Productivity
"Make it bouncy" is not a spec: prompt AI with exact motion data
Vague prompts get generic animations from Claude and ChatGPT. Give the model real keyframes, durations and easing, and get code that matches the motion you wanted on the first try.
Ask an AI model for "a smooth, bouncy card that expands into a voice chat room" and you will get something. It will probably run. It will also have a 300 ms ease-in-out on everything, because that is the average of every animation the model has seen.
The problem is not the model. It is the prompt. Words like smooth, snappy and bouncy are a long way from numbers. The model fills the gap with defaults, and you spend the next ten messages nudging it toward what you meant.
What the model actually needs
To rebuild an animation faithfully, a model needs two things:
- The motion. Each animated property, with its keyframes, durations, delays and easing.
- The look. The structure and styling of what is moving.
A description gives it neither precisely. A Capture Motion LLM prompt gives it both.
What is inside a Capture Motion prompt
On any gallery card, open View, then choose Copy > LLM prompt. You get three parts:
| Part | What it gives the model |
|---|---|
| A ready-made instruction | What to build, and to use the captured values |
| The capture JSON | Each animated property with keyframes, durations, delays and easing |
| Reference markup | The card's own markup, as a visual reference for structure and look |
If you customized the card first, your choices are already in the prompt: colour, text, speed and background. You do not have to describe them again.
For an animation on another website, capture it with the Chrome extension and use Copy for LLM. That prompt carries the motion data, so the model rebuilds the movement faithfully. It has to invent the visual design around it, so describe the look in a follow-up if it matters.
A worked example
Say you want a presence card for your team app, like the Voice Chat Expand card: a pill of faces opens on a spring into a room, the faces fly into a grid one after another, and a speaking ring hops between whoever is talking.
- Open Voice Chat Expand with View. Set Initials instead of photos, a Dark card, 5 people and the room name "Daily Standup".
- Copy the LLM prompt.
- Paste it as the first message of a new chat in Claude or ChatGPT. Do not trim it.
- Add one line about your stack: "Build this as a React component with Tailwind."
- Run what comes back, and compare it with the card.
Because the model has the real timing and the card's markup, the first answer is a faithful starting point rather than a rough sketch.
Follow-ups that keep the motion intact
Get a faithful copy first, then change it. If you ask for a copy and a redesign in one message, the model tends to drift on the timings. These follow-ups work well:
- "Replace the text with my own: [copy]. Do not change the animation."
- "Use my brand colours: primary #…, background #…. Keep the timings as given."
- "Move the durations, delays and easing curves into CSS custom properties at the top so I can tune them."
- "Add a prefers-reduced-motion version that fades in without movement."
Notice the phrase keep the timings as given. Models like to tidy numbers. A 520 ms duration becomes 500 ms, and a precise cubic-bezier() becomes ease-out. Saying it every time costs nothing.
Check it in two minutes
Before you build on the result, open the prompt's JSON next to the code and check each track. Look hardest at easing and delays. If a curve became a keyword, or a stagger lost its spacing, list what differs and ask the model to use the given values exactly. One correction message usually fixes it.
A quick trick: ask the model to "list each animated element with its property, duration, delay and easing as a table". It doubles as a timing check.
Why this is a productivity win
The saving is not one prompt. It is the conversation you no longer have. A vague prompt costs a dozen rounds of "a bit faster", "less bounce", "no, the other one". A data-backed prompt usually needs one or two. Across a project with ten animated components, that adds up to hours.
It also changes what you learn. When the model works from real values, you can read its code and see exactly how a spring or a stagger is put together, instead of reverse-engineering its guesses.
Try it now. Open capturemotion.dev, pick a card, customize it, and paste its LLM prompt into your favourite model. Copying the LLM prompt from the gallery is unlimited, and you do not need an account for it.
Questions
Does the LLM prompt work with any model?
It is written for Claude and ChatGPT. It is plain text, so you can paste it into any chat model, but results depend on the model.
Why does the model change my timings?
Models like to round numbers and swap custom curves for keywords. Add "keep the timings as given" to each follow-up, and check the code against the JSON.
