Back to Details

review-skills

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Review Skills

审核Skills

Review and analyze a skill against best practices for length, intent scope, and trigger patterns.
根据长度、意图范围和触发模式的最佳实践，对Skill进行审核和分析。

Prerequisites

前提条件

Before analyzing, read these resources to understand skill writing principles:
  1. ./references/skill-creator/SKILL.md
    - Core principles, anatomy, and progressive disclosure
  2. references/spec.md
    - Complete Agent Skills specification (required for compliance checks)
  3. references/validate.md
    - Validation checklist (used in Step 2)
  4. ./references/skill-creator/references/workflows.md
    - Workflow patterns (if relevant)
  5. ./references/skill-creator/references/output-patterns.md
    - Output patterns (if relevant)
在分析之前，请阅读以下资源以了解Skill编写原则：
  1. ./references/skill-creator/SKILL.md
    - 核心原则、结构和渐进式披露
  2. references/spec.md
    - 完整的Agent Skills规范（合规性检查必需）
  3. references/validate.md
    - 验证清单（步骤2中使用）
  4. ./references/skill-creator/references/workflows.md
    - 工作流模式（如适用）
  5. ./references/skill-creator/references/output-patterns.md
    - 输出模式（如适用）

Reference Examples from Anthropic (REQUIRED)

来自Anthropic的参考示例（必需）

You MUST read reference skills from Anthropic's repository before analyzing. This is essential for calibrating your review.
  1. Ensure cache is available: Check if 
    ./.cache/anthropics-skills/
    exists. If not (or if stale), run:
    bash
    python scripts/download_anthropics_skills.py
  2. Read at least 3 reference skills: Before analyzing, read these SKILL.md files from 
    ./.cache/anthropics-skills/skills/
    :
    Always read these high-quality examples:
    • pdf/SKILL.md
      - Well-structured workflow skill with clear triggers
    • docx/SKILL.md
      - Good example of document processing patterns
    • skill-creator/SKILL.md
      - Meta-skill showing best practices
    Then read 1-2 skills similar to the one being reviewed:
    • For workflow-based skills: 
      xlsx/SKILL.md
      , 
      pptx/SKILL.md
    • For tool/API skills: 
      mcp-builder/SKILL.md
    • For creative/design skills: 
      brand-guidelines/SKILL.md
      , 
      frontend-design/SKILL.md
    • For testing skills: 
      webapp-testing/SKILL.md
  3. Note patterns to compare: As you read, note:
    • How descriptions are structured (trigger patterns)
    • Length and depth of SKILL.md body
    • How references are organized and used
    • Balance between brevity and completeness
在分析之前，您必须阅读Anthropic仓库中的参考Skill。这对于校准您的审核至关重要。
  1. 确保缓存可用：检查
    ./.cache/anthropics-skills/
    是否存在。如果不存在（或已过期），请运行：
    bash
    python scripts/download_anthropics_skills.py
  2. 至少阅读3个参考Skill：在分析之前，请阅读
    ./.cache/anthropics-skills/skills/
    中的以下SKILL.md文件：
    务必阅读这些高质量示例：
    • pdf/SKILL.md
      - 结构清晰的工作流类Skill，具有明确的触发条件
    • docx/SKILL.md
      - 文档处理模式的优秀示例
    • skill-creator/SKILL.md
      - 展示最佳实践的元Skill
    然后阅读1-2个与待审核Skill类似的Skill：
    • 工作流类Skill：
      xlsx/SKILL.md
      、
      pptx/SKILL.md
    • 工具/API类Skill：
      mcp-builder/SKILL.md
    • 创意/设计类Skill：
      brand-guidelines/SKILL.md
      、
      frontend-design/SKILL.md
    • 测试类Skill：
      webapp-testing/SKILL.md
  3. 记录可对比的模式：阅读时，请记录：
    • 描述的结构方式（触发模式）
    • SKILL.md主体的长度和深度
    • 参考资料的组织和使用方式
    • 简洁性与完整性之间的平衡

