dongnguyenvie/bashcut-agent-kit / skills/beat-cut
beat-cut
VerifiedBeat cutting and punch-in reframes
The loop: measure → choose → cut → measure again → adjust. Numbers are ranges with their source (T07 pacing,
T03 shot grammar, T08 effects: the kit's research notes, naming the editors' kits they came from). Starting
points, not targets: a user's reference, measured with bc:style-study, wins.
1. Measure what drives the cuts
| Video | Cuts follow | Measure |
|---|---|---|
| Montage, travel teaser, MV, music-led | beat grid, music energy | beats detect, beats grid, audio energy |
| Talking, review, food with speech, tutorial | phrase edges; picture follows words | transcript words --heard, speech rate |
Speech-led: do not lay cuts on the grid first (one example: a grid-first draft had 10 silent gaps, 21 s).
Chain the spoken lines, cut picture at phrase ends (the gap before a word in transcript words --heard), then put
music under it. Denser speech (speech rate) carries more per shot, so shots hold longer (T07 §3). Before a music
hit after a line, leave 0.1–0.6 s of air (bestagentkits ≥0.12 s, video-use 400–600 ms; T07 §3).
Long pauses in one talking clip: Silence Markers cuts them (bashcut plugins run bashcut.silence-markers.remove --params '{"detect":"auto","minSilenceMs":400,"paddingMs":100}' with the clip selected; mark only adds section
markers). An older version refuses detect: leave it out there (plugins actions silence). Look at the long spans
in data.removed (media strip --media ID --from S --to S) before keeping it: a long "pause" can be a reaction, a
laugh or b-roll sound. Shorter minSilenceMs for a fast talker, more paddingMs for a slow one. Then section 6.
2. Read the music
Place the music first, then beats detect (job; needs an audio.beats provider), beats grid, audio energy
(job). Judge the grid before cutting on it. You cannot listen: a weak grid means ask the user, or cut on phrases.
| Fact | Strong | Weak → do |
|---|---|---|
fit.rmsErrorMs | small next to a frame (33 ms at 30 fps); kits gate at mean <10 ms (shotcraft), median <15 ms (jianshuo) (T07 §3) | tempo drifts: detect per song or section, cut on phrases |
confidence | stands out from the alternates (no published threshold) | rubato, ambient, speech-heavy: energy lifts/drops and phrases |
alternates | one tempo clearly stronger | similar: pick by how fast the music moves, ask the user. One example: 58.7 BPM detected, truth 117 (the bar) |
downbeats, phaseScores | one phase clearly highest | close: a full-band grid can lock onto hi-hats, every cut half a beat late (hyperframes; T07 §4). Ask the user to check a downbeat hits (ui seek F) |
strengths | — | "the hit table is a candidate pool, not a trigger" (shotcraft; T07 §2) |
Several songs: detect each media.
Find moments in the energy curve yourself: a lift is levelDb or fullness rising several dB over ~2 s against the 2 s before, a drop the reverse, a breath a 0.5–2 s dip before a lift. Snap each to the nearest beat or downbeat. Pointers, not cut points: they shape a section, build → breath → hit on a downbeat → hold (reelmimic; T07 §2). One breath of 0.5–2 s before the main hit, a payoff hold of ≥1 s (T07 §7).
3. Choose shot lengths and cut points
Pick a band per section, then move inside it by content:
| Genre / section | Median shot | Source |
|---|---|---|
| Fast short-form, montage, ads, food inserts | 0.6–2 s | assafkip, iart, znyupup (T07 §3) |
| Speech-led vertical | 2–4 s | T03 §3 |
| Narrative, vlog, b-roll travel | 2.5–6 s (travel 3–4.5 s) | higgsfield, vox-director, vlog plugin (T07 §3, T03 §3) |
| Calm vs urgent tone | 2–4 s vs 0.4–1 s | saas-motion-kit (T07 §3) |
| Hero hold | one per film, 1.5–2.5× the average | higgsfield (T07 §3) |
- Shorter: high-motion, low-information shots, energy lifts and drops; longer: dense speech, text to read, landscapes, calm tone, after a breath (T07 §3).
- Cuts per minute seen: dialogue 14–16 (davinci, "descriptive, NOT a target"), daily vlog 8–12, review 10–14, travel 13–16 (vlog plugin), food TikTok 30–40, cinematic vlog 12–20 (style-study) (T07 §3). Compare, don't aim.
- Beats per shot 2–8, fewer at peaks (opuscar); cutting every beat for over ~8 beats reads as noise (T07 §7). Seconds = beats × 60 / BPM (4 beats: 2 s at 120, 2.7 s at 90): choose the count from the band, not a pattern.
