Cutfarmer is built as a pipeline: each stage produces something the next one uses, and every automatic decision is one you can override in the app. All of it runs on your Mac.
Drop in a file or paste a link. Cutfarmer builds an edit-friendly proxy, extracts audio, transcribes with word-level timestamps on-device, and separates who's speaking from whom (diarization). A two-hour podcast ingests in minutes on Apple Silicon.
The engine maps the episode: shot boundaries, every visible face clustered into identities, voices matched to faces, audience reactions (laughter, applause) tagged on the timeline. This is the ground truth every later decision stands on.
A reasoning pass reads the transcript like an editor: where does a bit start, where does the payoff land, what stands alone without context? Candidates are scored for hook strength and self-containment, then re-cut to exact sentence boundaries. You can also give it a brief — "one clip per comedian, full set" — and it follows it.
For each clip, Cutfarmer plans the 9:16 crop frame-by-frame: who is speaking, where they are, when to hold, when to follow, when to cut. Identity-aware tracking keeps the same person framed across camera angles; quality gates verify a face is actually in frame during speech before a clip can ship.
Word-timed captions render in your chosen style with per-word highlights, and a hook headline is burned into the top band — placed to never collide with the speaker, the watermark, or platform UI.
A native AVFoundation renderer produces the final 1080×1920 video — hardware-accelerated, color-managed, loudness-normalized. The preview in the app is the same pipeline at a smaller bitrate, so what you see is exactly what exports.
Finished clips post to your connected TikTok and YouTube accounts (Instagram rides the same pipeline), with per-platform titles, descriptions, and hashtags. Performance data flows back in, so the next batch is ranked by what your audience actually watched.
Every stage writes its decisions to an inspectable log, and the app surfaces the moments worth your judgment: a clip the quality gate wasn't sure about, a framing call on a crowded stage, a title that needs your voice. Override once and the system remembers — user corrections become training signal for the next run.