I gave Claude a raw video file and told it: "trim it, add motion graphics." It did everything. No editing software, no manual prompts for each beat, no me clicking around. The intro you see at the top of the video was built end-to-end by the pipeline — and it's the kind of motion-graphics work that usually takes an editor an afternoon.
In this one I walk through the architecture behind it: how the trim layer works, why VAD on top of transcription beats either alone, and the part most people skip — the instruction layer that turns Claude from "popping text on the screen" into actually composing real motion graphics. I also show the moment where Claude built a beat I never asked for, just by reading what I said in the transcript. There's a small file-size detail in the takes folder that quietly tells the whole story too.
Drop your raw .mov in the repo, see what comes out the other side.