A Claude Code skill that downloads any video, extracts frames, transcribes audio, and hands everything to Claude — so it can answer questions about what it actually saw and heard, not what the title says.
This guide is reviewed for clarity, service accuracy, and AI-search readability. The next quarterly content review is tracked internally before unsupported metrics or client proof are added.
The gap Claude Code couldn't close
Claude Code can read a webpage, run a script, browse a repo, analyze an image. What it can't do natively is watch a video. Paste a YouTube link and it has to guess from the title, pull a transcript that misses half of what's on screen, or admit it can't help.
The /watch skill closes that gap. Paste a URL or a local file path, ask a question, and Claude downloads the video, extracts frames at an auto-scaled rate, pulls a timestamped transcript, and reads every frame as an image. When it answers, it has seen the video and heard the audio.
Installing the skill
bash# Claude Code (CLI) /plugin marketplace add bradautomates/claude-video /plugin install watch@claude-video # Codex / generic skills runtime git clone https://github.com/bradautomates/claude-video.git ~/.codex/skills/watch
For claude.ai (web): download watch.skill from the GitHub releases page and drop it into Settings → Capabilities → Skills.
How to use it
- Analyze a video:
/watch https://youtu.be/dQw4w9WgXcQ what happens at the 30 second mark? - Reverse-engineer content:
/watch https://youtu.be/abc123 what hook did they open with? - Debug from a recording:
/watch bug-repro.mov what's going wrong? - Summarize:
/watch https://youtu.be/xyz summarize this - Focused window:
/watch https://youtu.be/abc --start 0:45 --end 1:15 what changed on screen?
What the skill actually does
When you call /watch, Claude executes a five-step pipeline:
yt-dlpdownloads the video into a temp directory (local files are probed in place)ffmpegextracts frames at a duration-aware rate- Transcript comes from native captions via
yt-dlp(free), with Whisper API as fallback - Frames + transcript are handed to Claude, which reads each frame as an image in parallel
- Claude answers grounded in what's on screen and in the audio
Zero config on first run — yt-dlp and ffmpeg install automatically via brew on macOS. Whisper API key is only needed if a video has no captions.
The frame budget — why it matters for token cost
Every frame is an image token. The skill's auto-fps logic exists to avoid blowing your context budget on a long video:
- ≤30 seconds: ~30 frames — dense, essentially every key moment
- 30s–1 min: ~40 frames — still dense
- 1–3 min: ~60 frames — comfortable
- 3–10 min: ~80 frames — sparse but workable
- >10 min: 100 frames — use
--start/--endfor focused questions
What people actually use it for
- Content analysis — break down a competitor's hook structure, pacing, and CTA from the actual video, not the title
- Bug reproduction — watch a screen recording of something broken and identify the frame where the issue appears
- Video summarization — structure and key moments faster than 2x speed
- Ad creative analysis — paste a TikTok or Reel URL and ask what made the first 3 seconds work
Frequently asked questions
What video sources does the skill support?
Anything yt-dlp supports — YouTube, Loom, TikTok, X, Instagram, Vimeo, and hundreds of other platforms. Local .mp4, .mov, .mkv, and .webm files also work.
Do I need a Whisper API key? No, not for most videos. The skill first tries native captions via yt-dlp — free and instant. Whisper only kicks in for videos with no captions.
How do I get more detail on a specific moment in a long video?
Use --start and --end flags: /watch https://youtu.be/abc --start 2:30 --end 3:00 what's happening here? Focused mode gets a denser frame budget for the specified range.