Skillforge

Generate AI agent skills from documentation. When Claude fails a task, Skillforge auto-discovers docs, crawls them, and produces a reusable SKILL.md file.

Installation

Using pip (recommended)

pip install skillforge

From GitHub

pip install git+https://github.com/Ranoobaba/Skillforge.git

From source

git clone https://github.com/Ranoobaba/Skillforge.git
cd Skillforge
pip install -e .

Quick Start

1. Initialize (set up API keys)

skillforge init

This creates a .env file with your API keys:

• ANTHROPIC_API_KEY - Get one at https://console.anthropic.com
• FIRECRAWL_API_KEY - Get one at https://firecrawl.dev

2. Verify setup

skillforge doctor

3. Generate a skill

# Auto-discover documentation
skillforge run "use the Stripe API to create subscriptions"

# Or provide a seed URL
skillforge run "build CUDA kernels" --seed https://docs.nvidia.com/cuda

Commands

Command	Description
skillforge init	Set up API keys interactively
skillforge doctor	Verify configuration
skillforge run "task"	Generate a skill from a task description
skillforge list	List generated skills

Options for run

skillforge run "task" [OPTIONS]

Options:
  --seed URL           Seed documentation URL (auto-discovers if omitted)
  --model MODEL        Claude model to use (default: claude-sonnet-4-20250514)
  --max-attempts N     Max teacher retries (default: 10)
  --corpus-limit N     Max pages to crawl (default: 50)
  --stealth            Use stealth proxies for anti-bot sites (9x cost)
  --fast               Speed-optimized mode for demos
  --no-sandbox         Disable sandbox validation
  -v, --verbose        Verbose output

Examples

# Basic usage - auto-discovers Stripe docs
skillforge run "use the Stripe API to create subscriptions"

# With explicit seed URL
skillforge run "build CUDA kernels" --seed https://docs.nvidia.com/cuda

# Fast mode for demos (smaller corpus)
skillforge run "use Firecrawl to crawl webpages" --fast -v

# For sites with anti-bot protection
skillforge run "use the GitHub API" --stealth

Output

Generated skills are saved to skills/<skill-name>/:

skills/
└── use-the-stripe-api-to-create-subscriptions/
    ├── SKILL.md           # Main skill documentation
    ├── tests.json         # Verification test cases
    └── skill_manifest.json # Metadata

How It Works

1. Config Validation - Checks API keys
2. Source Discovery - Auto-discovers docs via Firecrawl search OR maps seed URL
3. Corpus Building - Crawls sources, saves as markdown
4. Teacher Session - Claude attempts the task; on failure, searches for missing info
5. Validation - Static analysis + sandbox checks
6. Skill Generation - Synthesizes SKILL.md from successful trace

Environment Variables

You can also set API keys via environment variables instead of .env:

export ANTHROPIC_API_KEY=sk-ant-...
export FIRECRAWL_API_KEY=fc-...

Requirements

• Python 3.12+
• Anthropic API key
• Firecrawl API key

License

MIT