Motion films that never cut to the next slide: every beat grows out of the one before, one continuous camera, continuity measured by an oracle. Product launch films and feature demos, made with a coding agent.
Rhythm and carry — what the first two films taught
Two films, six cuts, one verdict for two different reasons: 「PPT」. motion-web failed on rhythm. Pocket Weather had the rhythm and failed on carry.
Rhythm — the three cuts of motion-web
The same nine pages, the same 15 seconds, three completely different results. Cut three shipped.
Cut 1 — the card field (rejected: 「PPT 式一个一个展示」)
Nine pages as cards on a 2.5-D field, camera dollying past each for 0.74 s, depth blur, labels. Measured: 0.02 still frames, camera never stopped, every card the same dwell.
Why it failed: uniform cadence and thumbnails. A page's motion does not survive being 700 px wide and blurred for 0.7 s. The camera move made it feel like a gallery walk — which is a slideshow with a dolly.
Cut 2 — full-frame footage, hard cuts (rejected: 「也没有节奏变化 依然是 ppt」)
Seven pages full-frame, 1.55 s each, hand drawn back in, label pill, push-in per shot, spring settle on cuts. Better legibility, still rejected.
Why it failed: still a metronome. Seven equal shots with the same shape (cut, push-in, label) is seven slides even when each slide moves. No rests, no bursts, no story — a catalogue.
Cut 3 — the reference's structure (accepted)
| t | beat | length | motion |
|---|---|---|---|
| 0–1.6 | three words | 0.4 s per word | pops only |
| 1.6–3.0 | wordmark letters drop and squash, tag wobbles | 1.4 s | small |
| 3.0–4.8 | prompt card types, hand clicks send | 1.8 s | typing |
| 4.8–7.0 | build panel ticks 7 lines | 2.2 s | ticks |
| 7.0–9.6 | three pages fly through tilted, blurred in/out | 0.9 s each | burst |
| 9.5–12.1 | four micro cards, live sims, hand pokes each | ~0.85 s, overlapping | short |
| 12.1–13.25 | browser window, slow push-in | 1.15 s | rest |
| 13.25–14.3 | MOTION · IS · THE · MATERIAL | 0.25 s each | hits |
| 14.3–15 | black, wordmark | 0.7 s | still |
Measured: 0.54 still frames, one burst, one rest, hits at the end. Shot lengths from 0.22 s to 2.2 s.
Why it worked: it tells what the thing does (the checklist is the skill's real workflow; the micro cards are its mechanisms running live; the fly-throughs are its outputs) instead of showing outputs one after another; and the energy curve has shape.
The rules that fall out
- Shot lengths: coefficient of variation ≥ 0.25 or it is a metronome.
verify_promo.py --shotschecks. - At least 25 % of frames dead-still; at least one ≥ 1 s stretch under the low threshold.
- If the reference hits, at least one burst (≥ 3 big changes inside 1.5 s), not at a uniform interval. A concept without hits — one element transforming, one continuous camera move — has none, and that is not a flaw.
- Show mechanisms as components in cards when the page's motion is too small to read at card size.
- Show pages full-frame only for a hold, with the hand visible.
- Never end a shot the same way twice in a row (spring-out, blur-out, hard cut — alternate).
- Kill any flourish that is not the product's own behaviour (confetti, specks, sparkles).
Carry — the three cuts of Pocket Weather Club
Made in Codex with this skill, after motion-web. It had learned the rhythm lesson: v1 passed every rhythm leg. It was still rejected as PPT.
v1 — rhythm passed, rejected: 「像 PPT」
Key visual, character showcase, full-page recording, brand end card, 4K60. Cadence, rest, burst, not-flat and
audio all passed. The code shows what those legs cannot see: every beat is its own scene, and showing one sets
display:none on the rest. probe.py afterwards: two boundaries, both bare — carry score 0.00.
v2 — the first 6 s accepted, the rest sent back
Re-cut as a 1080p30 draft. Sunny lands on the headline and knocks its letters apart, hops to its slot, the cards spring up around it, the hand picks it, the page grows out of the card. Then three recorded pages one after another and a pocket lockup. The user kept the first 6 s and sent back the rest. probe.py: carry score 0.40 — carried at 2.97 s (Sunny) and 11.60 (the lockup); bare at 6.03 (the page growing out of the card, a seam v3 kept), 8.13 and 9.83 — page replaced by page, in the rejected half.
v3 — carried throughout, accepted
The first 6 s kept; the rest rebuilt so every beat grows out of the one before: a push into Sunny on the page, a
circular stage opening from Sunny until it is the frame, the hand dragging an elastic rail of moods while the words
lag, the cast gathering on arcs and folding into a turning seal as the name rises from its masks. probe.py:
carry score 0.75 — carried at 2.97 (Sunny), 8.27 and 9.17 (the hand), 11.30 (the cast); anchored at 7.27
(Sunny, unmoved while the stage opens); bare at 6.03. It failed not-flat (0.33), the leg v1 had passed.