Steps

步骤

Step 1: Receive the Skill to Review

步骤1：接收待审核的Skill

The user must provide a skill folder/path to review. If not provided, prompt:
"Please provide the path to the skill folder you want to review (e.g., 
.claude/skills/my-skill/
)"
用户必须提供待审核的Skill文件夹/路径。如果未提供，请提示：
"请提供您想要审核的Skill文件夹路径（例如：
.claude/skills/my-skill/
）"

Step 2: Validate Skill Structure

步骤2：验证Skill结构

Using the validation checklist (
references/validate.md
), verify the skill passes all basic checks:
  1. File Structure: SKILL.md exists
  2. Frontmatter Format: Valid YAML between 
    ---
    delimiters
  3. Allowed Properties: Only 
    name
    , 
    description
    , 
    license
    , 
    compatibility
    , 
    metadata
    , 
    allowed-tools
  4. Name Validation:
    • Hyphen-case only (lowercase, digits, hyphens)
    • No start/end hyphens, no consecutive hyphens (
      --
      )
    • Max 64 characters
    • Matches directory name
  5. Description Validation:
    • No angle brackets (
      <
      or 
      >
      )
    • Max 1024 characters
    • Non-empty
If validation fails: Stop the review and report the specific validation error(s). The skill must pass basic validation before proceeding with the full review.
使用验证清单（
references/validate.md
），验证Skill是否通过所有基础检查：
  1. 文件结构：SKILL.md存在
  2. 前置元数据格式：
    ---
    分隔符之间的YAML格式有效
  3. 允许的属性：仅包含
    name
    、
    description
    、
    license
    、
    compatibility
    、
    metadata
    、
    allowed-tools
  4. 名称验证：
    • 仅使用连字符格式（小写、数字、连字符）
    • 首尾无连字符，无连续连字符（
      --
      ）
    • 最多64个字符
    • 与目录名称匹配
  5. 描述验证：
    • 无尖括号（
      <
      或
      >
      ）
    • 最多1024个字符
    • 非空
如果验证失败：停止审核并报告具体的验证错误。Skill必须通过基础验证后，才能进行完整审核。

Step 3: Read the Skill

步骤3：阅读Skill

Read the complete skill structure:
  • SKILL.md
    (frontmatter and body)
  • Any files in 
    references/
    , 
    scripts/
    , 
    assets/
    directories
IMPORTANT: Only analyze the skill provided by the user.
阅读完整的Skill结构：
  • SKILL.md
    （前置元数据和主体）
  • references/
    、
    scripts/
    、
    assets/
    目录中的所有文件
重要提示：仅分析用户提供的Skill。

Step 4: Verify Spec Compliance

步骤4：验证规范合规性

Check that the skill follows the Agent Skills specification (
references/spec.md
). Verify:
检查Skill是否遵循Agent Skills规范（
references/spec.md
）。验证：

Directory Structure

目录结构

  • Skill is in a directory matching the 
    name
    field
  • Contains required 
    SKILL.md
    file
  • Optional directories follow conventions: 
    scripts/
    , 
    references/
    , 
    assets/
  • Skill位于与
    name
    字段匹配的目录中
  • 包含必需的
    SKILL.md
    文件
  • 可选目录遵循约定：
    scripts/
    、
    references/
    、
    assets/

Frontmatter Compliance

前置元数据合规性

FieldCheck
name
1-64 chars, lowercase alphanumeric + hyphens, no start/end hyphens, no 
--
, matches directory name
description
1-1024 chars, non-empty, describes what and when
license
If present, short (license name or file reference)
compatibility
If present, max 500 chars
metadata
If present, string keys to string values
allowed-tools
If present, space-delimited tool list
字段检查项
name
1-64个字符，小写字母数字+连字符，首尾无连字符，无
--
，与目录名称匹配
description
1-1024个字符，非空，描述用途和适用场景
license
若存在，需简短（许可证名称或文件引用）
compatibility
若存在，最多500个字符
metadata
若存在，字符串键对应字符串值
allowed-tools
若存在，空格分隔的工具列表

