Video Cut

Turn one unedited long video into a compact, watchable first cut — decisive, content-aware, and self-verified. Human (you, the model) makes the editorial decision of WHAT to keep; the bundled scripts do the mechanical precision (boundary alignment, dead-air reclaim, memory-safe render, self-check).

When to use

• A raw/unedited video (YouTube URL or local file) needs to become a tighter cut.
• The ask is "halve it", "cut the boring parts", "make a first/rough cut".
• NOT for: word-by-word filler micro-surgery, multi-cam B-roll assembly, color/audio polish — those are a later P1/P2 polish pass. This skill is P0: validate the diagnosis + produce a solid cut.

Dependencies

/video-understand is a prerequisite. Run it first and use its validated media, transcript, analysis, and timeline instead of recreating them in this skill. Before starting, verify that it is installed. If it is not, warn the user that this prerequisite is missing and stop before processing media.

Activate any environment that has these (do NOT assume a specific conda env name):

• yt-dlp (download), ffmpeg/ffprobe (cut/render/probe) on PATH
• Python with faster-whisper (CPU works: device=cpu, compute_type=int8) Check first: yt-dlp --version, ffmpeg -version, python -c "import faster_whisper".

Project protocol workflow

Use these durable project files:

work/cut/edit-plan.json          # hand-authored keep/drop decisions only
work/timeline.json               # generated precision ranges + source/program mapping
review/01-cut/cut-summary.md
review/01-cut/timeline-map.png
review/01-cut/boundary-review.mp4
review/01-cut/full-proxy.mp4  # optional whole-program pacing review

Each canonical decision has a stable id, action (keep or drop), start_s, end_s, reason, and semantic/transcript evidence_refs. Keep decisions chronological; V1 does not support reordered or duplicated source clips.

Reuse the shared evidence layer rather than transcribing or diagnosing again:

python skills/video-understand/scripts/transcribe.py work/cache/audio16k.wav work/understand/transcript
python skills/video-understand/scripts/analyze.py work/understand/transcript.json work/understand/analysis.json

The existing precision scripts accept the canonical plan. Expand word-safe boundaries, optionally assign linear varispeed, then generate the shared timeline:

python skills/video-cut/scripts/build_edit.py work/cut/edit-plan.json work/understand/transcript.json work/cache/edit-final.json
python skills/video-cut/scripts/assign_speed.py work/cache/edit-final.json work/understand/transcript.json
python skills/video-understand/scripts/build_timeline.py work/cache/edit-final.json work/timeline.json --fps-num SOURCE_FPS_NUM --fps-den SOURCE_FPS_DEN

Use the exact rational FPS from work/understand/media.json; never round 30000/1001 to 30/1. Review timeline-map.png and the short boundary-review.mp4. Generate full-proxy.mp4 only when whole-program pacing must be reviewed. Write the result to cut-summary.md and point the operation check report at that file.

Record the timeline contribution with its required source input, update the operation integer revision, and set based_on to the exact understanding revision consumed:

"render": {
  "kind": "timeline-transform",
  "input": "../input/original-video.mp4",
  "plan": "timeline.json"
}

The shared delivery renderer encodes the active sequence after all revision checks pass.

Standalone compatibility workflow

The legacy files and commands below remain supported during migration. build_edit.py accepts both edit_coarse.json and canonical decisions[]; cut_render.py accepts both edit_final.json and timeline.json; this skill's transcribe.py is a compatibility wrapper around ../video-understand/scripts/transcribe.py.

Keep all intermediates under work/, deliver to project root:

work/source.mp4            # downloaded original
work/audio16k.wav          # 16k mono audio for whisper
work/transcript.json/.srt  # word-level timestamps
work/analysis.json         # diagnosis (silences/fillers/runs)
work/edit_coarse.json      # *** YOUR editorial decision (hand-written) ***
work/edit_final.json       # expanded tight render segments (generated; carries per-segment speed)
work/selfcheck/            # cut re-transcript + join frames
first_cut.mp4              # THE DELIVERABLE
first_cut_report.md        # edit + self-check report

The pipeline (6 steps)

