script-cut

script-cut — text-based editing with forced alignment

The idea (say it back to yourself before using it)

You edit in Premiere by looking at the audio track and the words. This automates exactly that:

1. Script — the exact clean words you want the clip to say.
2. Find — forced-align that script to the audio so you know where every word physically is.
3. Cut — keep the word-spans, cut everything in between.

"Everything in between" is silence AND junk (fillers, false starts, restarts). Junk is never in the script, so it lives in the gaps between kept words and gets cut automatically. You never surgically remove errors after the fact — you only keep the words you chose, so errors can't leak through.

This is text-based editing (how Descript works), done with a real forced aligner so it's frame-honest.

Why align the FULL transcript, not just the clean script

A forced aligner cannot skip unscripted audio — it absorbs a stumble into a neighbouring word's boundary, inflating that word's span so the stumble rides along in the cut. So: align the faithful transcript (every word, incl. fillers/restarts), then KEEP only the curated subset and cut the rest. The kept words land on their true positions; the junk stays in the cut gaps.

Run it

# 1) one-time: install the toolchain (MFA + a python venv with numpy + num2words)
bash ${CLAUDE_PLUGIN_ROOT}/skills/script-cut/scripts/setup_toolchain.sh

# 2) cut: produces <out>/cut_spec.json  (precise, in SOURCE seconds)
"$VENV/bin/python" ${CLAUDE_PLUGIN_ROOT}/skills/script-cut/scripts/script_cut.py \
    --source  SOURCE.mp4 \
    --transcript words.json \      # word-level transcript: {"segments":[{"words":[{word,start,end}]}]} etc
    --spec    structure.json \     # {"segments":[{"in":sec,"out":sec, "n":1?, "cat":"LABEL"?}]} = rough marks
    --out     OUTDIR \
    --gap     0.20                 # cut any inter-kept-word gap >= this; bigger = keep longer pauses

# 3) render cut_spec.json with whatever renderer you use, e.g. for a numbered listicle short:
python ${CLAUDE_PLUGIN_ROOT}/skills/listicle-short/scripts/build_short.py \
    --source SOURCE.mp4 --spec OUTDIR/cut_spec.json --music TRACK.mp3 \
    --res 1080 --grade --per-seg --eye-lock --out RENDERDIR/

# 4) EAR-TEST IT YOURSELF (never hand QC to the user):
"$VENV/bin/python" ${CLAUDE_PLUGIN_ROOT}/skills/script-cut/scripts/qc_ears.py  --clip RENDER.mp4 --spec OUTDIR/cut_spec.json
"$VENV/bin/python" ${CLAUDE_PLUGIN_ROOT}/skills/script-cut/scripts/qc_clips.py --clip RENDER.mp4 --spec OUTDIR/cut_spec.json \
    --transcript RENDER_words.json    # the render's OWN transcript = ground truth of what it actually says