Body Content

主体内容

  • Markdown format after frontmatter
  • Recommended: step-by-step instructions, examples, edge cases
  • Under 500 lines (move detailed content to references)
  • 前置元数据后使用Markdown格式
  • 推荐：分步说明、示例、边缘情况
  • 不超过500行（详细内容移至参考资料）

Progressive Disclosure

渐进式披露

  • Metadata (~100 tokens): name + description loaded at startup
  • Instructions (<5000 tokens recommended): SKILL.md body loaded on activation
  • Resources (as needed): scripts/references/assets loaded on demand
  • 元数据（约100个token）：启动时加载名称+描述
  • 说明（建议不超过5000个token）：激活时加载SKILL.md主体
  • 资源（按需）：脚本/参考资料/资源文件按需加载

File References

文件引用

  • Use relative paths from skill root
  • Keep references one level deep (avoid deeply nested chains)
If spec violations found: Document them clearly in the review output with specific fixes.
  • 使用相对于Skill根目录的路径
  • 引用仅保留一层深度（避免深层嵌套链）
如果发现规范违规：在审核输出中清晰记录，并提供具体的修复方案。

Step 5: Analyze the Skill

步骤5：分析Skill

Perform analysis in four areas, comparing against the reference skills you read from Anthropic's repository:
从四个方面进行分析，并与您从Anthropic仓库中读取的参考Skill进行对比：

A. Length Analysis

A. 长度分析

Using the progressive disclosure guidelines from skill-creator, evaluate:
  • Word count in 
    description
    field
  • Line/word count in SKILL.md body
  • Number and size of reference files
  • Duplication between SKILL.md and reference files
使用skill-creator中的渐进式披露指南，评估：
  • description
    字段的字数
  • SKILL.md主体的行数/字数
  • 参考文件的数量和大小
  • SKILL.md与参考文件之间的重复内容

B. Intent Scope Analysis

B. 意图范围分析

Evaluate:
  • All intents the skill serves
  • Whether skill handles multiple distinct use cases
  • Whether splitting would improve triggering accuracy
  • Trade-offs: context efficiency vs. maintenance overhead
Questions to answer:
  • Does this skill try to do too much?
  • Are there distinct user intents that deserve separate skills?
评估：
  • Skill服务的所有意图
  • Skill是否处理多个不同的用例
  • 拆分是否能提高触发准确性
  • 权衡：上下文效率与维护成本
需要回答的问题：
  • 这个Skill的功能是否过于繁杂？
  • 是否存在值得拆分为独立Skill的不同用户意图？

C. Trigger Analysis (CRITICAL)

C. 触发分析（关键）

The 
description
field is the primary triggering mechanism. Evaluate it for three types of triggers:
Trigger TypeWhat to Check
User INTENTDoes it describe what the user wants to do? (e.g., "deploy", "create", "edit")
TECHNICAL contextDoes it mention code patterns, file types, imports? (e.g., "base44.entities.*", ".jsonc files")
Project stackDoes it mention frameworks, tools, file structures? (e.g., "Vite", "Next.js", "base44/")
Check:
  • Does description cover both intent-based AND technical triggers?
  • Is it specific enough to trigger correctly, but broad enough to not miss cases?
  • Are there gaps where the skill might not trigger when it should?
  • Does it clearly distinguish from similar skills?
Good trigger pattern example:
ACTIVATE when (1) INTENT - user wants to [action]; (2) TECHNICAL - code contains [patterns], uses [APIs]; (3) CONTEXT - project has [structure/files]
description
字段是主要触发机制。从三种触发类型进行评估：
触发类型检查项
用户意图是否描述了用户想要执行的操作？（例如："部署"、"创建"、"编辑"）
技术上下文是否提及代码模式、文件类型、导入？（例如："base44.entities.*"、".jsonc文件"）
项目栈是否提及框架、工具、文件结构？（例如："Vite"、"Next.js"、"base44/"）
检查：
  • 描述是否同时涵盖基于意图和技术的触发条件？
  • 是否足够具体以确保正确触发，但又足够宽泛以不遗漏场景？
  • 是否存在Skill应该触发但未触发的空白？
  • 是否能与类似Skill明确区分？
良好的触发模式示例：
当满足以下条件时激活：(1) 意图 - 用户想要[执行操作]；(2) 技术 - 代码包含[模式]，使用[API]；(3) 上下文 - 项目具有[结构/文件]

