SKILL.md 实战指南：从资源库选择到生产级 Skill 编写

my-skill/
├── SKILL.md              # Skill 的入口说明
├── scripts/              # 可执行脚本
│   └── process.py
├── templates/            # 模板文件
│   └── report.md
├── examples/             # 输入输出示例
│   ├── input.csv
│   └── output.md
└── assets/               # 需要复用的静态资源

---
name: csv-data-summarizer
description: Analyze CSV files, summarize key statistics, detect anomalies, and generate a Markdown report.
---

# CSV Data Summarizer

## When to use

Use this skill when the user provides one or more CSV files and asks for:

- data profiling
- summary statistics
- anomaly detection
- missing value analysis
- a Markdown data report

Do not use this skill for real-time database queries or files larger than the execution environment can safely process.

## Input requirements

The input should be one or more `.csv` files. Before analysis:

1. Check file encoding.
2. Inspect column names and row count.
3. Detect delimiter automatically if possible.
4. Report unreadable files before continuing.

## Workflow

1. Load the CSV file with Python.
2. Generate basic statistics for numeric columns.
3. Count missing values for every column.
4. Detect duplicated rows.
5. Identify obvious outliers using interquartile range.
6. Produce a Markdown report using `templates/report.md`.

## Commands

Run:

```bash
python scripts/analyze_csv.py --input <csv-file> --output report.md

这个例子不复杂，但已经具备生产级 Skill 的基本骨架：触发条件清楚、流程可执行、异常有处理、输出能检查。

## Skills 聚合网站：适合查找灵感和现成能力

如果目标是快速了解社区里已经有哪些 Skill，聚合网站比直接翻 GitHub 更方便。它们通常提供搜索、分类、更新时间和入口链接，适合用来找方向。

| 网站 | 地址 | 适合用途 |
|---|---|---|
| SkillsMP | https://skillsmp.com/ | 搜索不同类型的 Skills，快速了解分类 |
| AI Template Skills | https://www.aitmpl.com/skills | 查找模板化 Skill，适合看命名和描述方式 |
| Claude Marketplaces | https://claudemarketplaces.com/ | 浏览 Claude 生态里的 Skill 和插件资源 |

聚合网站适合“找样本”，但不适合直接判断质量。一个 Skill 是否真的好用，还要打开仓库看它的目录结构、`SKILL.md` 写法、脚本实现和更新记录。

## 官方 Skills：最适合学习目录组织和质量标准

### Anthropic 官方 Skills

地址：

```text
https://github.com/anthropics/skills

/plugin marketplace add anthropics/skills
/plugin install document-skills@anthropic-agent-skills

git clone https://github.com/anthropics/skills.git
cd skills

1. description 是否足够具体
2. 触发条件是否明确
3. 是否有可复用脚本
4. 是否有模板和示例
5. 是否写了失败处理
6. 是否有输出检查规则

https://github.com/eliasjudin/oai-skills

code-review-skill/
├── SKILL.md
├── checklists/
│   ├── security.md
│   ├── performance.md
│   └── maintainability.md
└── scripts/
    └── collect_diff.sh

Help users process documents.

Create, edit, inspect, and summarize Microsoft Word documents while preserving structure, headings, tables, and comments.

Do not use this skill for scanned PDFs that require OCR unless an OCR tool is available.

python scripts/extract_tables.py --input report.pdf --output tables/

# 数据分析报告

## 1. 数据概况

- 文件名：
- 行数：
- 列数：
- 时间范围：

## 2. 关键发现

| 发现 | 证据 | 影响 |
|---|---|---|

## 3. 风险和异常

## 4. 建议动作

Before final response, verify:

- every finding has file path and line reference
- security issues are separated from style suggestions
- uncertain findings are marked as assumptions
- no generated code is proposed without explaining trade-offs

https://github.com/anthropics/skills/tree/main/skills/skill-creator

我想创建一个 Skill，用来分析产品经理导出的需求文档。
输入通常是 Markdown 或 Word 文档。
它需要识别功能点、用户角色、接口依赖、待确认问题，并输出一份评审清单。
结果要按“阻塞问题、风险问题、建议优化”三类组织。

samples/
├── normal/
│   └── user-profile-requirement.md
├── edge/
│   └── requirement-with-missing-api.md
└── negative/
    └── meeting-summary-not-requirement.md

刚才这个 Skill 有两个问题：
1. 它把会议纪要误判成需求文档；
2. 它没有把接口依赖单独列出来。

请修改 SKILL.md：
- 增加“不适用场景”
- 在工作流中加入接口依赖提取步骤
- 在最终检查清单中要求列出待确认接口

# Skill Acceptance Checklist

## Discovery

- [ ] description clearly identifies the task
- [ ] unsuitable scenarios are listed
- [ ] required input types are explicit

## Execution

- [ ] all referenced scripts exist
- [ ] commands can run from the skill directory
- [ ] dependencies are documented

## Output

- [ ] output format is specified
- [ ] examples are provided
- [ ] final checklist is included

## Maintenance

- [ ] file names are readable
- [ ] templates are separated from instructions
- [ ] changes can be tested with sample inputs

## Output format

Respond in Chinese. Use the following table:

| Severity | File | Line | Issue | Recommendation |
|---|---|---|---|---|

Sort findings by severity:

1. Critical
2. High
3. Medium
4. Low

skills/
├── document/
│   ├── word-editor/
│   └── pdf-inspector/
├── development/
│   ├── code-review/
│   └── mcp-server-builder/
├── data/
│   ├── csv-profiler/
│   └── meeting-insights/
├── design/
│   ├── brand-guidelines/
│   └── theme-factory/
└── README.md

# Skills Library

| Skill | Category | Purpose | Status |
|---|---|---|---|
| code-review | development | Review Git diffs | stable |
| csv-profiler | data | Analyze CSV files | testing |
| brand-guidelines | design | Generate brand-consistent assets | draft |