How to write a motion prompt that actually reproduces

Oct 25, 2026

Browse any gallery of AI motion videos and a pattern jumps out: the videos are impressive, but most of the prompts behind them are a single line. One line can produce a great video — once. Run it again and you get a different video, because the prompt left almost every decision open: the subject, the look, the timing, what to avoid, and how the frames get rendered at all.

That is why this site exists. For every entry in our gallery we either wrote the prompt ourselves or rebuilt it from the video, then ran it and published the render next to it. The prompts that reliably produce good motion share five parts.

1. Inputs first

Start by telling the model what to ask you for: the subject, the states or shots, the palette, the music tempo. A prompt that begins by collecting inputs cannot silently invent the wrong product, the wrong colours, or the wrong length.

2. Direction — one concept, and a banned list

Give the video a single idea it can keep returning to ("one shape, never cut", "type is the whole show"). Then ban the defaults you do not want: glows, particle bursts, bouncy cartoon easing, gradients on everything, anything that looks like a template. Models reach for those clichés whenever a prompt leaves a gap; a banned list closes the gap.

3. Structure on a beat grid

Replace the word "dynamic" with a grid: a tempo, a bar count, and one event per beat. Cuts and morphs that land on beats feel designed; the same events at random times feel generated. List the shots or states in order, and make the last frame equal the first if the video should loop.

4. A build method, named explicitly

A motion prompt is not finished when it describes what the video shows — it must say how the video gets made. Our runs use HyperFrames: one composition where every frame is a pure function of time, seek-driven, rendered frame by frame. Naming the renderer and the seek rule is what makes the result repeatable on another machine, not just watchable once.

5. Gotchas and a self-review step

Every renderer has known failure modes: transforms anchored to the wrong point so letters collide, radius and size on different tweens so corners wobble, linear easing where a spring belongs. Writing the ones you know into the prompt — and instructing the model to render one frame per section, look at it, and fix problems before the full render — catches the errors that otherwise ship.

Start from a tested one

You do not have to write all five parts from a blank page. Every template on this site is a prompt we ran ourselves, with the variable parts turned into fields: change the subject, palette and size, and copy a version that keeps the structure. Start with the showreel template or browse all templates.

Method note: this article summarises the structure of the prompts behind our own gallery entries at launch — a small, growing sample. As the gallery grows past 50 tested entries we will publish the full data study, the way we count, and what changed.