1. Download

Highest accessible quality + audio. The default player client often fails (not-available / DRM / sign-in). The reliable fallback is the android client:

yt-dlp --extractor-args "youtube:player_client=android" \
       -o work/source.mp4 "<URL>"

If only a muxed 360p (format 18) is reachable without login, accept it and document the constraint — HD DASH is frequently DRM/login-gated. Then extract audio:

ffmpeg -y -i work/source.mp4 -ac 1 -ar 16000 work/audio16k.wav

2. Transcribe (word-level)

python scripts/transcribe.py work/audio16k.wav work/transcript [base.en]

Produces work/transcript.json (segments + per-word start/end) and .srt. faster-whisper, CPU/int8, VAD on, word_timestamps=True. Non-English: pass a multilingual model + --lang, e.g. python scripts/transcribe.py work/audio16k.wav work/transcript medium --lang zh (the .en models are English-only). Verify the language on a short sample first. Note: Whisper strips most vocalized "um/uh" — they are NOT transcript-measurable, so filler detection must combine the lexical word-search below with silence-gap analysis, not rely on the transcript text alone.

3. Diagnose ("see" + "read")

python scripts/analyze.py work/transcript.json work/analysis.json

Surfaces: leading/trailing dead air, inter-word silences ≥0.8s (with the word they follow), filler counts, wpm, speaking_ratio, near-duplicate repetitions, monologue runs >40s (ramble candidates).

To see the picture, sample a frame every ~2s into a contact sheet and read it yourself; confirm scene structure / hunt black or junk frames:

ffmpeg -y -i work/source.mp4 -vf "fps=1/2,scale=240:-1,tile=10x10" -frames:v 1 work/contact.png
ffmpeg -i work/source.mp4 -vf "blackdetect=d=0.1:pic_th=0.98" -an -f null -

For a single static talking-head shot there are no scene cuts — expect this, and note that the later jump-cuts are normal, not glitches.

4. Decide the cut — write edit_coarse.json BY HAND

This is the editorial core. Read the transcript + analysis, then write a coarse plan of keep-blocks and drop-spans that tile the entire timeline with no gaps/overlaps, each with in, out, and a one-line reason. Compress slow preamble to a few high-value nuggets; keep substance as large blocks; drop clear tangents/rambles. Preserve chronological order; the canonical V1 timeline rejects reordering. Aim to roughly halve runtime.

• Format spec + a real worked example: reference/edit_coarse.schema.md.
• Align every block boundary to a sentence/clause end — never start/stop mid-sentence or on a stammer. Use:

  python scripts/inspect_bounds.py work/edit_coarse.json work/transcript.json

  It accepts both legacy keep[] and canonical decisions[] plans and prints the words straddling each boundary so you can nudge them onto clean edges.

5. Expand → auto-speed → render

Turn coarse blocks into tight, dead-air-free render segments, normalize the speaking pace, then render:

python scripts/build_edit.py   work/edit_coarse.json work/transcript.json work/edit_final.json 1.5
python scripts/assign_speed.py work/edit_final.json   work/transcript.json   # adds per-segment speed
python scripts/cut_render.py    work/edit_final.json   work/source.mp4 first_cut.mp4

