界面已本地化;技能介绍与说明文档按所选语言由 AI 自动翻译,代码、命令、路径和专业术语保留原文,以原作者文档为准。
探索技能
开发与工程 · SKILL.md

spec-readiness-review

spec-readiness-review

【开发与工程】Reviews, repairs, and approves implementation spec readiness. Use when specs need semantic review, deterministic checks, readiness reports, approval gates, or gap repair before planning.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准

Ssebastianwessel / skillsApache-2.0

技能说明

GitHub

name: spec-readiness-review description: Reviews, repairs, and approves implementation spec readiness. Use when specs need semantic review, deterministic checks, readiness reports, approval gates, or gap repair before planning.

Spec Readiness Review

Review canonical specs after authoring and before planning. Own the readiness decision, deterministic checks, semantic judge evidence, and spec gap repair. Do not create implementation plans or implement code.

Standard

Approve only when planning agents can create tickets without deciding behavior, interfaces, failures, data, runtime, security, recovery, release, ops, tests, clean-rebuild strategy, generation mapping, general file/folder structure, or closed-boundary type semantics. Specs must define every capability end to end: actor/consumer, entrypoint/reachability, contracts, data, states, side effects, permissions, errors, recovery, observability, acceptance, verification, and final state or explicit N/A. Use the checklist walk to find omitted topics before planning. Third-party and dependency choices require current primary documentation, package/release metadata, or dated research evidence. For in-scope data shapes, approve only a catalog that distinguishes canonical semantic shapes from valid boundary/persistence/projection representations and records every allowed mapping.

Modes

  • Review/Approve: judge existing specs, run checks, write or update specs/.readiness-report.yaml, and route blocking findings.
  • Fix Gap: update canonical specs from explicit findings, relink dependents, prune stale duplicates, record impact, then rerun readiness review. Do not invent broad requirements; route broad authoring to spec-architect.

Workflow

  1. Confirm source specs exist and the requested scope is clear.
  2. Load references/readiness-gates.md, references/artifact-shapes.md, and references/spec-checklists.md, and references/representation-catalog.md; then load relevant high-level checklist indexes and only the detailed topic checklist files those indexes route to.
  3. Run the approval-time spec judge loop from references/readiness-gates.md.
  4. Verify the capability inventory and end-to-end definition chain across feature, UX/client, API, data lifecycle, state transition, async, and ops paths. Validate the reciprocal 00-traceability.yaml graph from requirement through capability and path to testable acceptance.
  5. Walk relevant checklist indexes and topic files; record covered, not_applicable, or gap evidence and block on any relevant gap.
  6. Verify dependency and third-party solution specs use current stable versions by default, official docs for schema/query/security/config/performance guidance, and explicit rationale for older pins or nonstandard choices.
  7. Run semantic judge evidence when available.
  8. Run node <spec-readiness-review-skill-root>/scripts/check_specs.mjs <spec-root>. Before approval, run node <spec-readiness-review-skill-root>/scripts/generate_spec_manifest.mjs <spec-root> <source-revision> after every canonical-source edit, then obtain fresh, content-bound approval evidence; never carry approval forward across a digest change.
  9. For contract-heavy boundaries, verify the clean-rebuild decision, generation map, generated outputs/checks, strong boundary type policy, handwritten remainder, and compatibility/migration stance.
  10. Verify every in-scope representation has a catalog entry, canonical owner, closed/open policy, consumer list, and approved mapping or N/A evidence.
  11. If gaps exist, update canonical specs only in Fix Gap mode or persist findings for spec-architect; do not approve.
  12. If all gates pass and human approval is recorded, write approved readiness with the matching manifest digest, an accountable approval reference, and semantic-judge evidence. Verify decision authority, scoped applicability, reuse inventory, and module-boundary artifacts before handoff.
  13. Self-audit assumptions, skipped checks, current-doc uncertainty, checklist uncertainty, and residual risk.

Stop Conditions

Stop without approval when specs are missing, scope is unclear, human approval is missing, approval evidence is not bound to the current manifest digest, semantic judge evidence is unavailable when required, deterministic checks fail, source contracts are ambiguous, generation mapping is incomplete, weak closed-boundary types are accepted without an open JSON leaf, or planning would need to invent behavior, file/folder placement, dependency versions, or third-party implementation guidance, a data shape, or a representation mapping.

Reference Map

  • references/readiness-gates.md: approval gates and judge loop.
  • references/artifact-shapes.md: expected spec tree and readiness report.
  • references/spec-checklists.md: root checklist index that routes to high-level checklist indexes and detailed topic files.
  • references/representation-catalog.md: required catalog shape and reuse decision rules for in-scope data representations.
  • scripts/check_specs.mjs: deterministic spec-shape checker.

Lifecycle Handoff

Authoring and broad spec updates start in spec-architect. Approved readiness hands off to spec-implementation-planner. Implementation review spec gaps return here for persisted findings and repair routing. If stage is unclear, use spec-driven-workflow.

我的收藏