What is claude-video?

claude-video adds a /watch command that gives Claude the ability to watch a video, not just read a transcript or guess from a title. You paste a video link or a local file and ask a question, and the tool downloads captions when they exist, pulls out frames from the video, transcribes the audio when there are no captions, and hands all of that, images and text together, to Claude so it can answer based on what actually happens on screen and in the audio. People use it to break down how a viral video or an ad opens, to work out what a competitor changed in a launch video, to figure out what is going wrong in a screen recording someone sent them without opening the file themselves, to summarize a long video quickly, to strip the hype out of a product announcement, or to turn a playlist of videos into a set of written notes. Under the hood, the tool first checks for existing captions, since those are free and fast, and only downloads and processes video or audio when needed. It extracts frames at different levels of detail depending on how much visual detail matters, and automatically limits how many frames it grabs based on the video length, so a short clip gets a dense set of frames and a long video gets a manageable sample instead of overloading Claude with images. A deduplication step drops frames that look almost identical to the last one kept, such as a slide that stays on screen for a long stretch, so the frame budget is spent on frames that actually differ from each other. It works with YouTube and many other video sites supported by the underlying download tool, as well as local video files. Installation is simple through Claude Code's plugin marketplace or a general skills installer, and required tools like yt-dlp and ffmpeg install automatically on first run on macOS, with instructions printed for other systems.