← back to guides

Claude watches YouTube for you 📺

From @techyai_cj. I stopped sitting through tutorials. This is a free skill called watch. You paste a video link, you ask what you actually want to know, and it comes back with the answer and the timestamps. Two commands to install and it works on any plan. I ran it on a four hour course to write this page, and the real result is further down.

Install it · two commands

paste these into Claude Code, one after the other

command 1 of 2

/plugin marketplace add bradautomates/claude-video

command 2 of 2

/plugin install watch@claude-video

That is the whole install. It is free and open source, MIT licensed. Not in Claude Code? You can add it to Cursor, Codex, Copilot and Gemini CLI with npx skills add bradautomates/claude-video -g, and on claude.ai you download watch.skill from the releases page and add it under Settings, then Capabilities, then Skills.

Use it

link first, then your question, in plain English

the shape of it

/watch https://youtu.be/VIDEO_ID what is actually worth knowing in here?

The question is the part people skip and it is the part that matters. Do not ask it to summarise. Ask it the thing you actually wanted to know before you clicked the video: what changed, what is the setting, does this work on Windows, is any of this new.

What you actually get back

a real run, not an example I made up

I pointed it at a four hour twenty seven minute Claude Code course on freeCodeCamp and asked for the things that actually matter with timestamps. This came back:

2:46:15

The code audit workflow

Turn Claude into a repeatable reviewer. Broad architecture pass first, then drill down. Score the findings, demand drop in fixes, write the report to a file, then paste it back and tell it to fix its own score.

1:32:12

session-notes.md, then start fresh

Long sessions rot. Have it write a session-notes.md with the decisions and the current state, then open a brand new session and say read that file and continue.

2:20:50

The CADE prompt formula

Context, Action, Details, Examples. Not "make it better" but "refactor this function to async await". A minute on the prompt saves ten on revisions.

1:22:12

permissions.deny for your secrets

Put your secrets folder, your .env and your build directories in permissions.deny, so Claude can see the whole app but never reads the sensitive files.

Four things out of four and a half hours, each one with a timestamp I can jump straight to. That is the whole pitch. I never opened the video.

Only watch part of it

when you already know roughly where the bit is

a section only

/watch https://youtu.be/VIDEO_ID --start 2:15 --end 2:45 what is he doing here?

This reads that window more closely and costs less, so it is the one to use when you want detail rather than a sweep.

What it actually reads

worth understanding, because it changes what you can ask

Captions. If the video has them, it pulls the transcript in a few seconds without downloading anything. Most public videos are covered here and it costs you nothing.

Frames. It also pulls still images out of the video and looks at them, so it can answer about what is on the screen and not just what was said. That is the difference between this and pasting a transcript somewhere.

It picks sensibly on its own. On the four hour course above it answered from captions alone, because a talking instructor does not need frames. Ask about something visual and it goes and gets them.

If it asks for something on first run

two things, both quick

1. It needs ffmpeg and yt-dlp. On a Mac it installs them for you. On Windows and Linux it prints the exact command to paste, so you are never guessing.

2. A Whisper key, but only for videos with no captions at all. Set a Groq key, which is the cheap one, or an OpenAI key. If the videos you watch have captions you can ignore this completely.

Ways I actually use it

A launch video or a keynote. Ask what is new and skip the hype, and ten minutes turns into four bullet points.

A long tutorial. Ask what is worth knowing if you already use the tool daily, and it tells you which parts to skip.

Your own screen recordings. Point it at a bug you recorded and ask when the interface breaks.

A video someone sent you that you are not going to watch. Ask if there is anything in it for you. Usually there is not, and now you know in twenty seconds.

It takes YouTube, Loom, TikTok, X, Instagram and Vimeo links, plus local files off your own machine.

I'm a software developer and I post what I actually run.
Follow @techyai_cj on TikTok and Instagram 🚀
More free guides: see all of them