ootto-watch 👀 — let Claude actually watch a video

A Claude skill that gives Claude eyes + ears on any online video. Most tools only read the transcript and miss everything visual — this hands Claude the frames and the words, aligned on one timeline, so it can reason about cuts, on-screen text, demos, hooks, and pacing.

Paste a link → it downloads the video, samples key frames, transcribes the audio, and Claude reads both — then gives you a full breakdown: hook, format, storytelling structure, pacing and visual layout. Pass several links to batch-analyze dozens at once and surface the patterns the best ones share.

Install

Claude Code (drop the skill in):

git clone https://github.com/Ootto-AI/ootto-watch ~/.claude/skills/watch-video

claude.ai (web): zip this folder and add it under Settings → Capabilities → Skills.

Then just tell Claude: "watch this video: ".

How it works

1. Download — yt-dlp pulls the video (+ native captions if present) from YouTube, Instagram, X, Vimeo, and 1000+ sites.
2. Frames — ffmpeg samples key frames, duration-aware (a 30s clip gets ~30 frames; a 30-min lecture stays sparse) and scales them down for fast vision.
3. Transcript — native captions when available (free); otherwise local faster-whisper (also free, CPU). --no-whisper to skip.
4. Watch — Claude reads the frames as images alongside the timestamped transcript (index.json) and answers grounded in both.

Requirements

• yt-dlp and ffmpeg/ffprobe on PATH
• (optional) pip install faster-whisper — only needed for videos without captions

Analyze & batch

• Analyze — after watching, Claude writes a structured breakdown (hook · format · structure · pacing · visual layout), every point grounded in a frame timestamp + transcript line. See SKILL.md → Analyze.
• Batch — pass several URLs to watch them all, then compare across them for shared hook shapes, beat lengths and visual moves — the patterns winners reuse. See SKILL.md → Batch.

Run it directly

python scripts/watch.py "https://youtu.be/…" --out watch_out --max-frames 60
# writes watch_out/frames/*.jpg, transcript.txt, index.json

# batch — many at once:
python scripts/watch.py "<url1>" "<url2>" "<url3>" --out batch_out
# writes batch_out/v01…/, batch_out/batch.json

License

MIT © Ootto-AI. Built by @jayantcreates.ai. This is an original implementation; the "give Claude video input" idea is a community pattern with several great open-source takes (e.g. bradautomates/claude-video, devinilabs/claude-watch) — go star those too.