Skip to content

/watch Skill — Setup Guide

Created 2026-05-09
Tags claude-codeskillssetup

The /watch skill lets Claude download a video, extract frames, pull a timestamped transcript, and answer questions grounded in the actual video content. Useful for research, reference videos, tutorials, or any time you want Claude to watch something and report on it.

Source: bradautomates/claude-video Install location: ~/.claude/skills/watch — global, available in every Claude Code project on the machine.


Requires two command-line tools. On a Mac with Homebrew, run once:

Terminal window
brew install ffmpeg yt-dlp
  • ffmpeg — extracts frames and audio from video
  • yt-dlp — downloads video and native captions from YouTube and most platforms

Run once per machine:

Terminal window
mkdir -p ~/.claude/skills
git clone https://github.com/bradautomates/claude-video ~/.claude/skills/watch

That’s it. The skill is immediately available as /watch in any Claude Code session on that machine.


Section titled “Groq API key (recommended — needed for local video files)”

YouTube videos have auto-generated captions and work without any API key. For local video files (.mp4, .mov, session recordings, etc.) the skill needs a transcription backend. Groq is free, fast, and uses whisper-large-v3 — the best Whisper model available.

Setup (takes 2 minutes):

  1. Go to console.groq.com/keys and sign up for a free account
  2. Generate a new API key
  3. Run these two commands in Terminal, replacing the key with yours:
Terminal window
mkdir -p ~/.config/watch
echo "GROQ_API_KEY=your_key_here" > ~/.config/watch/.env
chmod 600 ~/.config/watch/.env

That’s it. The skill detects the key automatically — no other configuration needed.

Note on security: Never share or commit the key. The .env file is outside any git repository and only readable by you (0600 permissions). If you ever need to rotate it, generate a new key at Groq and rerun the echo command above.


/watch <url>
/watch <url> your question here
/watch /path/to/local/video.mp4

Examples:

/watch https://www.youtube.com/watch?v=jbHB-rzKBAs
/watch https://youtu.be/abc123 what framework does he explain at the 5 minute mark?

For long videos (over 10 minutes), you can focus on a section:

/watch <url> --start 2:00 --end 5:00

  • Works with YouTube and most yt-dlp-supported platforms
  • Best accuracy on videos under 10 minutes (frame coverage gets sparse on longer videos)
  • Patrick has this installed on his Mac. Ask him if you want to see it in action before installing.