Cutting Rhythm

Cutting rhythm — why an accurate edit still feels wrong

Frame-accurate cuts in the right places still watch badly if the timing is wrong. Three levers do almost all the work, and they map onto three concrete things you can set:

Run edl --report on any cut list and it will tell you which of the three you're missing.

1. Split edits (J/L cuts) — the single biggest flow win

A cut where picture and sound change on the same frame is a butt cut. It announces itself. Real edits stagger the two so the viewer's ear carries them across the visual change and the cut disappears.

• J-cut — the next clip's sound arrives before its picture. You hear the answer starting under the end of the question. Use on almost every dialogue/interview cut.
• L-cut — the current clip's sound continues over the next picture. You keep hearing the speaker while you cut to what they're describing. Use for cutaways, B-roll, reactions.

One field expresses both — positive leads, negative lags:

{"src": "b.mp4", "start": 12.0, "end": 20.0, "audio_lead":  0.4}   // J-cut into this clip
{"src": "c.mp4", "start": 30.0, "end": 38.0, "audio_lead": -0.5}   // L-cut out of the previous

Typical amounts: 0.2–0.5 s for conversational cuts, 0.5–1.0 s for a documentary cutaway, up to ~2 s when laying B-roll under continuing narration. Below ~0.15 s it's inaudible — you've paid for nothing.

Where to use it: cuts within a continuous thought. Do not split a cut that's meant to be a hard break — a scene change, a chapter boundary, or the button at the end of a trailer wants the slam.

The renderer keeps video and audio on independent concat chains, so a lead never desyncs anything downstream — the leads telescope back to matching totals.

2. Where the cut actually goes

• Cut in the silence, not on the word. Get the what from transcribe --words and the where from speech-segments; the silence boundary is the truth (whisper word times jitter ±0.3–0.7 s). snap --to silence does this for a whole EDL at once.
• Leave the breath. The intake before a sentence belongs to that sentence. Cutting it off is the classic "why does this sound rushed and airless" mistake.
• Cut on motion, not around it. A cut placed while a subject is moving (a gesture mid-swing, a head turn, a step) hides itself; a cut between two still moments exposes itself. Never cut on the peak of a gesture — cut just after it starts or just before it lands.
• Match the outgoing and incoming energy. Cutting from a wide static shot to a fast handheld one reads as a mistake unless the audio motivates it.

3. Pause grammar — natural, not dead

"Tighten this up" almost never means cut content; it means shorten the gaps.

uv run video-agent tighten talk.mp4 -o edit.json --target-gap 0.5 --min-gap 1.0
uv run video-agent edl edit.json -o out.mp4

tighten removes time from the middle of each long pause, so the breath at the end of one sentence and the intake before the next both survive. Guidance for --target-gap:

Do not tighten uniformly. A pause after a punchline or a reveal is doing work — it's the beat the viewer laughs or thinks in. Widen those back out by hand in the EDL after running tighten. Uniform pacing is exactly what makes an edit sound machine-made.

This is distinct from filler-removal: that cuts spoken um/uh; this leaves every word and shortens the space between them. Run filler-removal first, then tighten.

4. Shot length — the snap

• Vary the holds. If edl --report says variation is under ~15%, the cut list is monotone no matter how good the picks are. Aim for a mix of short and long.
• Accelerate toward the climax. Trailers/montages want shot lengths that shorten as they build — e.g. 3.0, 2.6, 2.0, 1.5, 1.0, then the payoff.
• Hold the last shot. Endings want air. The final shot should be at or above the mean, never the tightest cut in the piece.
• Hold after a hit. A punchline, reveal or big visual needs a beat before the next cut, or it doesn't land.

Trailer structure that works: hook (≤2 s, in the first two seconds) → build (lengthening context shots) → turn (the problem/conflict) → climax (fastest cutting) → button (one held shot + title).

5. Cutting on the beat

uv run video-agent beats inputs/music.mp3                       # bpm + beat grid
uv run video-agent snap edit.json -o snapped.json \
    --to beats --ref inputs/music.mp3 --tolerance 0.4

snap --to beats moves each cut so it lands on a beat in the output timeline (cumulative running time) — which is what "cut on the beat" means; snapping source timestamps to beats is meaningless. Cuts with no beat inside --tolerance are left alone.

If the music bed starts partway into the track, pass its in-point as --offset (the same value as the EDL's music.start), or the grid is shifted.

Tempo is reported with a prior around 120 BPM to avoid the classic half-tempo lock. If the reported BPM looks like half or double what you expect, the grid is still musically valid (it's a metrical level) — just coarser or finer than intended.

6. The review loop

uv run video-agent edl edit.json -o /dev/null --report --dry-run   # pacing, no render
uv run video-agent edl edit.json -o draft.mp4 --draft              # fast 480p check
uv run video-agent edl edit.json -o final.mp4                      # deliverable

--report prints every shot length as a bar plus warnings for the three failure modes above. Read it before rendering — it's free, and it catches monotony that's invisible in JSON. --draft is for your own verification only; per the no-partial-previews rule the user sees one finished file.

Gotchas

• audio_lead needs material to reach into. A J-cut on a clip that starts at 0.2 s in its source has only 0.2 s to borrow — the renderer errors rather than silently shortening.
• No audio_lead on the first clip — nothing precedes it. It's ignored with a warning.
• A cutaway (vsrc) must be the same length as the audio it covers, or picture and sound come apart for the rest of the edit. The renderer now rejects a mismatch.
• Cut points are quantized to the frame grid by the renderer, so times produced by snap/tighten (which are not round numbers) can't drift picture against sound.
• Order: filler-removal → tighten → structure/split edits → beats snap → everything else. Tightening after you've set split edits invalidates the leads.