Blog · · Maxwell Prehoda
AI Motion Design Prompts: What 217 Real Prompts Get Right and Wrong
We went through every motion design video we could find that someone made with AI and posted with the prompt. After dropping reposts and posts with no prompt, that left 217 prompts, nearly all of them for Claude Opus 5.5, posted in the last weeks of September and early October 2026. They're all in our Motion Prompts directory, credited to the people who wrote them.
The short version: most of these prompts are one sentence of vibes. The videos still look good, because the model fills every gap with its own taste. That works right up until you need the video to be about something specific. Then the gaps are the problem.
The numbers
- The median prompt is 26 words. 82% are under 40 words.
- 57% give a length. "15-second" is the single most common detail anyone specifies.
- 44% lean on hype: "go all out", "incredible", "impressive", "stunning".
- 22% mention a product, brand, logo or website.
- 16% say anything about music or sound.
- 13% name a stack (HTML, Remotion, three.js, GSAP, SVG, Canvas).
- 8% mention easing or timing. 8% mention color or type. Only 5% give an aspect ratio.
One more number, with a warning attached. The 16 prompts over 150 words had a median of about 20,000 views. Prompts under 40 words had a median of about 600. Don't read that as "long prompts go viral." Most of the long ones are templates posted by accounts that already had an audience. What it does tell you is that the people who take motion seriously write long prompts.
Why one-line prompts work at all
The most-copied prompt in the set is a single sentence asking the model for a 15-second showreel that proves it's a great motion designer. Forty-one creators posted runs of it. (We wrote it up separately: the showreel prompt, explained.)
It works because it asks for nothing in particular. There's no subject to get wrong, no brand to match, no story to tell. "Go all out" is permission to spend tokens. The model picks the moves it's best at: kinetic type, geometric shapes, particles, a UI that assembles itself. You get a strong reel of techniques.
Effort helps more than wording. 48 of the posts say which effort setting they used, and max effort is the most common by a wide margin. One creator, @shneural, ran the showreel prompt in stock Claude Code on max reasoning, one shot, no edits. The model decided on its own to build the 3D parts in Blender, which happened to be installed. It cost about $81 at API prices.
Why they stop working
The moment you need your logo, your product, a specific length for a specific platform, or an actual message, a vague prompt gives you a generic video with your name on it. The model reaches for the same moves every time. Kinetic type shows up in nearly a third of the directory. Particles and shape morphs are everywhere.
Generic isn't wrong. It's just not yours. Nobody remembers a launch video that looks like the last twelve launch videos.
What the best prompts specify
The prompts that produce specific, usable work all cover the same five things. They don't always use these headings, but the parts are there.
1. Inputs: what you're giving it
Logo, screenshots, product copy, a song. The best template in the set, @twoclipping's morphing UI loop, opens with an inputs block that makes Claude ask for 8 to 12 UI states and a royalty-free song around 120 BPM before it does anything else. If you don't give the model real material, it invents placeholder material, and placeholders look like placeholders.
2. Direction: a short style brief and a banned list
Two or three reference words beat a paragraph of adjectives. The banned list matters more. The same template bans bouncy easing, particle bursts, glows, gradients on UI chrome, dead time and "anything that looks like a template." Telling the model what not to do removes its defaults, which is most of what makes AI motion look like AI motion.
3. Structure: beats, with times
Say what happens and when. @Gdgtify's "Build the Floor" timestamps every line of a 20-second piece, down to a two-second hold where nothing new appears. That level of detail is why it reads as directed instead of generated. You don't need timestamps for everything. You need an order, and you need holds.
4. Build: the stack and the format
Only 13% of prompts name a stack, and only 5% give an aspect ratio, so most people find out what they got when it's done.
Decide up front: one HTML file, Remotion, or three.js. Give the resolution and frame rate. If you'll export to video, ask for
every frame to be computed from time (a seek(t) function) so it renders cleanly frame by frame. Our post on
editing and exporting AI motion covers why.
5. Check: make it look at its own work
@aschapmann's Remotion launch prompt ends by telling Claude to render the video, check the frames itself, and fix anything that overlaps, clips or feels rushed before showing the result. That one sentence saves you two rounds of notes. Claude Code can screenshot frames and look at them. Ask it to.
A template you can copy
This is ours, built from the patterns above. Fill in the brackets and delete what you don't need.
Make a [length]-second motion graphics video about [subject] for [where it will be posted, e.g. X at 16:9].
Inputs: I've put [logo.svg / screenshots / brand colors] in this folder. Ask me for anything missing before you start.
Story: [beat 1] -> [beat 2] -> [beat 3] -> end card with [name and URL].
Style: [two or three reference words]. One typeface, [colors]. Springs with little overshoot. Hold every line long enough to read it twice.
Avoid: [particles, glows, lens flares, whatever you're tired of seeing].
Build: one HTML file, 1920x1080, 60fps. Compute every frame from a seek(t) function so it can be rendered frame by frame.
Before you show me anything: render one frame from each scene, look at them, and fix overlaps, clipped text and anything cramped.It's longer than 82% of the prompts we collected. That's the point. Every line replaces a guess the model would otherwise make for you.
Blunt advice
- Run the first draft on the highest effort you can afford. Use lower effort for small fixes.
- Give it real assets in the folder and refer to them by file name.
- Write the banned list first. It's the fastest way to stop the video looking generated.
- Ask for a story, even a three-beat one. A reel of techniques is not a message.
- Run the same prompt two or three times before you start editing. The first take varies a lot.
If you'd rather start from something proven, browse the directory by topic: product UI, explainers, charts, kinetic type, 3D. Every entry has the creator's original video next to the prompt, so you can see what it actually produced before you use it.