Back to Details

reviewing-skill

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Reviewing Skill

评审Skill

Review the target as a portable Agent Skills artifact when possible. If the target is only a skill-like instruction document, still review it and separately call out packaging gaps against the Agent Skills format.
尽可能将目标视为可移植的Agent Skills工件进行评审。如果目标只是类技能指导文档，仍需对其进行评审，并单独指出与Agent Skills格式相比存在的打包缺陷。

Review Goal

评审目标

Produce a concrete review that:
  • finds real issues, not generic advice
  • checks whether the skill uses the right design pattern or pattern mix
  • checks whether another AI could use the skill correctly and stably
  • recommends any missing design pattern from the five-pattern set when it would improve the skill
  • follows the required output template exactly
Load 
references/review-principles.md
before writing the review. Load 
assets/review-template.md
before writing the final report.
产出一份具体的评审报告，需满足：
  • 发现实际问题，而非泛泛建议
  • 检查该skill是否使用了正确的设计模式或模式组合
  • 检查其他AI能否正确且稳定地使用该skill
  • 当五种模式集中的某种缺失模式可提升skill质量时，建议添加该模式
  • 严格遵循要求的输出模板
撰写评审前，请先加载
references/review-principles.md
。 撰写最终报告前，请先加载
assets/review-template.md
。

Required Methodologies

必备评审方法

Use both review methodologies every time:
  1. Design patterns and principles Check the skill against the review principles and all five design patterns:
  • Tool Wrapper
  • Generator
  • Reviewer
  • Inversion
  • Pipeline
  1. Another AI usability and stability Imagine a competent AI activates the skill with no extra human coaching. Check whether that AI would:
  • know when to use the skill
  • know what file or resource to load next
  • know what order to follow
  • produce the required output shape
  • avoid ambiguous or unstable behavior
每次评审必须同时使用以下两种方法：
  1. 设计模式与原则 对照评审原则及全部五种设计模式检查skill：
  • Tool Wrapper（工具封装）
  • Generator（生成器）
  • Reviewer（评审器）
  • Inversion（反转模式）
  • Pipeline（流水线）
  1. 跨AI易用性与稳定性 假设一名合格的AI无需额外人工指导即可激活该skill，检查该AI是否能够：
  • 知晓何时使用该skill
  • 知晓下一步应加载哪个文件或资源
  • 知晓应遵循的操作顺序
  • 生成符合要求的输出格式
  • 避免模糊或不稳定的行为

Required Workflow

必备工作流程

  1. Identify the review target.
  2. Determine whether it is:
  • a packaged Agent Skill
  • a partial skill package
  • a skill-like instruction document
  1. If the target follows the official skill structure, read 
    SKILL.md
    first. If it does not, use the document the user identified as the review target.
  2. Read only the additional references, assets, or scripts needed to evaluate the skill.
  3. Review the target using both required methodologies.
  4. Evaluate all five design patterns, even if the target currently uses none of them.
  5. Recommend any pattern that should be added, removed, or made more explicit.
  6. Write findings with severity and file/line references.
  7. Render the final report using the template exactly.
  8. End by asking whether the user wants to implement the concrete changes you recommended.
  1. 确定评审目标。
  2. 判断目标类型：
  • 已打包的Agent Skill
  • 部分完成的skill包
  • 类技能指导文档
  1. 如果目标遵循官方skill结构，优先阅读
    SKILL.md
    。若不遵循，则使用用户指定的评审目标文档。
  2. 仅阅读评估skill所需的额外参考资料、资源或脚本。
  3. 使用上述两种必备方法评审目标。
  4. 评估全部五种设计模式，即使目标当前未使用任何模式。
  5. 建议应添加、移除或明确化的模式。
  6. 撰写包含严重程度及文件/行号引用的评审发现。
  7. 严格按照模板生成最终报告。
  8. 结尾询问用户是否希望落实你所建议的具体修改。

Severity Rules

严重程度规则

  • High: likely to cause the AI to misuse the skill, skip required behavior, produce wrong output, or fail unpredictably
  • Medium: meaningfully reduces review quality, portability, or reliability, but the skill may still work in some cases
  • Low: clarity, maintainability, or optimization issue with limited direct behavior risk
Do not inflate severity. Prefer fewer, sharper findings over long weak lists.
  • 高: 可能导致AI误用skill、跳过必要步骤、生成错误输出或出现不可预测的故障
  • 中: 显著降低评审质量、可移植性或可靠性，但skill在某些场景下仍可正常工作
  • 低: 清晰度、可维护性或优化问题，对行为的直接风险有限
请勿夸大严重程度。优先选择少量精准的发现，而非冗长的薄弱问题列表。

