code-review-and-qualit…
Reviews a code change for correctness, security, maintainability, architecture, performance, and verification gaps, or performs a…
Audits documentation for internal contradictions and/or mismatches with code, config, schemas, and approved contracts. Use for 文档审计、前后冲突、术语漂移、README 是否过时、文档和代码一致吗, including check-and-fix requests. Select evidence sources from the request; return prioritized findings, source
$ npx -y skills add hylarucoder/hai-stack --skill hai-audit-docs --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/hai-audit-docsContext preview
The summary Claude sees to decide when to auto-load this skill.
Audits documentation for internal contradictions and/or mismatches with code, config, schemas, and approved contracts. Use for 文档审计、前后冲突、术语漂移、README 是否过时、文档和代码一致吗, including check-and-fix requests. Select evidence sources from the request; return prioritized findings, source
name: hai-audit-docs description: | Audits documentation for internal contradictions and/or mismatches with code, config, schemas, and approved contracts. Use for 文档审计、前后冲突、术语漂移、README 是否过时、文档和代码一致吗, including check-and-fix requests. Select evidence sources from the request; return prioritized findings, source precedence, and repair ownership. Apply clear local repairs only when requested. Use hai-rewrite-doc for wholesale reconstruction, hai-prd for product shaping, and readme-beautifier for formatting only.
For Chinese readers, see `SKILL.zh_CN.md`. The English file is the execution source of truth.
A coherent document can be factually wrong; implementation can violate an approved requirement. Declare which authority settles each claim before deciding which side needs repair.
Internal agreement proves consistency, not runtime truth.
behavior with approved specs/contracts. Inspect related internal contradictions as useful.
available. State the chosen scope; if code is unavailable, mark claims unverified instead of implying they match.
then recheck affected claims and links. Do not silently change implementation to make docs true. Escalate unresolved product/authority choices; continue independent clear repairs.
These are internal modes, not questions the user must answer before work starts.
1. Identify the named document surface, audience, scenarios, and requested action. Do not expand named targets into sibling documents without reason or authorization. With no named targets, inspect README, setup/API docs, and examples relevant to the question. 2. Build a claim map: behavior, commands, defaults, lifecycle, permissions, terminology, assumptions, acceptance, and cross-links. Record source precedence per class of claim. 3. For internal evidence, read `references/internal-audit.md`; distinguish explicit supersession from unsupported claims. Absence of support alone does not prove a claim false or obsolete. 4. For implementation evidence, read `references/checklist.md`. Trace documented scenarios to implementation and inspect public/config contracts back toward documentation for important omissions. Cite both locations and identify generated versus canonical sources. 5. Classify each finding as document defect, implementation gap, internal conflict, or needs evidence/decision. If a doc describes approved future behavior, mark the implementation gap; do not rewrite it to bless current behavior. Consult `references/worked-examples.md` when authority is disputed. 6. Recommend update, move, merge, remove, split, or resolve. Separate a suggested repair from an applied edit. Preserve real contracts and unresolved claims explicitly. 7. If repairs were requested, make the bounded authorized edits and verify them. Route a wholesale rewrite to `hai-rewrite-doc` with the established evidence; product decisions belong to `hai-prd`, plan changes to `hai-goal`. 8. Read `references/output-template.md` and report scope, findings, actual edits, verification, and unresolved decisions without repeating a full report for each evidence mode.
Rate actual impact, not the presence of a sensitive keyword. A security-related typo alone does not become a release blocker.
inspection alone does not prove runtime acceptance.
让 Coding Agent 把软件迭代做对:从真实问题出发,完成改动,用证据确认结果。 支持:大多数能读取 Skills(SKILL.md)的 Coding Agent,例如 Claude Code、Codex。 hai-stack 由 @海拉鲁编程客 创建。它把一次软件迭代拆成判断、设计、动手、确认、沉淀五步,每一步都有对应的技能。
Reviews a code change for correctness, security, maintainability, architecture, performance, and verification gaps, or performs a…
Audits an entity model field by field: what should exist, what exists now, where each value belongs (column, config, computed…
Challenges an overly conservative, incremental, or compatibility-led proposal with a bold target model, a concrete…
Pressure-tests an ambitious proposal and returns a landing judgment, one minimum viable proof, explicit cuts, success/failure…
Investigates architecture and change complexity from repository runtime paths to bounded modules, React components, and shared…
Produces a ready-to-run ast-grep command or reusable YAML lint/codemod rule, validated against positive and negative fixtures.…