Back to Details

create-ultimate-skill

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Create Ultimate Skill

创建终极Skill

Creation Workflow

创建工作流

When creating a new skill based on user's request: $ARGUMENTS
当根据用户请求创建新Skill时：$ARGUMENTS

Phase 1: Define Use Cases

阶段1：定义使用场景

Ask the user:
  1. Use cases: "Describe 2-3 specific scenarios where you'd want this skill to activate. What would you say to Claude, and what should it do?"
  2. Non-use cases: "What should this skill NOT be used for?" (defines negative triggers)
询问用户：
  1. 使用场景："描述2-3个你希望激活该Skill的具体场景。你会对Claude说什么，它应该做什么？"
  2. 非使用场景："该Skill不应该用于什么场景？"（定义负向触发条件）

Phase 2: Choose Archetype and Scope

阶段2：选择原型与范围

Ask the user:
  1. Archetype (see Archetypes section below)
  2. Triggers: What phrases should activate this skill?
  3. Scope: What's in scope vs out of scope?
询问用户：
  1. 原型（见下文原型章节）
  2. 触发词：哪些短语应激活该Skill？
  3. 范围：哪些属于范围内，哪些属于范围外？

Phase 3: Fetch Latest Documentation

阶段3：获取最新文档

Fetch the latest official skill documentation to verify current fields and constraints:
WebFetch: https://markdown.new/https://code.claude.com/docs/en/skills.md
获取最新官方Skill文档以验证当前字段和约束：
WebFetch: https://markdown.new/https://code.claude.com/docs/en/skills.md

Phase 4: Choose Save Location

阶段4：选择保存位置

Use AskUserQuestion:
  • Personal (home folder): 
    ~/.claude/skills/<skill-name>/
    - available in all projects
  • Project folder: 
    .claude/skills/<skill-name>/
    - project-specific
  • Existing plugin: ask which plugin, then 
    plugins/<plugin-name>/skills/
使用AskUserQuestion：
  • 个人目录（主文件夹）：
    ~/.claude/skills/<skill-name>/
    - 所有项目均可使用
  • 项目目录：
    .claude/skills/<skill-name>/
    - 仅特定项目可用
  • 现有插件：询问具体插件，然后保存到
    plugins/<plugin-name>/skills/

Phase 5: Create the Skill

阶段5：创建Skill

Scaffold with init_skill.py or create manually:
bash
python .claude/skills/create-ultimate-skill/scripts/init_skill.py <skill-name> --path <location>
Then write SKILL.md applying all rules below. Review with the checklist before presenting to user.
使用init_skill.py初始化或手动创建：
bash
python .claude/skills/create-ultimate-skill/scripts/init_skill.py <skill-name> --path <location>
然后按照以下所有规则编写SKILL.md。在提交给用户前，使用检查清单进行审核。

Phase 6: Summary

阶段6：总结

When complete, provide:
  • Skill location and how to trigger it
  • What it does
  • Suggestions for improvement
  • Testing prompts the user can try
完成后，提供以下信息：
  • Skill的位置及触发方式
  • 它的功能
  • 改进建议
  • 用户可尝试的测试提示

Frontmatter Hard Rules

Frontmatter硬性规则

  • description
    MUST be a single-line plain string. NEVER use YAML multi-line indicators (
    >-
    , 
    |
    , 
    >
    , 
    |-
    ). Multi-line descriptions break frontmatter parsing.
  • allowed-tools
    MUST be a single-line comma-separated string (e.g. 
    allowed-tools: Read, Write, Edit
    ). NEVER use YAML list syntax.
  • description
    必须是单行纯文本字符串。绝不能使用YAML多行标识（
    >-
    , 
    |
    , 
    >
    , 
    |-
    ）。多行描述会破坏frontmatter解析。
  • allowed-tools
    必须是单行逗号分隔的字符串（例如
    allowed-tools: Read, Write, Edit
    ）。绝不能使用YAML列表语法。

Required Fields

必填字段

FieldConstraints
name
Max 64 chars, 
[a-z0-9-]
only, no "anthropic"/"claude"
description
Max 1024 chars, non-empty, no XML tags
字段约束条件
name
最多64个字符，仅允许
[a-z0-9-]
，不能包含"anthropic"/"claude"
description
最多1024个字符，非空，无XML标签

Optional Fields

可选字段

FieldPurposeExample
allowed-tools
Limit tool access
Read, Grep, Glob
model
Override model
sonnet
, 
opus
context
Isolation
fork
agent
Subagent type
Explore
, 
Plan
hooks
Lifecycle events
PreToolUse
, 
PostToolUse
, 
Stop
user-invocable
Slash menu visibility
false
to hide
disable-model-invocation
Block Skill tool
true
See references/api-reference.md for full field documentation.
字段用途示例
allowed-tools
限制工具访问权限
Read, Grep, Glob
model
覆盖默认模型
sonnet
, 
opus
context
隔离性设置
fork
agent
子Agent类型
Explore
, 
Plan
hooks
生命周期事件
PreToolUse
, 
PostToolUse
, 
Stop
user-invocable
Slash菜单可见性
false
表示隐藏
disable-model-invocation
阻止Skill工具
true
完整字段文档请参见references/api-reference.md

Archetypes

原型

ArchetypeBest ForStructure
CLI ReferenceTool documentationCommands grouped by function, minimal prose
MethodologyWorkflows, processesPhilosophy + THE EXACT PROMPT + examples
Safety ToolValidation, securityThreat model + risk tiers + rules
OrchestrationMulti-agent coordinationQuick start + APIs + integrations
See references/archetypes.md for templates.
原型适用场景结构
CLI参考工具文档按功能分组的命令，尽量少用叙述性文字
方法论工作流、流程理念 + 精确提示词 + 示例
安全工具验证、安全威胁模型 + 风险等级 + 规则
编排工具多Agent协调快速入门 + API + 集成
模板请参见references/archetypes.md

Description Rules

描述规则

  • Third person always ("Processes files" not "I help you")
  • Include WHAT it does AND WHEN to trigger
  • Specific trigger phrases users would say
  • Add negative triggers ("Do NOT use for...") to prevent over-triggering
  • Max 1024 characters
Template:
yaml
description: [What it does]. Use when [triggers]. Do NOT use for [exclusions].
Examples:
yaml
description: Extract text and tables from PDF files, fill forms, merge documents. Use when working with PDF files or when user mentions PDFs, forms, or document extraction.

description: Process and transform CSV data files. Use when working with CSV imports, exports, or data cleaning. Do NOT use for Excel files (.xlsx) or database queries.
  • 始终使用第三人称（例如“处理文件”而非“我帮你处理文件”）
  • 需包含功能描述及触发时机
  • 列出用户会使用的具体触发短语
  • 添加负向触发条件（“请勿用于...”）以避免误触发
  • 最多1024个字符
模板：
yaml
description: [功能描述]. Use when [触发时机]. Do NOT use for [排除场景].
示例：
yaml
description: Extract text and tables from PDF files, fill forms, merge documents. Use when working with PDF files or when user mentions PDFs, forms, or document extraction.

description: Process and transform CSV data files. Use when working with CSV imports, exports, or data cleaning. Do NOT use for Excel files (.xlsx) or database queries.

SKILL.md Body Rules

SKILL.md正文规则

  • Imperative voice ("Extract the data", "Run validation")
  • Never second person ("You should...")
  • Concise - challenge each line: "Does Claude need this?"
  • Under 500 lines, under 5,000 words (~2,000 ideal)
  • Put critical instructions at the top of the file
  • Use scripts for validation where possible - code is deterministic, language interpretation is not
  • "Take your time to do this thoroughly" is more effective in user prompts than in SKILL.md
  • 使用祈使语气（例如“提取数据”、“运行验证”）
  • 绝不能使用第二人称（例如“你应该...”）
  • 简洁凝练——对每一行都要自问：“Claude需要这个吗？”
  • 不超过500行，不超过5000字（理想为2000字左右）
  • 将关键说明放在文件顶部
  • 尽可能使用脚本进行验证——代码是确定性的，而语言解释具有不确定性
  • 在用户提示中使用“请仔细完成此任务”比在SKILL.md中更有效

