Blog · October 19, 2026 · Maxwell Prehoda

Motion Graphics Prompts: The Anatomy That Works, With Free Ones

Motion graphics prompts that work have five parts: inputs, direction, structure, build and check. Every one of our 320 free prompts is built that way, and each one has a video showing exactly what it made, so you can check the claim yourself.

Below is each part, why it's there, and short excerpts from three of ours. They're CC0, so take what you want.

What most AI motion graphics prompts look like

Short. Of the 193 community prompts with text in our Motion Prompts directory, the median is 25 words. Ours have a median of 1137.

A one-line prompt still makes a good-looking video, because the model fills every gap with its own taste. That works right up until you need the video to be about something. Then the gaps are the problem. We broke down the community's prompts in what real AI motion design prompts get right and wrong, and the best ones covered the same five parts, whether or not they used the headings.

The anatomy of a motion graphics prompt

1. Inputs: what it's about

The content goes first: the name, the words on screen, the colors, the length. Every input gets a default, and one line tells the model to use it instead of stopping to ask. Without that, Claude Code may stop to ask questions or invent filler, and filler looks like filler. Defaults also mean the prompt runs as is, so the preview shows what you'd get.

From Gradient Mesh Brand Film:

<inputs>
If I don't give you something, use the default and don't ask.
- Brand: Northwind
- Headline: "Move money the way you move data."
- Subline: "One API for payments, payouts and treasury."
- Mesh colors: #6A4CFF, #FF5C8A, #FFB23F, #18C2E6
- Ink: #0B1736
- Length: 10 seconds
</inputs>

2. Direction: numbers, not adjectives

"Snappy" is a guess. A cubic-bezier and a duration are an instruction. Direction is where the style lives, and the more of it you can put in numbers, the less of it the model has to make up.

From Cut-Paper Title Sequence:

Motion is jazz: everything lands on the beat grid (132 BPM; beat n starts at (n − 1) × 0.4545s; half-beats allowed), and scenes change with hard cuts on a beat. Shapes snap in over 0.18–0.3s on cubic-bezier(0.2, 0.9, 0.3, 1), hold, and snap out. Quantize paper motion to 12 steps per second (Math.floor(t * 12) / 12) so it feels animated on twos; cuts stay exact. The cheap version is smooth tweens under a paper texture; the real thing is rhythm, reduction and one strong symbol.

Look at the last sentence. It names the cheap version of the style so the model can avoid it. That one line does more than a paragraph of praise for the real thing. The same prompt then bans its defaults outright:

Banned: gradients, glows, blur, crossfades, photo textures, a second motif, rounded icon-like shapes, soft easing on cuts.

All 320 of our prompts carry a banned list. It's the fastest way to stop a video looking generated, because most of what makes AI motion look like AI motion is the same handful of defaults. Liquid Glass Interface puts it in one line: "Premium is refraction visible at the rims, highlights that answer the light, and heavy, calm motion. Generic is a blurred translucent box."

3. Structure: beats, with times

Say what happens and when. One event per beat, in order, with holds. Skip it and you tend to get a reel of effects instead of a piece with a point.

From Liquid Glass Interface:

<structure>
- 0.0–1.2: The background flows; stripes and word drift.
- 1.2–2.6: A 760×420 player panel springs up from below; stripes bend through its rim.
- 2.6–3.8: A weather panel slides in from the left, a notification pill from the right.
- 3.8–5.4: Two 96px round buttons drift together and fuse into a pill that stretches into a scrubber.
- 5.4–7.0: All panels glide 80px left as the light swings; highlights travel the rims.
- 7.0–8.4: Panels gather and merge into one 360px round lens as content fades.
- 8.4–10.0: The lens settles over the background word, which fills solid white under it; the tagline appears below. Hold.
For other lengths, scale beats; keep the fuse and a 1.5 s minimum final hold.
</structure>

The last line matters. It makes the prompt reusable at any length without rewriting the beats, and it keeps the final hold.

4. Build: one file, every frame from time

Name the stack and the format, and make every frame a function of time. If the animation runs on CSS transitions and timers, it plays fine in a browser but won't render cleanly frame by frame.

window.seek(t), t in seconds, sets every visual from t alone: no CSS transitions or animations, no setTimeout/setInterval, no state carried between frames. Randomness comes from a seeded PRNG; springs are closed-form functions of time.

All 320 of our prompts ask for that seek(t) function. It's what let us render every preview to video, and it's what lets you do the same. Here's how to export to MP4.

5. Check: make it look at its own work

Claude Code can screenshot frames and look at them. Ask it to, before it shows you anything. The general checks (overlaps, clipped text) catch the obvious problems. The style-specific line catches the one that would ruin this particular piece.

From Liquid Glass Interface:

<check>
Before showing me, screenshot seek(0) through seek(10), one frame per second, and look at each. Fix overlaps, clipped or cramped text, and anything generic. Stripes must bend only at the rims, and panel text must stay readable over bright colors. Confirm the final frame holds.
</check>

A motion graphics prompt template

The same five parts with the specifics taken out. Fill it in, or start from one of ours and change the inputs.

Make a [length]-second [kind of piece] about [subject].

<inputs>
If I don't give you something, use the default and don't ask.
- Name: [default]
- Headline: [default]
- Colors: [3 or 4 hex values]
- Length: [seconds]
</inputs>

<direction>
Palette, type and motion, in numbers: sizes, eases, durations.
One line on what premium looks like, and one on what generic looks like.
Banned: [the defaults you never want to see].
</direction>

<structure>
- 0.0-1.5: [one thing happens]
- 1.5-3.0: [the next thing]
- ...and a final hold.
For other lengths, scale the beats; hold the end at least 2 s.
</structure>

<build>
One self-contained index.html, 1920x1080. window.seek(t) sets every visual from t alone.
</build>

<check>
Before showing me, screenshot one frame per second and look at each. Fix overlaps, clipped text and anything generic.
</check>

Free motion graphics prompts to start from

All 320 of ours are in the Motion Prompts directory, each with its video, free under CC0. Browse them by style:

Or drag the whole inspiration board and open whatever stops you. Our guide to the board shows how to combine three pieces into one prompt.

When you reuse one, change the inputs, keep the direction, and only touch the structure if your length or story is different. Run it once before you edit anything. You'll learn more from the first render than from rereading the prompt.

Questions

What should a motion graphics prompt include?

Five parts: inputs (what it's about, with defaults), direction (the style in numbers, plus a banned list), structure (timed beats), build (the stack, and every frame computed from time) and check (screenshots the model reviews before it shows you).

Are these motion graphics prompts free?

Yes. All 320 of ours are CC0. Use them for anything, including client work, with no credit needed.

How long should a motion graphics prompt be?

Long enough that the model doesn't have to guess. Ours have a median of 1137 words. A one-liner can work when you want the model's own taste, like the viral showreel prompt, but not when the video has a job to do.

Do they only work with Claude?

We wrote them for Claude Opus 5.5 in Claude Code and rendered every preview with that setup. That's the only setup we can vouch for.