Step 6: Provide Recommendations

步骤6：提供建议

Summarize findings with actionable recommendations for:
  1. Spec Compliance: What needs to be fixed to follow the spec?
  2. Length: What should be trimmed or split?
  3. Intent Scope: Should it be split or combined?
  4. Triggers: How can the description be improved?
总结发现并提供可操作的建议，包括：
  1. 规范合规性：需要修复哪些内容以遵循规范？
  2. 长度：哪些内容应精简或拆分？
  3. 意图范围：是否应该拆分或合并？
  4. 触发条件：如何改进描述？

Output Format

输出格式

undefined
undefined

Skill Review: [Skill Name]

Skill审核：[Skill名称]

Reference Skills Compared

对比的参考Skill

  • [List the 3-5 Anthropic skills you read before this review]
  • [列出本次审核前您阅读的3-5个Anthropic Skill]

Summary

摘要

[1-2 sentence overview]
[1-2句话概述]

Validation Result

验证结果

  • Status: [Pass/Fail]
  • Details: [Validation output or errors]
  • 状态：[通过/失败]
  • 详情：[验证输出或错误]

Spec Compliance

规范合规性

  • Directory structure: [Pass/Fail - details]
  • Frontmatter fields: [Pass/Fail - details]
  • Body content: [Pass/Recommendations]
  • Progressive disclosure: [Pass/Recommendations]
  • File references: [Pass/Recommendations]
  • Assessment: [Compliant/Partially compliant/Non-compliant]
  • Fixes Required: [List of specific fixes if any]
  • 目录结构：[通过/失败 - 详情]
  • 前置元数据字段：[通过/失败 - 详情]
  • 主体内容：[通过/建议]
  • 渐进式披露：[通过/建议]
  • 文件引用：[通过/建议]
  • 评估：[合规/部分合规/不合规]
  • 需修复内容：[具体修复列表（如有）]

Length Analysis

长度分析

  • Description: X words
  • SKILL.md body: X lines / X words
  • Reference files: X files
  • Assessment: [Pass/Needs attention]
  • Recommendations: [Specific suggestions]
  • 描述：X字
  • SKILL.md主体：X行 / X字
  • 参考文件：X个
  • 评估：[通过/需关注]
  • 建议：[具体建议]

Intent Scope Analysis

意图范围分析

  • Intents served: [List]
  • Assessment: [Focused/Broad/Too broad]
  • Recommendations: [Split suggestions if applicable]
  • 服务的意图：[列表]
  • 评估：[聚焦/宽泛/过于宽泛]
  • 建议：[拆分建议（如适用）]

Trigger Analysis

触发分析

  • Intent coverage: [Yes/Partial/No]
  • Technical coverage: [Yes/Partial/No]
  • Stack coverage: [Yes/Partial/No]
  • Assessment: [Strong/Adequate/Weak]
  • Recommendations: [Specific description improvements]
  • 意图覆盖：[是/部分/否]
  • 技术覆盖：[是/部分/否]
  • 项目栈覆盖：[是/部分/否]
  • 评估：[优秀/充足/薄弱]
  • 建议：[具体的描述改进方案]

Overall Recommendations

总体建议

  1. [Priority 1 action item - spec compliance fixes if any]
  2. [Priority 2 action item]
  3. [Priority 3 action item]
  4. [Priority 4 action item]
undefined
  1. [优先级1行动项 - 规范合规性修复（如有）]
  2. [优先级2行动项]
  3. [优先级3行动项]
  4. [优先级4行动项]
undefined