Claude can't watch videos
If you paste a YouTube link to Claude, it sees nothing. At best it guesses from the title. Watch closes that gap: it downloads the video, cuts it into frames, grabs the transcript with timestamps, and hands it all to Claude.
At that point Claude answers like someone who actually watched the video, not someone who read the cover.
It's useful for three things above all: understanding how a video is put together, getting twenty minutes summarized in thirty seconds, and — the case below — turning a tutorial into a written manual.
Install: two lines, just once
Open Claude Code and paste these two commands, one after the other.
/plugin marketplace add bradautomates/claude-video
/plugin install watch@claude-video The first tells Claude where to find the skill, the second installs it. That's it. You don't have to download anything by hand or touch any folders.
Using Claude from the website instead of the terminal? Download the watch.skill file from the
project's GitHub releases
and drag it into Settings → Capabilities → Skills.
The first run sorts itself out
The first time you run /watch, Claude checks whether you have two programs installed.
- yt-dlp
is the program that downloads the video from the link you give it. Without it, Claude has nothing to watch.
- ffmpeg
is what cuts the video into frames and, if needed, pulls out the audio. It's the video Swiss Army knife.
On Mac, if you don't have them, it installs them for you via Homebrew. You don't have to do anything: just wait.
On Windows and Linux it doesn't install them for you: it prints the exact commands on screen to copy and paste. Paste them, then run /watch again.
You'll never be asked to remove anything. If you see a message about libraries, it's asking to add them, not take them away.
When you need a key, and when you don't
This is the only real choice you have to make, and it's the part that confuses people most.
If the video already has subtitles — almost every public YouTube video does, even the automatic ones — Watch grabs them straight from there. Free, instant, no setup. In the vast majority of cases it ends here and there's nothing else to do.
If the video has no subtitles — typical of your own files, a screen recording, a downloaded video — Watch has no text to read. In that case it extracts just the audio and sends it to a transcription service, and that needs an API key. Two options:
- 1 Groq — recommended: faster and cheaper. Get the key at console.groq.com/keys.
- 2 OpenAI — the alternative, at platform.openai.com/api-keys.
Claude asks for it exactly when it needs it and saves it itself in a protected config file: you don't have to figure out where to put it.
If you'd rather not add it, you can still go ahead: Claude will watch only the images, without the audio. On a tutorial with a lot of talking, though, you'll miss half the content.
From YouTube tutorial to written manual
This is the reason it's worth installing. You find a good but long tutorial, and you want the written version, searchable, without watching it again every time. Paste this.
/watch https://youtu.be/... turn it into a step-by-step written manual, with timestamps next to each step Claude downloads the video, extracts the frames, grabs the subtitles, and then looks at the images. That's the point: in tutorials the important things are on screen, not in the audio.
The narrator says "click here," and here is visible only in the frame. With the transcript alone you'd get a manual full of "click here" with no idea where.
The result is a document with the steps in order, what was on screen at each point, and the timestamp to go back to the video if a step isn't clear. If the video is long, ask for one chunk at a time: "give me the manual for just the part between 4:00 and 9:00." It comes out much better.
Two limits worth knowing upfront
It works best under 10 minutes. It's not a rule, it's math: the frames Claude can look at are limited (100 max). On a 5-minute video it grabs one every few seconds and misses nothing. On a 40-minute one, those same 100 frames are spread thin, with twenty seconds between each where anything can happen. For long videos, always ask for a specific section.
It burns through context. A lot. Every frame is an image, and images weigh far more than text. A 5-minute video eats up a big chunk of the conversation. That's fine as long as you know it: ask the right question the first time instead of re-asking three times.
What leaves your computer
Worth being upfront about, since we're talking about your own videos. The video is never uploaded anywhere: downloading, cutting frames, it all happens on your machine. The only thing that leaves is the audio, and only if subtitles are missing and you've added a transcription key. If subtitles exist, no request goes out at all.