video-tutorial-grabber

CLI tool that extracts code from video tutorials using AI vision analysis.

How it works

1. Frame extraction - Uses ffmpeg to capture frames at fixed intervals or via scene change detection
2. Deduplication - Removes near-duplicate frames using perceptual hashing (pHash) to reduce API costs
3. AI analysis - Sends frames to Claude's vision API which reconstructs the complete source code

Prerequisites

• Node.js >= 18
• ffmpeg installed and in PATH
• Anthropic API key

Installation

npm install
npm run build

Usage

# Set your API key
export ANTHROPIC_API_KEY=sk-ant-...

# Basic usage - extract frames every 5 seconds
npx vtg tutorial.mp4

# Custom interval (every 10 seconds)
npx vtg tutorial.mp4 --interval 10

# Use scene detection (captures frames only when screen changes)
npx vtg tutorial.mp4 --scene-detection

# Scene detection with custom sensitivity (lower = more frames)
npx vtg tutorial.mp4 --scene-detection --scene-threshold 0.2

# Save output to specific file
npx vtg tutorial.mp4 -o output/app.ts

# Keep extracted frames for inspection
npx vtg tutorial.mp4 --keep-frames

# Use a different Claude model
npx vtg tutorial.mp4 --model claude-sonnet-4-20250514

Options

Option	Description	Default
-i, --interval <seconds>	Seconds between frame captures	5
-s, --scene-detection	Use scene change detection instead of fixed interval	off
-t, --scene-threshold <0-1>	Scene detection sensitivity (lower = more sensitive)	0.3
-d, --dedup-threshold <n>	Deduplication strictness (lower = stricter)	5
--no-dedup	Disable frame deduplication
-o, --output <path>	Output file path	<video>_code.<ext>
-k, --api-key <key>	Anthropic API key	$ANTHROPIC_API_KEY
-m, --model <model>	Claude model to use	claude-sonnet-4-20250514
-b, --batch-size <n>	Max frames per API request	20
--keep-frames	Keep extracted frames after analysis	off
--frames-dir <path>	Directory for extracted frames	frames

Tips

• Scene detection is recommended for most tutorials - it captures meaningful changes and skips static moments
• For long videos (>30 min), increase --interval or use --scene-detection to reduce frame count
• If results are noisy, try a stricter dedup threshold (--dedup-threshold 3)
• For higher accuracy, use --model claude-sonnet-4-20250514 (costs more but better at reading code)