reel-machine

Raw clips in, scheduled reels out. You make the editorial calls (which moment, which words, which hook); the scripts do the cutting, drawing, rendering, checking and scheduling. A reel is one JSON spec, so every render is reproducible and a fix is a one-line change.

Done means: the render exists, finish.sh passed (voice EQ, jump cuts, pace gate, face check), avsync.py passed, the sensitivity pass and the second-model caption review both cleared, the owner approved anything on the show-first list, and the three posts exist in Buffer (drafts until the owner arms them).

Setup (once per machine)

Needs ffmpeg, ffprobe and Python 3. Any machine with those renders; no GPU. The pace gate and jump cuts use Silero VAD through faster-whisper (CPU, about 2 MB model, downloaded with the package); the face check uses OpenCV 4 (opencv-python-headless<5: OpenCV 5 dropped the Haar cascades from the main package).