File And Line References

文件与行号引用

Every finding must include:
  • severity
  • explanation of the issue
  • file path
  • line number or the closest precise location available
Prefer embedding the location directly into the sentence, for example by naming the file, line number, and a short quoted phrase from the reviewed text. Do not hide all location detail at the end of the bullet. Never quote or reproduce secrets, credentials, tokens, API keys, passwords, private keys, or other sensitive values from the reviewed material. If a finding depends on sensitive content, describe it generically and redact the value, for example 
token [REDACTED]
or 
credential-like string
.
If the target lacks file structure, cite the best available location description and say that exact line references are unavailable.
每条发现必须包含：
  • 严重程度
  • 问题说明
  • 文件路径
  • 行号或最接近的精确位置
建议将位置直接嵌入句子中，例如提及文件名、行号及评审文本中的简短引语。请勿将所有位置细节隐藏在项目符号末尾。 切勿引用或复制评审材料中的机密信息、凭证、令牌、API密钥、密码、私钥或其他敏感内容。如果发现依赖敏感内容，请进行通用描述并编辑敏感值，例如
token [已编辑]
或
类凭证字符串
。
如果目标缺乏文件结构，请引用最佳可用的位置描述，并说明无法提供精确行号引用。

Pattern Review Rules

模式评审规则

Always evaluate each pattern explicitly:
  • Tool Wrapper: Would the skill be stronger if it taught a library, domain, policy, or internal system through references?
  • Generator: Would the skill benefit from templates, reusable output forms, or artifact generation?
  • Reviewer: Does the skill need a rubric, checklist, scoring rule, or findings format?
  • Inversion: Should the skill force clarifying questions or structured input gathering before acting?
  • Pipeline: Should the skill enforce an ordered workflow with gates or checkpoints?
The pattern assessment may appear inside 
Suggestions
, but it must show an explicit result for all five patterns, such as 
applies
, 
does not apply
, or 
already present
. If a pattern applies, recommend it directly in the review rather than leaving the implication implicit.
必须明确评估每种模式：
  • Tool Wrapper: 如果该skill通过引用传授库、领域知识、政策或内部系统，是否会更完善？
  • Generator: 该skill是否能从模板、可复用输出表单或工件生成中获益？
  • Reviewer: 该skill是否需要评分标准、检查清单、评分规则或发现格式？
  • Inversion: 该skill是否应在执行前强制要求澄清问题或收集结构化输入？
  • Pipeline: 该skill是否应强制执行带有关卡或检查点的有序工作流？
模式评估可置于
建议
部分，但必须明确显示所有五种模式的结果，例如
适用
、
不适用
或
已存在
。如果模式适用，请在评审中直接推荐，而非隐含其含义。

Another-AI Stability Checks

跨AI稳定性检查

Ask these questions during review:
  • Is the trigger description specific enough that an AI would load the skill at the right time?
  • Are the first actions explicit after activation?
  • Are references and assets named concretely?
  • Are mandatory steps clearly marked?
  • Are outputs constrained enough to be repeatable?
  • Are there loopholes that let an AI skip key steps?
  • Does the skill separate "when to use" from "how to do it" clearly enough for discovery and execution?
评审期间需询问以下问题：
  • 触发条件描述是否足够具体，使AI能在正确时机加载该skill？
  • 激活后的首个操作是否明确？
  • 参考资料和资源是否有具体名称？
  • 强制步骤是否有清晰标记？
  • 输出是否受到足够约束以确保可重复性？
  • 是否存在可让AI跳过关键步骤的漏洞？
  • 该skill是否将“何时使用”与“如何执行”清晰分离，以便于发现和执行？

Output Contract

输出约定

Your final output must match 
assets/review-template.md
. Do not replace the requested sections with a different structure. Write findings in natural prose. Explain impact directly instead of using the literal label 
Why it matters:
. When citing reviewed content, prefer short non-sensitive phrases. Redact any sensitive material instead of quoting it.
最终输出必须与
assets/review-template.md
匹配。 请勿用不同结构替换要求的章节。 用自然散文撰写发现。直接说明影响，而非使用字面标签
重要性：
。 引用评审内容时，优先选择简短的非敏感短语。编辑任何敏感材料，而非直接引用。