Video Review Skill

Claude Code can't watch videos. But it can look at hundreds of screenshots.

Extract keyframes from any video and map them to the transcript - giving Claude full visual context.

The Problem

You get a client video (Loom walkthrough, screen recording, meeting) but Claude can't watch it.

The Solution

1. •Extract 100-400+ keyframes using scene change detection
2. •Map each frame to the transcript timestamp
3. •Feed the screenshots + transcript to Claude

Now Claude understands the UI, workflows, and pain points.

Usage

# Local video file (client recording, Loom export, etc.)
python video_review.py ./client-walkthrough.mp4 ./output 0.1

# YouTube URL (downloads video + transcript)
python video_review.py "https://youtube.com/watch?v=VIDEO_ID" ./output

Scene Threshold

Output

output/
├── frames/           # 100-400+ keyframes
│   ├── scene_0001.jpg
│   └── ...
├── frames.json       # Timestamps for programmatic use
└── review.md         # Screenshots + transcript markdown

Python API

from video_review import VideoReviewer

reviewer = VideoReviewer(
    scene_threshold=0.1,  # 10% pixel change
    min_interval=1.0,     # Min 1 sec between frames
    max_interval=30.0     # Max 30 sec gap
)

result = reviewer.process("./video.mp4", "./output")
print(f"Extracted {result.frame_count} frames")

Requirements

• •ffmpeg: brew install ffmpeg
• •yt-dlp: pip install yt-dlp (only for YouTube URLs)

Triggers

• •"review this video"
• •"extract frames from video"
• •"analyze client walkthrough"
• •"process screen recording"
• •"give Claude video context"