rollShort product video looks simple and isn't. Twenty seconds is roughly six scenes, and every one of them has to arrive, land, and leave without the viewer noticing the machinery. This is a breakdown of the grammar we use when directing them, useful whether you're building your own or just want to know why some reels feel expensive and others feel like a slideshow.
Time is a grid, not a stopwatch
The first rule: scene lengths are musical, not arbitrary. Pick a tempo, then measure every scene in bars rather than seconds.
At 122 BPM, a bar is about two seconds. A scene of two bars is four seconds. When cuts land on bars, the edit feels intentional even when the viewer has no idea why. When they land at 3.7 seconds because that's how long the animation took, it feels amateur, and again the viewer cannot say why.
This has a practical consequence: you cannot cut to the beat after the fact. The timing has to be the grid the scenes were built on.
Nothing on screen is ever completely still
A flat background reads as a missing asset. Every frame should have some low-amplitude life in it:
- Two large radial washes in the brand color, drifting slowly in different directions.
- Film grain at low opacity.
- A vignette so the edges don't compete with the center.
None of these are perceptible on their own. Together they're the difference between "video" and "PowerPoint." The same rule applies to elements that have finished animating: after a card arrives, it should keep breathing: a fraction of a degree of rotation, a percent of scale, staggered so no two objects drift in sync.
One entrance, used everywhere
A common failure in generated video is a different transition for every element: one slides, one spins, one bounces. It reads as a template demo.
Pick one entrance and use it for everything: elements rise a short distance while fading in, with a slight overshoot at the end. Stagger the delay across items so a list assembles instead of appearing. That's it. Variety comes from timing and scale, not from a catalog of effects.
The exception that earns its keep is the mask reveal: type that wipes in behind a moving clip edge rather than fading. Used once or twice in a reel it feels crafted. Used everywhere it feels busy.
Depth has to be real
Parallax faked with two layers moving at different speeds looks like two layers moving at different speeds. Real depth needs three things working together:
- A perspective projection on the stage rather than 2D transforms.
- A camera move applied to the whole composition: a slow dolly forward, a drift left.
- Per-layer depth that changes how far each element travels through that move.
When those combine, near objects sweep past faster than far ones because of the projection, not because you animated them differently. That's the cue the eye reads as three-dimensional space.
The product itself is the best asset
Stock illustration is the enemy. The strongest scenes in any product reel are made of the product's own pixels: not a screenshot of the whole page, but individual components lifted out of it, a pricing card, a stat, a button, a countdown.
The trick that makes this look like motion design rather than a chopped-up screenshot is the cutout. Capture each element with a transparent background, including its rounded corners, so it composites as a real object with its own silhouette. Then those objects can overlap and occlude each other in 3D space. A pricing card passing in front of a stat block reads as a designed composition; two rectangular screenshots overlapping reads as a mistake.
One caveat learned the hard way: bare text should keep its background. Cut out dark type and composite it onto a dark stage and it vanishes.
Narration leads; the picture follows
If there's a voice, it is the spine of the edit, not a caption track laid over finished visuals. Three rules follow from that:
- Write one performance, not six lines. A 35–55 word arc distributed across scenes, rendered as a single take, so the prosody carries across cuts. Six separately-generated sentences sound like six sentences.
- Scenes stretch to fit speech, quantized up to the nearest bar. Speech length sets scene length, not the reverse.
- Visual hits land on emphasized words. Take the word-level timings from the render, find the word each scene is really about (usually a number, or the word the on-screen type repeats) and snap the scene's sound and motion accents onto it.
That last one is the highest-leverage detail in this entire piece. When a whoosh and a scale-punch land exactly on the word "instantly," the whole thing feels directed. When they're 200ms off, it feels automated, even though nobody can articulate what's wrong.
Silence matters too. If a director deliberately leaves a scene unnarrated, that's a beat, not a gap to fill with a caption read aloud.
Duck the music once, not per sentence
When narration plays, the music should drop about 3.5 dB, and it should be one envelope across the whole passage, not a dip per line. Passages less than about 1.2 seconds apart should merge into a single duck. Otherwise the score audibly pumps up and down between sentences, which is the single most common giveaway of an automated mix.
Master the final mix to −14 LUFS with a −1 dBTP ceiling, which is what social platforms normalize to anyway. And normalize one-shot sound effects by peak, not RMS: a tick is mostly silence, so RMS-matching it to a two-second riser makes the tick shout.
None of these are opinions about taste. They're the difference between a viewer thinking "nice product" and thinking "nice video about a product I now want." If you'd rather not implement any of it, that's what we built: paste a URL and the whole grammar above gets applied to your product automatically. Or just watch what it produces on real sites in Dailies.