evidence as absent. Never convert “not checked,” “cannot access,” or “would probably pass” into a pass.

Hard rules

  • Verify the live tree and current external state; memory and prior summaries are retrieval aids only.
  • Distinguish built, locally tested, independently reviewed, committed, pushed, continuous-integration green, merged, deployed, and visibly verified.
  • Do not bypass hooks or rename a claim to evade project doctrine.
  • Re-run the affected gates after remediation or scope changes.
  • Do not write persistent memory unless the user explicitly authorizes it. If authorized, capture only the verified verdict and durable lesson after the artifact exists.
  • Founder or owner approval requires separately attributable evidence. Agent prose is not ratification.
  • If the task only asks for a status report, audit the evidence needed for that report without mutating files or external state.

Output

Return:

  1. the literal VERDICT and SUPPORTED WORDING fields;
  2. the promise matrix with file, line, command, and external-state evidence as applicable;
  3. the verification commands and results;
  4. remaining risks or blockers;
  5. the completion-audit path when an artifact was required.

Use references/origin-paper-tigers.md when training or explaining why existence is not invocation. It is background evidence, not a substitute for the current audit.

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.