The numbers
| film | verdict | carry score | carried · anchored · bare | rhythm legs |
|---|---|---|---|---|
| motion-web cut 3 | accepted | 0.56 | 5 · 0 · 4 | all pass |
| Pocket Weather v1 | rejected | 0.00 | 0 · 0 · 2 | all pass |
| Pocket Weather v2 | first 6 s accepted | 0.40 | 2 · 0 · 3 | all pass |
| Pocket Weather v3 | accepted | 0.75 | 4 · 1 · 1 | not-flat fails (0.33) |
Continuity separates the verdicts; no rhythm leg does. So verify fails a carry score under 0.5 and warns under 0.7. motion-web passes at 0.56 with four bare boundaries — the setup words giving way to the wordmark (1.6 s), the checklist to the first fly-through (7.1), one fly-through to the next (8.0), the micro cards to the hold (12.3) — and would be a better film carried.
Curves separated nothing: 57–72 % of all travel was on soft curves (t80 ≥ .55) in all four films, median t80 .64–.71. All four were rendered without motion blur, and three have moves far over 80 px/frame at 30 fps (v2 664, motion-web 385, v3 354).
How the probe decides, and where it is blind
A boundary is a 0.4 s window in which most of the on-screen area changes identity; fades through the ground are
bridged. Carried: something survives and moves ≥ 2 % of the frame diagonal or scales ≥ 10 %. Anchored:
the only survivor is large (≥ 3 % of the frame) and unchanged. Bare: nothing survives. Exempt: inside a burst
of hard cuts (≥ 3 within 1.5 s), the last 1.2 s, and times listed in __meta.cuts. Score = (carried +
anchored / 2) / boundaries considered.
It tracks identity, so a carry built from two elements is invisible to it. At Pocket Weather 6.03 s the page is a
new image fading in at the card's size while the card fades out: a viewer sees the card become the page, the
probe sees a replacement — and the user accepted that seam twice. Build a carry as one element changing
(morphRect, one layer scaled by zoomThrough), or list the time in __meta.cuts with the reason in the beat sheet.
Two more blind spots, from the camera films:
- One continuous world has few or no boundaries. knockon, a top-down machine under a moving camera, scored continuity 1.00 over 0 boundaries; clearing did it over 3 and unbroken's split over 1. There the carry was judged by eye. A PASS over fewer boundaries than the beat sheet has is not evidence.
- rest reads energy at 320 px. It over-reads small detail on a dark ground: a dot on black (one-dot v1, still 0.844) or bone-white brush strokes (unbroken, 0.628). It also counts any camera push as motion: clearing's first render failed at 0.178 because every hold had a slow push. Make the rests truly still; then the leg is honest.
The rules that fall out
- Write the carry into the beat sheet: for each boundary, the element that survives it and what it does.
- The carries that worked: the thing pressed becomes the result, a stage opens from the subject, the hand drags the next option in, the cast folds into the lockup (motion-library.md → carry). A fade or a slide between unrelated scenes is not a carry.
- The hand is the cheapest carrier: keep it on screen across a boundary it caused.
- Page after page is a slideshow even when each page moves (v2 at 8.13 and 9.83, motion-web at 8.0).
- Bare cuts only in a burst of hits or into the end card.
- Run probe.py on the first playable storyboard. Its
barelines are the beats that will read as slides. - Rests are dead still: no
drift, no settling camera. A push in a hold is motion to verify.
Camera on the beat — the domino film's tree (8–12 s)
Five splits on beats 20–28 turn one line of phones into 32. Two cuts of the camera, one lesson.
v6 — one smooth move (a low three-quarter behind the fronts, widening over 4 s; lines at constant speed): 「7-12s, 看着有点无聊呢,没啥节奏变动」. Every split landed on a beat, and nothing on screen moved with it.
v7 — a new set-up snapped to on every split, sides alternating, each higher and wider, a stiff spring (ω 8.5), a
small shake per split; and each generation's lines starting slow and surging into the next split (acc, references/3d.md). The
surge stayed. The camera was sent back: 「除了镜头左右摆来摆去…有点生硬了」. Equal-sized jumps every 0.86 s are a
metronome, and alternating sides crossed the line on every beat — the lines' screen direction flipped each time, so the
eye re-found the picture five times in four seconds.
v8 — accepted (「牛逼,可以了」). One direction for the whole stretch, and the beat carried by speed, not by position:
- motivated movement first: swoop off the previous set-up onto the trunk and ride the hero front FPV (0.42 m behind, 11 cm up, fov 44). At each split the hero line veers and the camera banks with it — roll = −0.5 × the front's turn rate, clamped to ±0.22 rad, on a softer spring than the move. The turn has a reason, so it isn't a jolt;
- then a stepped pull-back: from split 3, one step per split — 0.24 s,
cubicOut, starting 40 ms early — back and up along a single direction, landing exactly on the next section's set-up (make that set-up a function,set32(t), and interpolate look / direction / log-distance towards it). Between steps the camera drifts with the fronts.
What it suggests (a way of thinking, not a camera recipe — the right move depends on what the section is about):
- A camera that reframes on every beat becomes the metronome. Hitting the beat by jumping sides also flips the picture's direction each time; the eye pays for that. Movement with a reason (following the thing that turns) and accents in speed, scale or lens read as rhythm without the jolt.
- The subject can carry the rhythm too. A front that surges into its beat makes the beat visible even from a still camera; a constant-speed chain only lands on it.
- Think about what each part of the stretch is for — here, close for force, far for the count — and where it has to end up for the next section to begin.