clipcut

Captioned vertical shorts from a raw take. Offline. Nothing renders until you say so.

pip install git+https://github.com/dancolta/clipcut
clipcut takes/ --speed 1.15

Not a clip finder. Not a caption burner. Not a cloud editor. It never rewrites what you said and never cuts for pace — only silence.

---

The point: it shows you the edit before it makes it

=== EDIT PLAN ===

  cold-email-quotes-reviews   (IMG_4451.MOV)
     length  105.4s -> 96.6s (-8.8s, 8%)
     speed   1.15x -> 84.0s final
     cuts    9   captions 52
     bleep   1 word(s) censored + beeped
     CHECK   1 possible mis-transcription(s):
               41.2s  'shadowing'  (conf 0.44)  -> did you mean 'scheduling'?
     gates   all pass

re-run with --yes to render

That CHECK line is the whole product. An automatic editor that mishears can't as can will burn the opposite of your meaning into 90 seconds of video, and you find out after publishing. clipcut flags it and stops.

Nine gates run before anything renders. The load-bearing one: a cut is a range of word indices, never a timestamp — so a cut landing inside a word isn't a bug to fix, it's a state that cannot be constructed.

What it does

Transcribes on-device (faster-whisper) → trims silence → burns word-by-word captions with the spoken word in an accent colour → masks and bleeps strong profanity → optionally retimes speed and ducks a music bed under the voice → one MP4 at −14 LUFS / −1 dBFS.

No account, no upload, no API key.

Install

You need ffmpeg built with libass. The plain Homebrew ffmpeg formula is not, and will fail at render time rather than at startup.

brew install ffmpeg-full        # macOS  (not `brew install ffmpeg`)
sudo apt install ffmpeg         # Debian/Ubuntu ships libass already
pip install git+https://github.com/dancolta/clipcut

First run pulls the whisper weights (~1.5 GB) into ~/.cache and never re-downloads. Captions default to Outfit (OFL) — install it or pass --font anything you have. No fonts or music are bundled.

Flags that change the output

Flag	Does
--speed 1.15	Retimes video, audio and captions together, inside the existing encode pass
--vocab "Grafana, Postgres"	Biggest single lever on accuracy for proper nouns. Worth more than a bigger model
--music bed.mp3	Your file, sidechain-ducked under the voice
--scripts notes/	Names clips from your written scripts and checks suspect words against them

--accent '#22D3EE', --font, --max-pause, --no-censor, --max-duration and the rest are in clipcut --help.

All nine gates — what each one blocks

All blocking and run before render, except where noted.

Gate	Blocks
no-mid-word-cut	a cut boundary landing inside a spoken word
no-orphan-fragment	a cut leaving half a clause behind
caption-cut-coherence	a caption showing a word the edit removed
caption-coverage-continuity	frames inside a line with no caption — reads as flicker
caption-sentence-integrity	one line carrying the end of a sentence and the start of the next
burned-text-fidelity	the ASS field-count bug that renders a phantom leading comma
protected-beats-intact	a phrase you marked uncuttable being cut
bleep-applied	a word masked in the caption but left audible
burned-text-recall	OCR of the rendered frames disagreeing with the script — advisory, does not block

python tests/test_gates.py

Six cases are adversarial and are supposed to fail. Weaken a gate and they stop failing, and the suite tells on you. PRs that make them pass will not be merged.

Config — clipcut.toml, env vars, precedence

built-in defaults  →  clipcut.toml  →  CLIPCUT_* env vars  →  CLI flags

clipcut walks up from the working directory looking for clipcut.toml, then falls back to ~/.config/clipcut/clipcut.toml. See clipcut.toml.example.

[clipcut]
out_dir  = "./shorts"
speed    = 1.15
accent   = "#A855F7"
margin_v = 400          # caption baseline, measured up from the bottom
vocab    = "Grafana, Postgres, Terraform"

margin_v is the one you will actually change — it is tuned to clear a centred talking head, and your framing is not mine. clipcut --print-config dumps what actually resolved.

Two findings from real use — audio clipping, and music licensing

Retiming after loudness normalisation clips. A 1.15x speed-up pushed a clean −0.8 dBFS master to +0.4 dBFS in testing. clipcut runs atempo before loudnorm so the signal being measured is the signal that ships.

Music on a Short over 60s is a licensing decision, not a taste one. A YouTube Short longer than 60 seconds carrying a Content ID claim is blocked globally, not merely demonetised. clipcut ships no music; --music takes a file you have cleared.

Licence

MIT. Built on faster-whisper and ffmpeg.