Cut cold-open trailers from long videos with a human approval step

Oluyale wanted the short teaser that plays before a podcast or YouTube intro without scrubbing an hour of footage. He had Claude Code design the agent first, using his own design skill to write a spec and draw the architecture on Miro, then build it to that spec. The resulting agent downloads a video with yt-dlp, transcribes it with Whisper into numbered sentences, and makes one structured Claude call that proposes moments by sentence ID with scores for hook, emotion and surprise. Code then checks every proposal, orders the clips for rising tension and stops for the person to approve the exact list. Only then does ffmpeg cut, caption and render 16:9 and 9:16 versions. He ran it on episodes of his own course from the command line and a Streamlit page.

On a 7:35 episode the agent found 12 verified moments from one Claude call in 68 seconds for about $0.105. He swapped one clip before approving and the 5-clip cut rendered at 61.8 seconds in both formats. In one run Claude proposed a 24-second clip despite a 20-second limit; the verifier rejected it. A five-case test set first showed cuts of only one or two long clips, fixed by passing the target length into the prompt. A plain build-me-an-agent prompt, tried as a comparison, ran out of output tokens before the code was finished.

Have the model return sentence IDs, not timestamps, so code owns the numbers. The approval is tied to a hash of the ordered clip list, so editing it after approval blocks the render. Screen recordings with no face fall back to a blurred letterbox for vertical. The post also promotes the author's free design skill and course.