repo-standards-review

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Reviewing a repository against its standards

根据仓库标准审查代码仓库

This skill is guidance, not a complete checklist. Verify the PR's live base and head, then read the diff and enough surrounding context to understand the design before judging. A short review with one substantiated blocker beats a list of nits.

For code-facing reviews — lifecycle, concurrency, invariants, disposal, and other implementation semantics — also load the code review checklist.

本Skill为指导文档，并非完整检查清单。 请先验证PR的实时基准分支和目标分支，再阅读差异内容及足够的上下文以理解设计后再进行评判。一份包含确凿阻塞问题的简短审查，远胜于一堆琐碎的意见。

针对面向代码的审查——如生命周期、并发、不变量、资源回收及其他实现语义——还需参考代码审查清单。

Sources of truth

权威依据

The repository's root
```
AGENTS.md
```
: skill authoring, validation, and quality-gate rules.
The repository's documentation standard (usually
```
docs/AGENTS.md
```
): placement and prose discipline.
```
prose-standard
```
: required coverage and editorial judgment.
```
trim-cot-leakage
```
and
```
prune-prompt-pollution
```
: reasoning-transcript and prompt-pollution smells respectively.
The repository's Agent Note rules (usually
```
.agents/notes/README.md
```
): format and scope. Treat disagreement with a note as a design discussion, not an automatic veto.
The repository's bilingual pairing contract and terminology table (usually
```
docs/i18n/README.md
```
and
```
docs/i18n/terminology.md
```
).

仓库根目录下的
```
AGENTS.md
```
：Skill创作、验证和质量门禁规则。
仓库的文档标准（通常为
```
docs/AGENTS.md
```
）：内容放置规范与行文准则。
```
prose-standard
```
：要求的覆盖范围与编辑判断标准。
```
trim-cot-leakage
```
和
```
prune-prompt-pollution
```
：分别指代推理记录泄露和提示信息冗余的问题。
仓库的Agent Note规则（通常为
```
.agents/notes/README.md
```
）：格式与适用范围。将与Note的分歧视为设计讨论，而非直接否决。
仓库的双语配对约定与术语表（通常为
```
docs/i18n/README.md
```
和
```
docs/i18n/terminology.md
```
）。

Blocking requirements

阻塞性要求

New prose receives semantic review. Critically review every added or changed
```
SKILL.md
```
, description, doc, Agent Note, and comment with
```
prose-standard
```
. Automated checks do not establish coverage, accuracy, or placement.
Skills follow the layout and manifest. A new skill has a valid
```
name
```
(hyphen-case) and a trigger-focused
```
description
```
; it is listed in the skill manifest; and its install name is unique across categories. Flag platform-specific metadata, runtime branding, and hard local paths unless justified.
Bilingual pairs update together. A change to either side of a pair updates the counterpart and keeps both switcher lines; the documentation gates are green. A green pairing check does not prove translation quality.
Every non-trivial change carries an Agent Note in the same diff, updated to shipped present-tense state when a proposal is implemented.
Required evidence exists. The author ran the relevant checks for the diff (documentation gates, skill validator,
```
git diff --check
```
); review covers the semantic gaps those cannot detect.

新增文案需经过语义审查。 使用
```
prose-standard
```
严格审查所有新增或修改的
```
SKILL.md
```
、描述、文档、Agent Notes及注释。自动化检查无法确保内容的覆盖性、准确性和放置合理性。
Skill需遵循布局与清单规范。 新Skill需有有效的名称（连字符格式）和聚焦触发条件的描述；需在Skill清单中列出；且其安装名称在各分类中唯一。标记特定平台的元数据、运行时品牌信息和硬编码本地路径，除非有合理理由。
双语配对内容需同步更新。 修改配对内容的任意一方时，需同步更新另一方并保留双方的切换链接；文档门禁需显示为绿色。但门禁显示绿色并不代表翻译质量合格。
所有非微小变更需附带Agent Note，且当提案落地时，需将Note更新为已发布的现在时态表述。
需提供必要的验证证据。 作者需针对差异内容运行相关检查（文档门禁、Skill验证器、
```
git diff --check
```
）；审查需覆盖这些工具无法检测到的语义漏洞。

Manual checks

手动检查项

Intent and contract: trace both sides of every changed rule or workflow; confirm the change matches the PR description and any Agent Note.
Trigger quality: a description's triggers are behavior-based and one per distinct branch, not a keyword pile or repo identity the body already carries.
One home per fact: no rule restated in a second location without a link to its owner; no reasoning transcript, change narration, or duplicated rationale.
Scope and necessity: challenge speculative generality and unrelated edits; each change maps to a current consumer or a recorded decision.
Scripts: deterministic helpers handle failures loudly and state non-obvious contracts; generated files are not hand-edited.
Test/validation strength: assertions fail on the intended regression rather than restating the implementation.

意图与约定： 追踪每一项修改规则或工作流的前后变化；确认变更内容与PR描述及相关Agent Note一致。
触发条件质量： 描述中的触发条件需基于行为，且每个不同分支对应一个触发条件，而非堆砌关键词或重复正文已包含的仓库标识。
事实唯一来源： 同一规则不得在第二个位置重复表述，除非链接至其原始定义；不得包含推理记录、变更说明或重复的理由阐述。
范围与必要性： 质疑无依据的通用性设计和无关编辑内容；每一项变更需对应当前使用者或已记录的决策。
脚本： 确定性辅助工具需明确提示失败信息，并说明非显式的约定；生成的文件不得手动编辑。
测试/验证强度： 断言需在预期的回归场景下失败，而非仅重复实现逻辑。

Reporting findings

报告审查结果

State the defect, location, impact, and evidence. Separate blockers from suggestions and omit issues a green check already enforces. When receiving review, verify each claim and fix or rebut it on technical grounds without performative agreement. When the review surfaces dead, duplicated, speculative, or over-built surface, hand those candidates to

find-simplifications

to turn them into proposed Agent Notes or

TODO(tag)

markers.

明确说明缺陷、位置、影响及证据。将阻塞问题与建议分开，并忽略已通过绿色检查确认的问题。在接收审查反馈时，需逐一验证每一项主张，基于技术理由进行修复或反驳，而非敷衍认同。当审查发现冗余、重复、无依据或过度设计的内容时，可将这些候选内容提交给

find-simplifications

工具，转化为拟议的Agent Notes或

TODO(tag)

标记。