Anything to AI Skill Generator

Transform YouTube videos and websites into AI agent skills that actually work.

The Problem

The web contains millions of programming tutorials, but this knowledge is:

1. Inaccessible to AI assistants — AI agents cannot watch videos or efficiently parse long pages
2. Time-consuming to extract — A 10-minute video takes 10 minutes to watch; long docs take ages to read
3. Often outdated — APIs change, libraries deprecate, best practices evolve
4. Not actionable — Narrative format ("so what we're gonna do...") ≠ executable instructions

The Solution

This Actor uses a Claude agent with the extract-text skill to:

1. Extract content from YouTube videos, playlists, or websites
2. Identify technologies, frameworks, and libraries mentioned
3. Verify against current documentation using RAG (Context7)
4. Transform content into actionable instructions
5. Generate a ready-to-use SKILL.md file

Input

Field	Type	Required	Description
urls	array	Yes	YouTube video/playlist URLs, website URLs, or any mix
skillName	string	No	Name for the skill (auto-generated if not provided)
prompt	string	No	Custom instructions (e.g., "Focus on deployment", "Beginner-friendly")
anthropicApiKey	string	Yes	Your Anthropic API key
context7ApiKey	string	Yes	Your Context7 API key for RAG verification
verifyTechnologies	boolean	No	Check technologies against current docs (default: true)
proxyConfiguration	object	No	Proxy settings for URL access
maxRetries	integer	No	Max validation retries (default: 2, max: 5)

Output

The Actor outputs:

1. Dataset: JSON with skill metadata and content
2. Key-Value Store: The generated SKILL.md file (and a .zip package for multi-file skills)

Dataset Schema

{
  "skill_name": "generated-skill-name",
  "skill_content": "---\nname: ...\ndescription: ...\n---\n# ...",
  "source_urls": ["https://youtube.com/...", "https://example.com/..."],
  "tools_used": ["extract_text", "resolve_library", "get_library_docs"],
  "tools_used_history": ["extract_text", "resolve_library", "get_library_docs"],
  "verify_technologies": true,
  "validation": {
    "is_valid": true,
    "warnings": [],
    "errors": []
  }
}

Example Usage

Single Video

{
  "urls": ["https://www.youtube.com/watch?v=VIDEO_ID"],
  "anthropicApiKey": "sk-ant-...",
  "context7ApiKey": "c7-..."
}

Playlist with Custom Instructions

{
  "urls": ["https://www.youtube.com/playlist?list=PLAYLIST_ID"],
  "skillName": "fastapi-deployment",
  "prompt": "Focus on production deployment steps only. Target experienced Python developers.",
  "anthropicApiKey": "sk-ant-...",
  "context7ApiKey": "c7-...",
  "verifyTechnologies": true
}

Multiple Sources (YouTube + Website)

{
  "urls": [
    "https://www.youtube.com/watch?v=video1",
    "https://www.youtube.com/watch?v=video2",
    "https://docs.example.com/guide"
  ],
  "skillName": "complete-react-guide",
  "prompt": "Merge into one cohesive skill covering all topics",
  "anthropicApiKey": "sk-ant-...",
  "context7ApiKey": "c7-..."
}

How It Works

┌─────────────┐     ┌─────────────┐     ┌─────────────┐     ┌─────────────┐     ┌─────────────┐
│   EXTRACT   │────▶│  IDENTIFY   │────▶│   VERIFY    │────▶│  TRANSFORM  │────▶│  GENERATE   │
│   content   │     │technologies │     │  up-to-date │     │   content   │     │    skill    │
└─────────────┘     └─────────────┘     └─────────────┘     └─────────────┘     └─────────────┘

Skill Structure Decision

The agent automatically decides whether to generate a single file or multi-file skill based on content complexity:

┌─────────────────────────────────────┐
│          Analyze Content            │
└──────────────────┬──────────────────┘
                   │
         ┌─────────┴─────────┐
         ▼                   ▼
   Short/Simple?        Complex/Long?
         │                   │
         ▼                   ▼
   Single SKILL.md    Analyze what to split:
                      ├── Theory → references/
                      ├── Code → scripts/
                      └── Variants → examples/

Decision Criteria

Criterion	Single File	Multi-File
Content length	Short / straightforward	Long / multiple topics
Code examples	1-2 small snippets	Complete runnable scripts
Theory vs practice	Mostly practical	Heavy theory + practice
Estimated word count	< 2000 words	> 3000 words
Multiple workflows	No	Yes (different use cases)

Multi-File Structure (Progressive Disclosure)

Following Anthropic's skill best practices:

skill-name/
├── SKILL.md              # Core instructions (<5k words, always loaded)
├── references/           # Deep docs (loaded on-demand)
│   ├── concepts.md       # Theory, explanations
│   └── api.md            # API documentation
├── scripts/              # Executable code
│   └── example.py        # Complete runnable examples
│   └── requirements.txt  # Required when Python scripts are included
└── assets/               # Templates, images (used in output)

Why this matters:

• SKILL.md is always loaded (~tokens cost)
• references/ only loaded when Claude needs them (saves tokens)
• scripts/ can be executed without loading into context
• assets/ copied to output, never loaded

Technology Verification

The agent uses Context7 RAG to verify that technologies mentioned in the content are still current:

• Checks library versions against latest documentation
• Identifies deprecated APIs or changed patterns
• Updates code examples to current best practices

Content Transformation

Video narration is transformed into actionable instructions:

Video Content	Skill Content
"So basically what we're doing..."	(removed)
"You want to run this command..."	bash command
"The important thing here is..."	Prerequisites section
"If you get this error..."	Troubleshooting section

Requirements

• Anthropic API Key: Get one at console.anthropic.com
• Context7 API Key: Get one at context7.com

Limitations

• YouTube videos without subtitles fall back to the video description
• English-focused (translation may reduce accuracy)
• Context7 coverage varies by library

Validation and Retries

If the generated skill fails validation, the Actor retries with the validation errors included in the prompt (up to 2 retries). This helps the model self-correct and produce a usable skill package. tools_used reports only the final attempt, while tools_used_history includes tools from all attempts.

License

MIT