Structure Rules

结构规则

skill-name/
├── SKILL.md              # Core guidance only
├── references/           # Detailed docs (on-demand)
├── examples/             # Working code samples
└── scripts/              # Executable utilities
  • References ONE level deep only - no chains
  • Long files (>100 lines): include TOC
  • Never include README.md in skill folder (SKILL.md IS the readme)
  • Forward slashes only (no Windows paths)
skill-name/
├── SKILL.md              # Core guidance only
├── references/           # Detailed docs (on-demand)
├── examples/             # Working code samples
└── scripts/              # Executable utilities
  • 参考文档仅允许一级深度——不能有嵌套链
  • 长文件（>100行）：需包含目录
  • Skill文件夹中绝不能包含README.md（SKILL.md即为说明文档）
  • 仅使用正斜杠（不使用Windows路径）

Anti-Patterns

反模式

Anti-PatternFix
Vague descriptionsSpecific triggers + negative triggers
Deeply nested referencesOne level deep
README.md in skill folderDelete it - SKILL.md IS the readme
>-
or `		` in description
YAML list for allowed-toolsComma-separated string
反模式修复方案
模糊的描述添加具体触发词和负向触发条件
深度嵌套的参考文档限制为一级深度
Skill文件夹中包含README.md删除它——SKILL.md即为说明文档
描述中使用
>-
或`		`
allowed-tools使用YAML列表格式改为逗号分隔的字符串

Testing

测试

Define success criteria before building:
  • Quantitative: Trigger rate >90% of relevant queries, correct tool sequence, task completes without user intervention
  • Qualitative: Output matches expected format, no hallucinated commands, consistent across sessions
Debug technique: Ask Claude "When would you use the [skill name] skill?" - the response reveals how Claude interprets the description.
Test three areas:
  1. Triggering: Direct trigger phrases, paraphrased requests, and edge cases that should NOT trigger
  2. Functional: Run each use case end-to-end, check output format and accuracy
  3. Performance: Compare skill-assisted vs manual - does the skill save time and improve quality?
Pro tip: Iterate on a single challenging task until Claude succeeds, then extract the winning approach into the skill. This grounds design in proven patterns rather than theory.
在构建前定义成功标准：
  • 量化标准：相关查询的触发率>90%，工具序列正确，无需用户干预即可完成任务
  • 定性标准：输出符合预期格式，无幻觉命令，跨会话保持一致
**调试技巧：**询问Claude "你会在什么时候使用[Skill名称]？"——回复会揭示Claude对描述的理解方式。
测试三个方面：
  1. 触发测试：直接触发短语、改写后的请求、以及不应触发的边缘情况
  2. 功能测试：端到端运行每个使用场景，检查输出格式和准确性
  3. 性能测试：对比使用Skill和手动操作——Skill是否节省时间并提升质量？
**专业提示：**反复迭代单个具有挑战性的任务，直到Claude成功完成，然后将成功方法提炼到Skill中。这会让设计基于已验证的模式而非理论。

Iteration Signals

迭代信号

SignalFix
Skill never activatesAdd more trigger phrases to description
Wrong skill activatesAdd "Do NOT use for..." negative triggers
Output is wrongRefine SKILL.md instructions, add examples
Output is inconsistentReduce degrees of freedom, be more prescriptive
信号修复方案
Skill从未激活在描述中添加更多触发短语
错误的Skill被激活添加“请勿用于...”负向触发条件
输出错误优化SKILL.md说明，添加示例
输出不一致减少自由度，更具指令性

