A good explainer video shows the real product doing one real thing per scene, and uses timing — not arrows, not captions — to carry the meaning. Across sixty-odd produced videos and the review notes behind every one of them, the distilled ranking criterion came down to a single sentence: the best are visually diverse, dynamic, clean, with some visual harmony.
That sentence sounds vague until you measure it against actual graded rankings. "Diverse" tracked how many different surfaces were alive on screen at once. "Dynamic" tracked whether the beat shapes varied or repeated. "Clean" was a ranking axis of its own — visual artifacts cost an otherwise-loved video two places. The rest of this guide unpacks the rules underneath those words, each one paid for by a video that shipped wrong.
The picture teaches; the voice explains
The screen owns motion and state. The narration owns words. When a step fails, the picture shows a red ✗ and a ring; the voice says why it matters. A label may name a thing ("Table") — it may never explain it.
The quick test: watch the video muted. If you can read the whole script off the screen, the video is a slide deck with a fade, and the voice and the text are competing for the same job. Every on-screen sentence the narration already says should be deleted. This division of labor is also why the script comes after the visuals, never before — narration written first produces pictures that illustrate, and illustration is the weakest thing a picture can do.
Sync is meaning
When two things on screen are causally related, they must change together — or at a deliberate, fixed offset. A ring lighting on a block as a row lands in a table says "this block wrote that row" without a single connector line. Viewers infer causality from timing, not from arrows.
There's a tuned number here: cause precedes effect by about 0.7 seconds. Simultaneous cause-and-effect reads as coincidence; a 0.7-second gap reads as consequence. That constant wasn't designed — it was mined from the renders that graded well. And the inverse rule matters just as much: an animation that can't name its sync partner is decoration. Cut it or pair it.
One thing is focal at a time
Attention is a budget. If three things animate at full strength, the viewer picks one at random — usually the wrong one. The fix is dimming: something is always moving, but everything non-focal drops to roughly a third of full strength. Dimming is how a film points. Stillness is not focus; stillness is a dead frame.
There's a hard floor, too. In one rejected build, the climax was a one-word text change inside a small box on a wide, never-moving frame — maybe 1/40 of the screen. The review verdict was four words: "don't see anything cool." The rule that came out of it: if the payoff is smaller than about 1/40 of the frame, it did not happen. Push the camera in, scale the element up, or restage the event.
Holds are earned
Narration almost always outlasts animation, and the lazy default is to let everything settle early and sit frozen while the voice finishes. One measured bad build spent 44% of its runtime fully static — 52 seconds of settled frames in a 93-second video. It passed every mechanical check and still died in review.
The distinction is alive versus dead. A hold after a payoff is a breath. A hold before anything has happened is dead air. An alive hold has ambient motion budgeted into it — a slow camera drift, a ticking cell, a working dot on the still-live element — and it costs about one line of effort per scene. Dead holds get capped at 3 seconds, and the better fix is to re-pace the beats so they land as the voice names them, so nothing settles early at all.
The camera moves between ideas, and cuts within one
Camera moves signal "new idea"; cuts happen inside an idea. Every camera ease settles before the beat it frames — a camera drifting during a payoff steals the payoff. And when the content on screen is maximal, the camera does nothing.
This discipline is measurable. When the same topics were built twice — once rejected for visual quality, once accepted — the rejected takes had zero named camera framings: geometry improvised per scene, so content got cropped or cut in half. The accepted takes defined 35–59 named framings over one fixed layout and moved a single continuous camera between them. Same tooling, similar amounts of code. The camera work separated them.
Show the real product
The same built-twice comparison found exactly three differences between rejected and accepted takes, and the biggest was surface choice. Every rejected take invented its display surface — a custom table hand-built while the real table component sat unused, a floating list, a fake inbox. Accepted takes used the product's real surfaces 4 to 13 times each. Rejected takes: zero.
Viewers who know the product notice invented UI instantly. Viewers who don't still sense the video is a brochure rather than a demonstration. Spectacle should come from real things visibly running — a queue flipping status row after row, a table filling, parallel work landing. If a concept can't be shown truthfully with the product's real surfaces, the concept is wrong. Change the concept, not the honesty. There's a full guide on this: show the real product.
Vary the rhythm
Uniform timing reads as a slideshow. Real systems don't complete in order at even intervals — so parallel work that finishes in a visibly shuffled order reads as a real system, and events arriving at compressing intervals read as a system under load.
Even stagger speed carries meaning. In our timing constants, 0.35 seconds per item is presentation cadence — a narrator pointing at things one by one. 0.14 seconds per item is machine cadence — a system sweeping a selection. Use one where the other belongs and the tone breaks: the machine looks sluggish, or the narrator looks frantic.
One more graded finding: a technically clean build that repeated the same good beat five times was rejected with a one-word verdict — "boring." Repetition of a good beat is still boring. The ranked-best builds varied the beat shape across a single run: a pull, a fan, a lane close-up, a scrambled finish.
The frame should remember
Fresh events pulse to full strength, then decay to about 35% — and stay. The record of what happened is made of leftover light, and permanent full-strength color is reserved for the one signal that must keep burning.
A video that resets after every beat teaches nothing cumulative. The loved builds all had a surface that visibly fills: a status column flipping row after row, evidence cards stacking with each action, a table populating as work lands. Accumulation makes progress visible, and it gives every later frame more meaning than the one before it.
Spend your runs carefully
Viewers give a machine's first run full attention, the second run less, the fifth almost none. Use the minimum number of end-to-end runs that proves the point — one graded build went from seven runs to three and got stronger. A second run is justified only when its outcome differs from the first in exactly the dimension the lesson needs.
The strongest closing frame in our corpus is a run that returns nothing: the video re-fires its opening query and selects zero rows, because every row had already been processed. The empty result was the entire thesis, proved by contrast. This only works because each scene carried one idea and the contrast was staged on purpose.
A short checklist
- Muted, can you read the script off the screen? Fix the text walls.
- Pause at random. Landing on a fully static frame more than a third of the time means it's a slideshow with a soundtrack.
- Can every animation name its sync partner? No partner, no animation.
- Is the climax at least 1/40 of the frame? If not, it didn't happen.
- Does any surface fill up across the video? If nothing accumulates, nothing was accomplished.
- Are the surfaces real? If a viewer who knows the product would flinch, so will everyone else — they just won't know why.
FAQ
What's the single most common failure? The dead hold: everything animates in a scene's first three seconds, then freezes while the voice finishes. It's the default failure of template-made and machine-assembled video, and it's why the pause-at-random test works so well.
How long should a good explainer be? Most products are best served by 60–90 tight seconds — roughly 6–8 scenes of 8–12 seconds, one idea each. Length has its own guide: explainer video length.
Can you judge a video before it's fully produced? Largely, yes — from stills. In our production process, reviewing two static frames (the money shot and the final frame) before any motion catches the majority of fatal errors at about 5% of the cost of a finished render. Wrong surfaces, broken framing, and invented UI are all visible in a single frame.
Does following these rules guarantee a good video? No. One build passed every written gate we had and was still graded worst of its batch. The rules eliminate the known ways to fail; the remaining gap is taste, and taste has to be watched for — a scene is done when its rendered frames have been looked at, not when the checklist is green.
If you'd rather judge finished directions than briefs, send us your product's URL and pick from twenty.