Skill Writer
2026-04-01
新闻来源:网淘吧
围观:31
电脑广告
手机广告
技能撰写器
为ClawdHub注册表编写结构良好、有效的SKILL.md文件。涵盖技能格式规范、前言模式、内容模式、示例质量和常见反模式。
使用时机
- 从头开始创建新技能时
- 将技术内容构建为智能体技能时
- 编写能被注册表正确索引的前言时
- 为不同类型技能选择章节组织方式时
- 在发布前审查自己的技能时
SKILL.md格式
技能是包含YAML前言的单个Markdown文件。智能体按需加载并遵循其指令。

---
name: my-skill-slug
description: One-sentence description of when to use this skill.
metadata: {"clawdbot":{"emoji":"🔧","requires":{"anyBins":["tool1","tool2"]},"os":["linux","darwin","win32"]}}
---
# Skill Title
One-paragraph summary of what this skill covers.
## When to Use
- Bullet list of trigger scenarios
## Main Content Sections
### Subsection with examples
Code blocks, commands, patterns...
## Tips
- Practical advice bullets
前言模式
名称(必填)
技能的标识符。必须与发布时使用的名称匹配。
name: my-skill
规则:
- 小写字母,用连字符连接:
csv-pipeline、git-workflows - 无空格无下划线
- 保持简短且描述性(1-3个词)
- 发布前检查slug冲突:
npx molthub@latest search "your-slug"
描述(必填)
最重要的字段。这是用于:
- 注册表进行语义搜索(向量嵌入)的索引依据
- 智能体读取以决定是否激活此技能
- 用户在浏览搜索结果时看到的内容
# GOOD: Specific triggers and scope
description: Write Makefiles for any project type. Use when setting up build automation, defining multi-target builds, managing dependencies between tasks, creating project task runners, or using Make for non-C projects (Go, Python, Docker, Node.js). Also covers Just and Task as modern alternatives.
# BAD: Vague, no triggers
description: A skill about Makefiles.
# BAD: Too long (gets truncated in search results)
description: This skill covers everything you need to know about Makefiles including variables, targets, prerequisites, pattern rules, automatic variables, phony targets, conditional logic, multi-directory builds, includes, silent execution, and also covers Just and Task as modern alternatives to Make for projects that use Go, Python, Docker, or Node.js...
有效描述的模式:
[What it does]. Use when [trigger 1], [trigger 2], [trigger 3]. Also covers [related topic].
元数据(必填)
符合以下架构的JSON对象:clawdbot架构:
metadata: {"clawdbot":{"emoji":"🔧","requires":{"anyBins":["make","just"]},"os":["linux","darwin","win32"]}}
字段:
emoji:在注册表列表中显示的单表情符号requires.anyBins:技能所需的CLI工具数组(至少有一个必须可用)操作系统支持的平台列表:"linux","darwin"(macOS),"win32"(Windows)
请谨慎选择requires.anyBins:
# Good: lists the actual tools the skill's commands use
"requires": {"anyBins": ["docker", "docker-compose"]}
# Bad: lists generic tools every system has
"requires": {"anyBins": ["bash", "echo"]}
# Good for skills that work via multiple tools
"requires": {"anyBins": ["make", "just", "task"]}
内容结构
"何时使用" 章节
此章节必须紧接在标题段落之后。它向代理(及用户)说明本技能适用的具体场景。
## When to Use
- Automating build, test, lint, deploy commands
- Defining dependencies between tasks (build before test)
- Creating a project-level task runner
- Replacing long CLI commands with short targets
规则:
- 4-8 个要点
- 每个要点应描述具体场景,而非抽象概念
- 以动词或动名词开头:"自动化...", "调试...", "转换..."
