Cut Video

Cut Video

A [Claude Code](https://claude.com/claude-code) skill that tightens long-form recordings — removes silences, ums, fillers, and dead air **while preserving laughs and comedic pauses**.

**More accurate than cutting from Whisper alone.** Cuts are driven by [Montreal Forced Aligner](https://montreal-forced-aligner.readthedocs.io/) (MFA) word boundaries (~10–20ms precision, with true inter-word silences as explicit intervals) rather than Whisper timestamps (±100–300ms, with pauses embedded inside word durations). Whisper is used only to produce the transcript *text*; MFA aligns it to the audio. The result: tighter cuts, no clipped word onsets/tails, and reliable silence detection.

Point it at any video file and it will:

1. Make a proxy — transcodes HEVC/4K → 1080p H.264 with hardware decode so every subsequent step runs in seconds, not minutes (portrait 9:16 sources handled too).
2. Transcribe with Whisper — tiny.en, for the transcript text only.
3. Force-align with MFA — snaps the transcript to the audio at ~10–20ms precision; this drafts the take structure and the cut list.
4. Plan the cuts — applies a tone-aware silence heuristic (aggressive / balanced / sentimental / documentary), strips fillers, drops false starts and retakes. Shows you the plan before touching the video.
5. Ground-truth every keep boundary — re-transcribes each kept region in an isolated window before rendering. On retake-heavy footage Whisper collapses repeated lines and MFA's global alignment drifts by seconds; isolated windows catch that, plus dropped words and flubs hiding inside stretched words. Pauses >0.2s inside kept takes get trimmed to ~0.1s here too.
6. Render & verify — ffmpeg trim+concat (not select, which misaligns audio), then re-transcribes the output and scans for surviving stutters or clipped words before shipping.

There's also an optional **br