Key Highlights
- Runs 24/7: 3 channels posting daily across YouTube, TikTok, Instagram and Facebook
- An LLM reads full transcripts and cuts complete thoughts under 60 seconds
- Argmax routing sends each clip to the one channel that wants it most — ~10x more usable clips for the same spend
- Fail-closed safety gate; schedules up to 14 days ahead
- 210 commits, 37 merged PRs, 30 test suites
Overview
Fablebench is a hands-off media pipeline running on a DigitalOcean droplet (n8n + Postgres). It watches source channels, reads every new upload's full transcript, picks a complete, self-contained thought under 60 seconds, and produces a finished vertical clip: reframe, captions, mood-matched music, per-channel branding.
The routing brain
Instead of one filter per channel, every clip is scored against every channel and routed by argmax to the single channel that wants it most. A database key keeps each clip on exactly one channel. The same spend now yields about 10x more usable clips.
Autopilot
A systemd timer runs the loop: pre-vet sources with an LLM, cut, edit, pass a fail-closed safety gate, write a caption, and schedule up to 14 days ahead across YouTube, TikTok, Instagram and Facebook. An operator dashboard lets Claude edit the montage timeline on request.