- 不要逐字重复描述字段的内容
主要内容章节
按任务而非概念组织。代理需要为特定情况找到合适的命令。
## GOOD: Organized by task
## Encode and Decode
### Base64
### URL Encoding
### Hex
## BAD: Organized by abstraction
## Theory of Encoding
## Encoding Types
## Advanced Topics
代码块
每个部分都应至少包含一个代码块。没有代码块的技能只是观点,而非工具。
## GOOD: Concrete, runnable example
```bash
# Encode a string to Base64
echo -n "Hello, World!" | base64
# SGVsbG8sIFdvcmxkIQ==
```
## BAD: Abstract description
Base64 encoding converts binary data to ASCII text using a 64-character alphabet...
代码块最佳实践:
- 始终指定编程语言(
bash、python、javascript、yaml、sql等) - 在命令下方以注释形式展示输出使用实际值
- ,而非foo
/bar(应使用myapp、api-server、真实IP格式等), real IP formats) - 首先包含最常见的情况,然后是变体
- 为非显而易见的标志或参数添加内联注释
多语言覆盖
如果某项技能适用于多种语言,请使用一致的章节结构:
## Hashing
### Bash
```bash
echo -n "Hello" | sha256sum
JavaScript
const crypto = require('crypto');
crypto.createHash('sha256').update('Hello').digest('hex');
Python
import hashlib
hashlib.sha256(b"Hello").hexdigest()
Order: Bash first (most universal), then by popularity for the topic.
### The "Tips" Section
End every skill with a Tips section. These are the distilled wisdom — the things that save hours of debugging.
```markdown
## Tips
- The number one Makefile bug: using spaces instead of tabs for indentation.
- SHA-256 is the standard for integrity checks. MD5 is fine for dedup but broken for cryptographic use.
- Never schedule critical cron jobs between 1:00-3:00 AM if DST applies.
规则:
- 5-10个要点
- 每条提示都是一个独立的见解(不依赖于其他提示)
- 优先考虑陷阱和非显而易见的行为,而非基本建议
- 避免使用“始终遵循最佳实践”的陈词滥调
技能类型与模板
CLI工具参考
适用于关于特定工具或命令系列的技能。
---
name: tool-name
description: [What tool does]. Use when [scenario 1], [scenario 2].
metadata: {"clawdbot":{"emoji":"🔧","requires":{"anyBins":["tool-name"]}}}
---
# Tool Name
[One paragraph: what it does and why you'd use it.]
## When to Use
- [4-6 scenarios]
## Quick Reference
[Most common commands with examples]
## Common Operations
### [Operation 1]
### [Operation 2]
## Advanced Patterns
### [Pattern 1]
## Troubleshooting
### [Common error and fix]
## Tips
语言/框架参考
适用于关于特定语言或框架中模式的技能。
---
name: pattern-name
description: [Pattern] in [language/framework]. Use when [scenario 1], [scenario 2].
metadata: {"clawdbot":{"emoji":"📐","requires":{"anyBins":["runtime"]}}}
---
# Pattern Name
## When to Use
## Quick Reference
[Cheat sheet / syntax summary]
## Patterns
### [Pattern 1 — with full example]
### [Pattern 2 — with full example]
## Cross-Language Comparison (if applicable)
## Anti-Patterns
[What NOT to do, with explanation]
## Tips
工作流程/过程指南
适用于关于多步骤过程的技能。
---
name: workflow-name
description: [Workflow description]. Use when [scenario 1], [scenario 2].
metadata: {"clawdbot":{"emoji":"🔄","requires":{"anyBins":["tool1","tool2"]}}}
---
# Workflow Name
## When to Use
## Prerequisites
[What needs to be set up first]
## Step-by-Step
### Step 1: [Action]
### Step 2: [Action]
### Step 3: [Action]
## Variations
### [Variation for different context]
## Troubleshooting
## Tips
反模式
过于抽象
# BAD
## Error Handling
Error handling is important for robust applications. You should always
handle errors properly to prevent unexpected crashes...
