Skill Extractor

Extract successful experiences and lessons learned from the current conversation context, and automatically generate reusable Claude Code Skill files.

Background

This is a "meta-skill" — a skill used to create other skills. By analyzing the tasks completed in the current conversation, extract key steps, common pitfalls, and best practices, and solidify one-time problem-solving into repeatable workflows.

Instructions

You are an experience extraction assistant, helping users transform work experiences in conversations into reusable Claude Code Skills. Please follow these steps:

Step 1: Analyze current conversation context

Review the entire conversation history, identify and extract the following information:

Original Requirement: What problem did the user initially want to solve?
Execution Steps: What key steps were taken to complete the task?
Lessons from Failures: Which attempts failed? What were the reasons for failure?
Successful Solution: What was the final solution adopted?
Key Findings: What important precautions, checkpoints, or tips are there?

Step 2: Collect user input

⚠️ Required: Use the AskUserQuestion tool to collect necessary information.

Use the AskUserQuestion tool to ask the user:

Skill Name: What is the name of this skill?
- Ask the user to input it, recommend using kebab-case format (e.g.,
```
api-caller
```
  ,
```
file-processor
```
  )
Core Function Description: A one-sentence description of what this skill does
- Ask the user to input it or make recommendations based on analysis results
Trigger Words: What words will users use to trigger this skill?
- Ask the user to input it, recommend 2-3 each in Chinese and English
Save Location:
- Options:
  - "Current plugin directory skills/ (Recommended)"
  - "Custom path"

Step 3: Decide whether a script is needed

Automatically determine whether a supporting script is needed based on the nature of the task (do not ask the user):

Situations where a script is needed:

Calling external APIs (such as OpenRouter, OpenAI, etc.)
Complex file processing (binary files, encoding conversion, etc.)
Dependence on specific libraries (such as ffmpeg, whisper, etc.)
Involving network requests, data parsing, etc.

Situations where no script is needed:

Pure prompt-driven workflows
Mainly combinations of bash commands
Simple file operations

Script language selection:

Prioritize Python, using uv script format (inline dependency declaration)
Use Bun when JavaScript is needed (built into Max, faster than Node.js)

Python uv script template:

python

#!/usr/bin/env -S uv run --script
# /// script
# requires-python = ">=3.11"
# dependencies = [
#     "requests",
#     "其他依赖",
# ]
# ///

import sys
# 脚本内容...

Step 4: Generate SKILL.md content

Based on the analysis results and user input, generate SKILL.md in the following format:

markdown

---
name: {skill-name}
description: {功能描述}。Use when user wants to {触发词1}, {触发词2}, {trigger1}, {trigger2}.
---

# {Skill Title}

{简要说明这个 skill 的功能和来源}

## Background

> 此 skill 从 [具体任务描述] 的实践经验中提取。

{描述这个 skill 解决的问题场景}

## Prerequisites

{列出依赖条件，如环境变量、工具等}

## Instructions

{详细的执行步骤，来自成功的方案}

### Step 1: {步骤标题}

{具体操作}

### Step 2: {步骤标题}

{具体操作}

...

## Common Pitfalls

{从失败经验中提取的常见陷阱}

| 错误做法 | 正确做法 | 原因 |
|---------|---------|------|
| ❌ {错误示例} | ✅ {正确示例} | {解释} |

## Best Practices

{从成功经验中提取的最佳实践}

- {实践1}
- {实践2}
- ...

## Troubleshooting

{常见问题及解决方案，来自实际遇到的错误}

**问题：{问题描述}**
- 原因：{原因分析}
- 解决：{解决方案}

## Example Interaction

{一个典型的使用示例}

用户：{示例输入}

助手：
1. {步骤1}
2. {步骤2}
...

Step 5: Write to file

Use the Write tool to write the generated content to the corresponding directory:

skills/{skill-name}/SKILL.md
skills/{skill-name}/{script-name}.py  # If a script is needed

If a Python script is created, add executable permissions:

bash

chmod +x skills/{skill-name}/{script-name}.py

Step 6: Update documentation (optional)

Ask the user if they need to update CLAUDE.md to register the new skill:

Update the project structure diagram
Update the Skills table

Step 7: Show results

After completion, inform the user:

The path of the created file
How to use the Skill
List of trigger words

Output Quality Guidelines

The generated SKILL.md should:

Executable: Clear steps that can be directly followed
Valuable: Contains truly useful experiences, not general statements
Pitfall-focused: Highlight failed experiences and common pitfalls
Code examples: Key steps should have code or command examples
Concise and clear: Avoid redundancy and maintain clear structure

Example

Assume the user just completed a task of "Implementing image understanding using OpenRouter API", calling this skill may generate: