Clipping Agent Stack

Scrubbing a two-hour podcast for the good 40 seconds, transcribing it, cutting it to 9:16, and captioning it takes two to three hours per video. Most people quit before the second one.

One command turns a long video into ready-to-post vertical clips: download with yt-dlp, transcribe with Whisper locally, merge the transcript into 15–60 second windows, score each window for viral potential, cut 1080x1920 clips with ffmpeg, caption, review, post.

Scoring runs free and local by default. Set an OpenRouter or OpenAI key and an LLM refines the picks with per-clip reasoning for about a cent a video. Both paths are fully implemented — neither is a stub.

One honest note about captions: plenty of Homebrew ffmpeg builds ship without libfreetype and libass, which means no drawtext and no subtitles filter. The pipeline probes your build at runtime — if the filters are there you get burned-in captions, and if they are not it says so plainly and writes an SRT sidecar instead. An SRT is written every run either way, and every upload platform accepts one.

• The whole pipeline is one 418-line file — stdlib plus ffmpeg, yt-dlp, and Whisper
• $0 per video with the heuristic scorer; pennies with LLM refinement
• Runs entirely on your machine — nothing uploads anywhere
• 10–20 minutes of automated work per video, plus about five minutes of human review