FREE TOOL
Free, no emailLet your AI watch a video.
Your AI can read a page and run a script, but it cannot watch a video. Paste a YouTube link and it guesses from the title, or it reads a transcript that misses everything happening on screen.
This hands it both tracks at once. It pulls a transcript with timestamps out of the audio, takes a picture every time the picture changes, and passes the lot back. Ask what happened at the 30 second mark and you get an answer from what is on screen.
That gap matters most when the visual is the whole point. A hook is three seconds of framing and text that no transcript captures. An ad is a layout. A tutorial is a screen.
You need an AI that runs on your computer, and about ten minutes to install it.
Download it.
One folder, 40 KB, and the licence is inside it.
Download the skillOn Claude Code you can skip the download and install it in one line instead.
The one line install.
/plugin marketplace add seanng23/claude-observe /plugin install observe@claude-observe
HOW TO INSTALL IT
How to install it.
Unzip it.
You end up with a folder called observe holding one instruction file and the scripts it runs.
Move that folder into your skills folder.
On Claude Code that is ~/.claude/skills/observe. Codex, Cursor and the rest read their own skills folder, and each one names it in its own setup notes.
Ask it to watch something.
Type /observe, the link and your question. The first run installs the three helpers it needs and tells you what it is doing.
WHAT IT DOES
What it does.
- Reads the words first, free where it can.
- Captions are pulled before anything downloads. A video with no captions falls back to a paid transcription service, and you decide whether to add a key for that.
- Takes pictures where the picture changes.
- Frames come out on scene changes rather than on a timer, and near-identical ones are thrown away so the budget goes on distinct content.
- A dial for how much detail you want.
- Four settings, from words only through to every frame it can find. Set it once and it stays set.
- Pin an exact second.
- Presenters say look at this over a shot that barely moves, so scene detection skips it. Name the timestamp and it grabs that frame.
- Works on files on your disk.
- Point it at an .mp4 you already have. A subtitle file sitting next to it gets used as is, with nothing sent anywhere.
- Remembers what it already did.
- Run it again on the same video and the transcript comes back free.
BEFORE YOU START
What you need.
- An AI that runs on your computer.
- Claude Code, Codex, Cursor or anything else that reads a skills folder. This one does not work in a browser tab, because it has to download and open files on your machine.
- Python, plus two small programs that handle video and audio.
- The setup installs all of it for you on a Mac. On Windows and Linux it prints the exact commands to run.
- Nothing to configure to start.
- Captions cover most public videos, so there is nothing to pay for and no key to add until you hit a video that has none.
Built on claude-video by bradautomates, MIT licensed. The original wrote the download, frame and transcript pipeline. This version adds the detail dial, frame deduplication, forced timestamps, caption language priority, subtitle pickup and transcript caching. The source is on GitHub.
LEVEL THREE
This is a level three skill.
Level three is a job written down once and run forever. This one is written in code rather than English, but it does the same thing: you stop explaining the job and start asking for the result.
See what we doThe Skill Prompt writes you one of these in your own words, for a job only your business does.
