bradautomates/claude-video: Claude video ingestion pipeline for frames + transcription

I spent some time looking at bradautomates/claude-video (GitHub Trending). It’s a small, pragmatic pipeline that gives Claude a way to process arbitrary videos: a `/watch` entry point downloads the video, extracts frames, transcribes the audio, and then hands the derived artifacts to Claude for analysis.

This is useful in the exact spot where “multimodal” demos stop being actionable: when you have an engineering workflow that starts from a raw video source (recordings, walkthroughs, support calls, internal demos) and you want a repeatable way to turn it into structured understanding.

For example, if I’m debugging a reported UI issue from a recorded screen capture, I can run the pipeline to get frame samples plus a transcript, then ask Claude to summarize the sequence, identify failure moments, and pull out the relevant steps.

What I’d check first:

- The `/watch` command implementation (how it downloads, where it stores intermediate files) - How frames are extracted (sampling rate/strategy, resolution, any dedupe) - The transcription component and output format (timestamps, speaker info, etc.) - The integration point where frames + transcript are packaged and sent to Claude (prompt shape, context limits)

If you’re building production-ish RAG/agent tooling, this is the kind of “glue” repo that’s worth studying: it’s not trying to invent a new model—just making video input legible to Claude in a repeatable way.