Give Claude Eyes: A Free Skill That Lets It Watch Any Video

bash# macOS (Homebrew)
brew install ffmpeg yt-dlp
pip install -U openai-whisper

# Windows (winget)
winget install Gyan.FFmpeg
winget install yt-dlp.yt-dlp
pip install -U openai-whisper

bashmkdir -p ~/.claude/skills/watch-video

markdown---
name: watch-video
description: Lets Claude "watch" a video by pulling frames from the first 3 seconds and at every cut, plus a timestamped transcript, then breaking down the hook, cuts, on-screen text and call to action. Use when the user shares a video file or a public video link and asks what happens in it, why it worked, or how to reuse its structure.
---

# Watch a video

Claude Code reads images, not video files. Turn the video into frames plus a timestamped transcript, then read both together.

## Rules
- Only download public videos the user has the right to save, for private study. Never re-upload or redistribute them.
- If the user gives a local file, skip the download and copy it to video-notes/video.mp4.
- Describe only what is in the frames and the transcript. If something can't be seen in a frame, say so instead of guessing.
- Keep every file inside ./video-notes/ and run every command from inside that folder.

## Steps
1. Check the tools: `ffmpeg -version`, `yt-dlp --version`, `whisper --help`. If one is missing, give the user the install command (macOS: `brew install ffmpeg yt-dlp` and `pip install -U openai-whisper`) and stop.

2. Get the video and its length:
   mkdir -p video-notes/frames && cd video-notes
   yt-dlp -f "bv*[height<=720]+ba/b[height<=720]/b" --merge-output-format mp4 -o "video.%(ext)s" "VIDEO_URL"
   ffprobe -v error -show_entries format=duration -of default=nw=1:nk=1 video.mp4

3. Hook frames, two per second for the first 3 seconds:
   ffmpeg -hide_banner -loglevel error -y -t 3 -i video.mp4 -vf "fps=2,scale=720:-2" -q:v 3 frames/hook_%02d.jpg

4. One frame at every cut, with its timestamp:
   ffmpeg -hide_banner -y -i video.mp4 -vf "select='gt(scene,0.3)',showinfo,scale=720:-2" -fps_mode vfr -q:v 3 frames/cut_%04d.jpg 2> frames/cuts.log
   grep -o 'pts_time:[0-9.]*' frames/cuts.log | cut -d: -f2 > frames/cut_times.txt
   Line N of cut_times.txt is the time in seconds of cut_000N.jpg.
   If you get fewer than 3 cuts, also take one frame every 2 seconds:
   ffmpeg -hide_banner -loglevel error -y -i video.mp4 -vf "fps=1/2,scale=720:-2" -q:v 3 frames/every2s_%04d.jpg

5. Contact sheet of the first 16 cuts, for a fast overview:
   ffmpeg -hide_banner -loglevel error -y -pattern_type glob -i 'frames/cut_*.jpg' -vf "scale=320:-2,tile=4x4" -frames:v 1 frames/contact_sheet.jpg

6. Transcript with timestamps (skip if the video has no speech):
   ffmpeg -hide_banner -loglevel error -y -i video.mp4 -vn -ac 1 -ar 16000 audio.wav
   whisper audio.wav --model base --output_format srt --output_dir .
   This writes audio.srt.

7. Read every hook frame, the contact sheet and the cut frames with the Read tool. If there are more than 40 cut frames, read every second one. Then read audio.srt.

## Output
Write the breakdown to video-notes/breakdown.md and show it:
- **Hook (0 to 3s):** what is on screen in each hook frame, the first spoken line, and any on-screen text word for word.
- **Cuts:** how many, the average seconds between cuts, and a table of time, what changes on screen, and what is being said.
- **On-screen text and captions:** style, position, and how often they change.
- **Structure:** hook, setup, payoff and call to action, with timestamps.
- **Why it works:** 3 to 5 reasons, each tied to a specific frame or line. Label anything that is your interpretation.
- **Reuse this:** 3 ideas the user could apply to their own video without copying the creator's words or footage.

promptBreak down this video and tell me why the hook works: [video link]

Use watch-video on ~/Desktop/my-reel.mp4. How many cuts are there, and where does the pace slow down?

Watch these two videos and compare their first 3 seconds frame by frame: [link 1] [link 2]