• build_edit.py splits each block at internal pauses > SPLIT_GAP (4th arg, e.g. 1.5s) and trims every boundary INTO the silence (LEAD 0.10s / TRAIL 0.30s) so no word is clipped and dead air is reclaimed. Lower SPLIT_GAP = tighter but choppier; for a slow, deliberate speaker prefer ~1.5, not 0.6 (avoids twitchy micro-cuts).
• assign_speed.py (auto varispeed) — groups render segments by decision_ref, measures the speaking pace of each hand-authored keep decision, and writes one shared speed factor plus each segment's resulting out_dur back into edit_final.json. Pace already inside the comfortable band stays at 1.0; out-of-band pace moves only to the nearest deadband edge. Adjacent decision blocks differ by at most 0.08x to avoid "pumpy" joins. It changes pace only — never which content is kept. Read the printed table before rendering:
  • Metric: words/min (English) — auto-switches to chars/min for CJK (zh/ja/ko).
  • Defaults are first-run starting points, not validated constants: deadband 145–185 WPM and speed clamped to 1.0–1.3. This is speed-up-only: slow decision blocks can be tightened, while natural-fast delivery remains at 1.0. Short decision blocks (<3.5s / <6 words) are too noisy to measure and inherit the global factor.
  • The rate is measured after dead-air reclaim, so it reads higher than gross WPM and is closer to articulation rate. The printed-table review is load-bearing: eyeball it and tune deadband to the speaker before rendering, don't trust the defaults blind.
  • Knobs (k=v): mode=segment|global|off (default segment; global=one factor for the whole video, off=disable), deadband=lo,hi, min_speed=/max_speed=, min_seg=/min_words=, lang=auto|en|cjk. To skip the feature entirely, just don't run this step — cut_render.py treats a missing speed as 1.0.
• cut_render.py uses N separate seeked inputs + concat filter, single re-encode. This is mandatory: trimming N times off ONE decoded input makes ffmpeg buffer the whole decoded stream (tens of GB for a long video). See reference/pitfalls.md. When segments carry a speed, it re-times each input (setpts video / atempo audio) and normalizes fps/timebase/sample-rate per input before concat so A/V stays locked; with no speed it falls back to the exact original graph.
• Final comparison: video-edit-compare reads the canonical timeline plus the actual final delivery. It projects final pixels back to source time, stretches varispeed clips to their source duration, and shows dropped ranges as black.

6. Self-check — do not declare done until this passes

# protocol precision regression check (run from the repository root)
python skills/video-cut/scripts/check_project_protocol.py
# re-transcribe the finished cut
ffmpeg -y -i first_cut.mp4 -ac 1 -ar 16000 work/selfcheck/cut_audio16k.wav
python scripts/transcribe.py work/selfcheck/cut_audio16k.wav work/selfcheck/cut_transcript
# content + filler + boundary check
python scripts/verify_cut.py work/edit_final.json work/transcript.json work/selfcheck/cut_transcript.json
# frame / black / join check (join times use post-speed out_dur, so they stay accurate)
python scripts/selfcheck_frames.py work/edit_final.json first_cut.mp4 work/selfcheck
# A/V sync + no residual dead air
ffprobe -v error -select_streams v:0 -show_entries stream=duration -of csv=p=0 first_cut.mp4
ffprobe -v error -select_streams a:0 -show_entries stream=duration -of csv=p=0 first_cut.mp4
ffmpeg -i first_cut.mp4 -af silencedetect=noise=-32dB:d=1.5 -f null -

Pass thresholds:

Honest filler caveat: paragraph-level cutting will NOT achieve "almost no fillers". Per-minute density can even rise because the retained substance is conversational. Report this honestly; word-level filler excision belongs to P1/P2 (and would make a single-shot talking head visibly choppy).

Finally

Write first_cut_report.md: source summary, what was cut & why, the self-check table, and any honest caveats (quality ceiling from 360p source, filler density, etc.). If varispeed was applied, disclose it: how many segments were re-timed and the speed range (from assign_speed.py's table / speed_params), so the reader knows the cut is not 1:1 real-time. Note too that verify_cut.py's filler-per-minute figure reads slightly higher after speed-up (same words, shorter minutes) — a reporting artifact, not added fillers.

Key principles (the hard-won ones)

• You decide content; scripts decide frames. Hand-write edit_coarse.json; let build_edit.py do precision and assign_speed.py do pace.
• Speed touches pace, never content. Auto-varispeed gives all render segments from one decision_ref the same speed, moves out-of-band pace only to the nearest deadband edge, limits adjacent decisions to a 0.08x delta, and clamps output to 1.0–1.3. atempo preserves pitch. Review the table; mode=off or skip the step to disable.
• Never clip a word. Boundaries always fall inside silence.
• Memory-safe render only (N seeked inputs + concat). Never multi-trim one input.
• Be decisive. Make reasonable calls and proceed; don't over-ask.
• Self-check is non-negotiable. Re-transcribe + frame-check before claiming done.
• See reference/pitfalls.md for the full list of traps and fixes.