20cuts

Resources · Craft · 7 min read

What makes a good explainer video?

A good explainer video shows the real product doing one real thing per scene, and it uses timing to carry the meaning.

You have probably watched a draft that felt lifeless without being able to say why. The animation was smooth, the colors were right, the script said true things. Across sixty-odd graded builds of our own, polish almost never explained the rankings. Diversity, motion, cleanliness, and harmony did, and each cost us a video that shipped wrong before we could state it as a rule.

Picture a 90-second explainer for a deploy tool: a queue of incoming commits, three build workers, and a status table that fills in as builds finish.

The picture teaches, the voice explains

Most videos get written script-first, and that order produces the genre's most common failure: a screen that restates the voice in text at the same moment. When both channels carry the same words, they compete for the same job, and the viewer reads instead of watches.

The screen owns motion and state, and the voice owns reasons. When a build fails, the worker's ring turns red and a ✗ lands in the status table while the voice explains why a failed build stops the release. A label may name a thing ("Queue"). The moment it starts explaining, it has taken the narrator's job.

Timing carries the causality

Film has no arrows. When a viewer concludes that a worker wrote a row into the status table, the belief came from timing: the ring lit, then the row landed. Viewers infer cause and effect from when things happen, so timing decisions are meaning decisions.

In the renders that graded well, cause led effect by about 0.7 seconds. Simultaneous events read as coincidence, a long gap reads as two unrelated events, and at around 0.7 seconds the second event reads as a consequence of the first. We did not pick the number; the approved renders kept doing it.

Watch the same cause-and-effect pair twice. On the left, the result lands 0.7 seconds after the trigger fires — you read causation. On the right, both fire on the same frame — you read coincidence. Rendered by the same system that makes our videos.

One thing is focal at a time

When three things move at full strength at once, the viewer picks one at random, usually the wrong one. Dimming is the fix: everything that is not the current subject drops to about a third of its strength. When worker two is the story, workers one and three keep working at 35% brightness, and the eye goes where the video points.

Size has a floor too. One of our builds staged its climax at about 1/40 of the frame, and no viewer could feel that it happened (the rejection, verdict included). Three fixes for an undersized payoff:

A hold is alive or dead

Narration runs longer than animation, so every video has stretches where the picture is done and the voice is still talking. One of our measured builds froze for nearly half its runtime and died in review (the measurement). A short hold after a payoff gives the viewer a breath. Before anything has happened, the same hold is dead air.

Two holds on the same card. The alive one keeps a value ticking and breathes once with a soft ring. The dead one is pixel-frozen — the timer under it shows how fast three seconds of stillness starts to cost you the viewer.

The camera moves between ideas and cuts within one

A camera move tells the viewer the subject has changed. A cut just changes the angle on the subject they already have. We built the same topics twice, and the rejected takes improvised their geometry scene by scene and kept cropping content, while the accepted takes moved one continuous camera between named framings over a fixed layout (the counts). The tooling and code volume were similar, so the camera work carried most of the grade.

Show the real product

The built-twice comparison exposed a larger difference. Every rejected take invented its display surface — a hand-built table while the product's real table sat unused — and the accepted takes drew on real surfaces throughout (the exact counts). Viewers who know the product notice invented UI instantly, and viewers who don't still sense they are watching a brochure.

Real things visibly running are spectacular on their own: the queue drains commit by commit, the workers finish out of order like real machines, the table fills row by row. If a concept cannot be shown truthfully with real surfaces, change the concept, not the honesty.

Shuffled timing reads as a real system

Real systems do not finish their work in order at even intervals, so parallel work that lands in a visibly shuffled order reads as real. The speed of a stagger carries tone as well: a slower cadence reads as a narrator pointing, a faster one as a machine sweeping. And even a well-made beat wears out — one technically clean build repeated its best beat five times and drew the one-word review "boring" (the exact cadences and rhythm rules).

Six chips revealed twice. At 0.35s apart each chip is its own event — a narrator pointing. At 0.14s apart the same six read as one sweep — a machine selecting. Same elements, different meaning, and the only variable is the gap.

The frame keeps a record

In the builds reviewers loved, fresh events pulsed to full strength, decayed to about 35%, and stayed there, so the frame accumulated a visible record of what had happened. A video that resets to a clean frame after every beat teaches nothing cumulative. By second 80 of the deploy video, the status table itself is the proof that work happened.

Each result flares to full green, then decays to a 35% residue — and stays. By the end of the run the frame carries a visible history: three things happened here, in this order.

Repeated runs spend attention

Viewers give a machine's first end-to-end run their full attention, the second less, and the fifth almost none, so use the fewest full runs that prove the point; one graded build got stronger going from seven runs to three. The strongest closing frame in our corpus re-fired the opening query and selected zero rows, because every row had already been processed. The empty result proved the thesis by contrast, and the staging depended on each scene carrying one idea.

A short checklist

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. The pause-at-random test catches it.

How long should a good explainer be? Most products fit 60 to 90 tight seconds — roughly six to eight scenes of 8 to 12 seconds, one idea each (length).

Can you judge a video before it's fully produced? Largely yes, from stills. Two static frames per scene, reviewed before any motion exists, catch most fatal errors at about 5% of a finished render's cost.

Does following these rules guarantee a good video? No. One build passed every written gate we had and still graded worst of its batch. The rules remove the known ways to fail; the remaining gap is taste, so we treat a scene as done only after its rendered frames have been looked at.

To see this on your own product, send us the URL. Twenty short candidate videos come back, and you pick the ones worth finishing.

See the answer for your product instead of the average:

Get my 20 free videos