- Vary the rate: one acceleration, one breath, a hero hold (T07 §4). Open with the hook for short form, establish first for narrative, real estate or a travel arrival (T03 §3); the plan says which.
One example (travel/food vlog, not a rule): cold open 6 beats (~2.6 s), busy sections 4 (2 on peaks, ~1.6–2.0 s),
quiet ending 4 on slow music (~3.5–4 s); a uniform 3.6 s cut felt flat whatever the content; pieces under ~0.6 s
were merged. The cold open was ~8 s of three shots (the ending, the loudest, the quietest), then chronological
with a section marker and a place label (bc:captions-text) at each location change.
4. Cut in one apply
Plan in beats and seconds, convert to frames last (shotcraft): beat frame = music item at + (beat second −
source-in second) × fps at speed 1 (beats grid also gives downbeatFrames). Per section: cut points → (clip,
in-point) from its moments in turn → one timeline apply, labelled per section. Spread in-points across sources;
never reuse one spot. roll nudges a cut onto a beat or word gap; slip changes what a clip shows.
5. Punch-in reframes from measured headroom
The most effective change when the footage has one angle. Read each clip's scale (timeline get, review shots).
- Headroom is
maxZoomNative(largest zoom before upscaling), per clip: ~2 for 4K in a 1080 landscape project, 1.0 for same-resolution footage (every punch-in upscales), ~1.125 for 4K landscape filling 1080×1920. - Step by purpose (T08 §3): size change at a cut 1.15–1.3, snap punch on a word 1.1–1.15, slow drift 1.03–1.07 over the shot. Aim for one described size step (wide → medium → close, T03 §7), within headroom where you can.
- Past headroom is a judgement, not a limit: look at
ui frame F(andui frames --compare source --items CLIP) first. One example: same-resolution footage looked acceptable at 1.28, soft beyond. - Pan and tilt need margin (none at zoom 1.0 with
frameCoverage1). Place the crop on what matters, seen in the frame, never a blind centre (bc:effects, Reframe). - Never punch in on signs or text (one example: 1.28 cut "Lẩu Bò Nồi Đất" to "Nồi Đấ").
- A patch replaces the whole
transform: send zoom, pan and tilt together.
Neighbouring cuts keep the same framing only by choice. One edit's rotation (zoom, pan, tilt): (1.00,0,0) (1.28,60,−25) (1.14,−50,15) (1.22,0,30) (1.00,0,0) (1.18,70,10) (1.26,−65,−20) (1.10,40,25). Talking heads: punch-ins every 4–6 s seen (vlog talking-head recipe; ghost-editor 4–6 per 60 s reel; T08 §3), timed to sentence starts, not a clock.
6. Measure again, then adjust
review measure (job), review shots --summary, review sync --bins, review sync --events impacts, and
review sync --events cuts,sfx --rendered after an export of this revision.
- Rhythm (overall, per section): median, cv, cuts per minute against your band. 6+ consecutive shots varying
under ~0.15 cv (davinci) or a
modeshare over ~60% (saas-motion-kit) is monotony: break it unless the calm is intended (T07 §3). - Runs of the same described size and move: 2 → look; 3+ → change size, angle or relation, or say why (a motif, a locked-off talk) (higgsfield; T03 §3). A talking head is one setup: punch-ins make the change.
- Cut facts:
sameSetupwith a smallsourceGapSecondsis a jump cut (punch-in or cutaway);sameFramingmeans a missing punch-in;motion,cameraMoveshow stillness and repeated moves. - Beat sync: hard cuts within ±1–2 frames; 3 is perceptible (shotcraft pass ≤3 f, ideal ≤1.5 f; T07 §3). A
median offset away from 0 with tight p10–p90: the grid's phase or the whole music is off, shift once or recheck
the downbeat, not cut by cut. A wide spread: cuts are off the grid,
rollthem. - Whole-picture hits (
impacts): a cut on the beat changes when, not how hard. Hits on every kick read as the camera shaking. One promo study: ~2–4 per 30–60 s, on the strongest beats, ~16 beats apart; other accents moved one element (title, card, number), not the frame. MV, event and hype edits use many more on purpose: decide by genre, say why whenperMinuteis high. - Word sync: a cut
insidea word clips it; move it into the gap. - Rendered:
lagMs,driftMsPerMinute: does the export land where the timeline does?
Then look: ui frame F on three consecutive cuts in one scene (three shot sizes?), review window F across a
doubtful cut, review run for problems. Adjust with roll/slip, measure again; stop after 3 rounds or when a
round changes nothing, and say what is left and why. Lock the cut before bc:audio-mix.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 197k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |