AGENT SKILL

outcome-quality-enforcement

Enforce the highest credible quality on a consequential team outcome through reference calibration, observable success contracts, native operation, independent criticism, root-cause revision, and quality-convergence stopping. Use when an accountable lead commissions creative or ambiguous work, judges a major deliverable, sees repeated near-misses, or must decide whether more iteration is warranted. Do not use to turn routine low-risk work into ceremony.

License 许可
Apache-2.0
最近更新
2026-11-14
Stars
★ 12
Forks
1

01 前言 Front Matter

name
outcome-quality-enforcement
description
Enforce the highest credible quality on a consequential team outcome through reference calibration, observable success contracts, native operation, independent criticism, root-cause revision, and quality-convergence stopping. Use when an accountable lead commissions creative or ambiguous work, judges a major deliverable, sees repeated near-misses, or must decide whether more iteration is warranted. Do not use to turn routine low-risk work into ceremony.

03 SKILL.md 正文

Outcome quality enforcement

The accountable lead owns the accepted outcome. Staff own production. A completed Attempt is evidence to judge, not proof that the outcome is good.

Form the quality contract

Before expensive production, retrieve rather than guess:

  • the real product, operating environment, repository, data, and constraints;
  • authoritative product or brand truth;
  • the strongest accepted incumbent or gold-standard reference;
  • prior failures as compact observations, without inheriting rejected solutions; and
  • the specialist skill or evaluator that matches the outcome.

Resolve any ambiguity that changes the solution class, cost, authority, or acceptance bar. Otherwise make a reversible assumption and record it. Keep one concise durable contract containing:

  1. the observable outcome and whose reality must change;
  2. priorities and explicit trade-offs;
  3. references ordered by authority;
  4. the incumbent or minimum quality floor;
  5. exclusions and authority boundaries;
  6. native evidence required for acceptance;
  7. the independent evaluator and verdict language;
  8. the owner-attention envelope; and
  9. the convergence and blocker conditions.

Give producers the minimum sufficient context: the contract, exact inputs, environment, and relevant examples. Do not replay the full lead conversation or make them reverse-engineer the outcome from a checklist.

Run the evidence loop

Use the smallest Work graph that preserves real accountability:

  1. Commission Staff to produce a coherent end-to-end candidate or one high-information slice.
  2. Operate the exact candidate in its real medium. Capture native evidence before interpreting it.
  3. Commission a fresh-context independent evaluator when judgement matters. Give it the contract, ordered references, exact candidate, and factual evidence—but not producer rationale.
  4. Require a decisive verdict and ranked consequential gaps. Check that the evaluator actually tested the promised outcome; a weak verifier creates false confidence.
  5. Commission the smallest root-cause repair with a credible improvement hypothesis, then evaluate the new exact candidate again.

Budget evidence context deliberately. Keep large screenshot, trace, log, and capture corpora as durable linked outputs; give the evaluator compact manifests, representative native views, and commands or locators for retrieving the rest. When capture and judgement together would saturate one session, separate evidence production from a fresh synthesis pass rather than carrying an ever-growing transcript. Verification commands must be non-interactive and must not install packages, rewrite lock files, or otherwise change the candidate they purport to judge.

Continue while a consequential gap and credible improvement path remain. Attempt limits bound local execution failure; they do not lower the team outcome. Create sparse revision or replacement Work as needed. After the same important failure twice, change the approach, context, skill, model, evaluator, or team seam. Use a clean reset when the trajectory is anchored to rejected assumptions or form.

Prefer one coherent root fix over accumulated patches. Complexity must buy observable quality, reliability, or future reasoning leverage. Convert repeated failures into a reusable skill, test, evaluation, tool, or compact durable observation.

Converge honestly

Accept only when:

  • the contract is proven against the exact native artifact;
  • the result clears the stated floor and ordered references;
  • independent review accepts it when required;
  • the verifier and evidence are themselves credible;
  • no consequential unresolved gap remains; and
  • the owner is asked only for authority or irreducible judgement, with a ready-to-use review target.

Do not stop because a build is green, a producer is confident, an arbitrary loop count elapsed, or the lead is tired of iterating. Stop when quality converges or when a concrete external constraint makes further machine progress impossible. State residual uncertainty and disagreement plainly.

04 同仓库更多 Skills

BlueprintLabIO
BlueprintLabIO/restless-core

fixture-note-etiquette

How the company writes a Fixture Notes message. Used only by the connections smoke.

★ 12
BlueprintLabIO
BlueprintLabIO/restless-core

web-native-review

Prepare or independently review a rendered website against its outcome contract and a mature ground-truth reference. Use for landing pages, product sites, blogs, documentation sites, or other browser-native outcomes; do not use to judge source code alone or turn mechanical layout facts into a taste score.

★ 12
BlueprintLabIO
BlueprintLabIO/restless-core

scenario-evidence

Run a versioned Runtime scenario, probe its real capabilities, preserve compact evidence, and prepare its native review target. Use for a project outcome that benefits from repeatable setup/run/assertion/evidence—not to choose a plan, team, or business verdict.

★ 12
BlueprintLabIO
BlueprintLabIO/restless-core

research

Answer a question the company needs to decide on with a sourced brief - market facts, a technology, a regulation, an audience, a price, a supplier. Use when an answer must be found and checked on the web or in documents before acting; do not use for a quick lookup a single source settles, or for competitor landscapes (use competitor-research).

★ 12
BlueprintLabIO
BlueprintLabIO/restless-core

reference-led-reset

Reset a rejected or repeatedly drifting outcome into a genuinely fresh Work Attempt grounded in ordered authoritative references. Use when the owner asks to start over, when qualitative defects survive revision, or when a mechanically valid result conflicts with the product or brand source of truth. Do not use for an ordinary bounded repair that should preserve the current candidate.

★ 12
BlueprintLabIO
BlueprintLabIO/restless-core

publish-git

Safely reconcile, verify, and publish Git refs through the governed Restless effect path. Use for pushes or deployments driven by a Git push; do not use for read-only fetches, local Git work, or provider API administration.

★ 12