cookcut

The problem

Cooking content lives on short-form video, and editing it is the bottleneck: forty minutes of phone clips have to become thirty seconds that open on the finished dish. I wanted a tool that does the first cut for me — locally, with no uploads.

What I built

cookcut: a local web app and CLI that turns raw cooking footage into platform-ready video.

• Backend: FastAPI + ffmpeg. Ingest is hash-deduplicated and streamed, so it never loads a whole video into memory. Scene detection and audio-event detection pick the moments; faster-whisper transcribes speech for captions.
• The edit: hook-first — open on the finished dish — then pace the shots, cap how much of the runtime any single source clip can take, and render 9:16, 1:1, 4:5 or 16:9.
• Review: review.py scores every cut on hook, pace, runtime and coverage, so "is this cut any good" is a test, not a feeling.
• Frontend: a Next.js editor to nudge cuts. CLI: one-click, batch and watch-folder modes.

The hard part: real footage

Everything passed on synthetic test clips. The first real iPhone run found seven bugs. Three favourites:

• Double rotation. ffmpeg autorotates portrait clips from the display matrix; the code was also applying a transpose. Every portrait clip came out sideways. Fix: -noautorotate on any input that feeds a rotation filter.
• Audio events found nothing. The threshold was "noise floor + 12 dB", but real kitchen audio only spans 3–18 dB. Now the threshold is relative to each clip's own dynamic range.
• True peak over the ceiling. AAC reconstructs peaks 0.6–1.2 dB above the PCM it's given, so loudness normalisation now aims lower and verification measures the encoded file.

Outcome

• 131 seconds of real footage → a 31-second vertical cut, 21 shots, opening on the dish, passing every structural check.
• 18 tests, now with real-footage fixtures where synthetic ones lied.