# GOOD
## Error Handling
```bash
# Bash: exit on any error
set -euo pipefail
# Trap for cleanup on exit
trap 'rm -f "$TMPFILE"' EXIT
### Too narrow
```markdown
# BAD: Only useful for one specific case
---
name: react-useeffect-cleanup
description: How to clean up useEffect hooks in React
---
# GOOD: Broad enough to be a real reference
---
name: react-hooks
description: React hooks patterns. Use when working with useState, useEffect, useCallback, useMemo, custom hooks, or debugging hook-related issues.
---
大段文字缺少示例
如果任何章节超过10行没有代码块,就说明文字内容过多。请用示例来分解内容。
缺少交叉引用
如果你的技能提到了另一个有独立技能的工具或概念,请注明:
# For Docker networking issues, see the `container-debug` skill.
# For regex syntax details, see the `regex-patterns` skill.
过时的命令
验证每条命令在当前工具版本下是否有效。常见陷阱:
- Docker Compose:
docker-compose(版本1)对比docker compose(版本2) - Python:
pip对比pip3、python对比python3 - Node.js:CommonJS(
require)对比ESM(import尺寸指南
指标
| 目标 | 过短 | 过长 | 总行数 |
|---|---|---|---|
| 300-550 | < 150 | > 700 | 章节 |
| 5-10 | < 3 | > 15 | 代码块 |
| 15-40 | < 8 | > 60 | 提示 |
| 5-10 | < 3 | > 15 | 少于150行的技能可能缺少示例。超过700行的技能应拆分为两个技能。 |
发布清单
发布前,请验证:
前置元数据是有效的YAML
- Frontmatter is valid YAML— 通过粘贴到YAML验证器中进行测试
- 描述以该技能的功能开头— 不要以“此技能...”或“一个用于...的技能”开头
- 每个部分至少包含一个代码块— 主要内容中没有纯文本部分
- 命令实际有效— 在干净环境中测试
- 没有遗留占位符值— 搜索
TODO、FIXME、example.com用作真实URL - Slug(唯一标识符)可用—
npx molthub@latest search "your-slug"返回无精确匹配项 requires.anyBins列出真实的依赖项— 技能命令实际调用的工具- 存在“提示”部分— 附带5个以上可操作、非显而易见的要点
发布
# Publish a new skill
npx molthub@latest publish ./skills/my-skill \
--slug my-skill \
--name "My Skill" \
--version 1.0.0 \
--changelog "Initial release"
# Update an existing skill
npx molthub@latest publish ./skills/my-skill \
--slug my-skill \
--name "My Skill" \
--version 1.1.0 \
--changelog "Added new section on X"
# Verify it's published
npx molthub@latest search "my-skill"
技巧
- 描述
字段是你的技能搜索排名的关键。花在它上面的时间应该超过任何一个单一的内容部分。要包含用户可能搜索的具体动词和名词。以最常见的用例开头。如果80%的用户需要"如何编码Base64",就把这个放在"如何在MessagePack和CBOR之间转换"之前。 - 每个代码示例都应该是可以直接复制粘贴的。如果需要未显示的设置,请补充设置步骤。
- 为智能体编写,而不是为人类。智能体需要可以逐步遵循的明确指令。避免使用"你可能想要考虑"——直接说"当Y时,执行X。"
- 通过让智能体在实际任务中使用你的技能来测试它。如果智能体无法按照指令产生正确结果,那么该技能就需要改进。
- 优先使用
- bash
代码块表示命令,即使在特定语言的技能中也是如此。智能体通常通过shell操作,bash代码块表示"运行这个"。不要重复 - --help
--help已提供的。专注于模式、组合以及那些--help不会教的内容。 - 按语义版本管理你的技能:补丁版本用于修正拼写错误,次要版本用于新增章节,主要版本用于重构。注册表会追踪版本历史。
文章底部电脑广告
手机广告位-内容正文底部
上一篇:Ai Compound 1.0.1
下一篇:AI Data Analysis


微信扫一扫,打赏作者吧~