Utility Scripts

实用脚本

init_skill.py

init_skill.py

bash
python .claude/skills/create-ultimate-skill/scripts/init_skill.py <skill-name> --path <location>
Creates SKILL.md with template, plus scripts/, references/, assets/ directories.
bash
python .claude/skills/create-ultimate-skill/scripts/init_skill.py <skill-name> --path <location>
创建带模板的SKILL.md，以及scripts/、references/、assets/目录。

package_skill.py

package_skill.py

bash
python .claude/skills/create-ultimate-skill/scripts/package_skill.py <path/to/skill> [output-dir]
Validates and creates distributable zip. Checks frontmatter, naming, ensures no TODOs remain.
bash
python .claude/skills/create-ultimate-skill/scripts/package_skill.py <path/to/skill> [output-dir]
验证并创建可分发的zip包。检查frontmatter、命名规则，确保没有遗留的TODO项。

Skill Review Mode

Skill审核模式

When reviewing an existing skill (triggered by 
--review
or "review my skill"):
  1. Read SKILL.md frontmatter, body, and supporting directories
  2. Categorise issues by severity:
    • Critical: Won't trigger or wrong output (missing description, vague triggers, broken refs)
    • Major: Significantly reduces effectiveness (>5,000 words, no progressive disclosure)
    • Minor: Polish (inconsistent terminology, missing TOC, no negative triggers)
  3. Generate structured review report:
markdown
undefined
当审核现有Skill时（由
--review
或“审核我的Skill”触发）：
  1. 阅读SKILL.md的frontmatter、正文及支持目录
  2. 按严重程度分类问题：
    • 严重：无法触发或输出错误（缺少描述、模糊的触发词、损坏的参考）
    • 主要：显著降低有效性（>5000字、无渐进式披露）
    • 次要：优化问题（术语不一致、缺少目录、无负向触发条件）
  3. 生成结构化审核报告：
markdown
undefined

Skill Review: [skill-name]

Skill Review: [skill-name]

Summary

Summary

[Assessment, word counts, structure]
[Assessment, word counts, structure]

Description Analysis

Description Analysis

Current: [description] Issues: [with severity] Suggested: "[improved version]"
Current: [description] Issues: [with severity] Suggested: "[improved version]"

Issues by Severity

Issues by Severity

Critical ([count])

Critical ([count])

  • [File]: [Issue] - [Fix]
  • [File]: [Issue] - [Fix]

Major ([count])

Major ([count])

  • [File]: [Issue] - [Fix]
  • [File]: [Issue] - [Fix]

Minor ([count])

Minor ([count])

  • [File]: [Issue] - [Suggestion]
  • [File]: [Issue] - [Suggestion]

Rating: [Pass / Needs Improvement / Needs Major Revision]

Rating: [Pass / Needs Improvement / Needs Major Revision]

Priority Fixes

Priority Fixes

  1. [Highest impact]
  2. [Second]
  3. [Third]
undefined
  1. [Highest impact]
  2. [Second]
  3. [Third]
undefined

Review Checklist

审核检查清单

  • Use cases: 2-3 concrete scenarios defined
  • name: lowercase, hyphens, ≤64 chars
  • description: single-line plain string, third person, specific triggers, ≤1024 chars
  • description: negative triggers if needed
  • SKILL.md body: <500 lines, <5,000 words
  • No README.md in skill folder
  • References: one level deep
  • Forward slashes only
  • Examples concrete, not abstract
  • 使用场景：已定义2-3个具体场景
  • name：小写、连字符、≤64字符
  • description：单行纯文本字符串、第三人称、具体触发词、≤1024字符
  • description：按需添加负向触发条件
  • SKILL.md正文：<500行、<5000字
  • Skill文件夹中无README.md
  • 参考文档：仅一级深度
  • 仅使用正斜杠
  • 示例具体，而非抽象

Latest Documentation

最新文档

Before finalising any skill's frontmatter, fetch the latest official documentation to check for new fields, changed constraints, or updated rules:
WebFetch: https://markdown.new/https://code.claude.com/docs/en/skills.md
在最终确定任何Skill的frontmatter之前，获取最新官方文档以检查是否有新字段、已更改的约束或更新的规则：
WebFetch: https://markdown.new/https://code.claude.com/docs/en/skills.md

Further Reading

扩展阅读

  • Skill Archetypes
  • Official API Reference
  • Skill原型
  • 官方API参考