How-to

Recreate a website animation with Claude or ChatGPT

Give Claude or ChatGPT the real timings instead of a vague description. What is in the Capture Motion LLM prompt, how to use it, and how to check the result.

· 5 min read

Ask a model to "make a card that slides in smoothly, like the one on that site I saw" and you will get a card that slides in. It will use ease-in-out, a round 300 ms, and nothing else. The result is fine and forgettable, because the model is guessing at exactly the parts that made the original feel good: the duration, the delays between parts, and the easing curve.

The fix is to stop describing the animation and give the model the data. This guide shows how to use the Capture Motion LLM prompt with Claude or ChatGPT, what goes into it, how to steer the result, and how to check that the timings survived.

Where the prompt comes from

There are two places to get it, depending on your source.

Copying needs a free account. Capturing and previewing in the extension do not.

What is in the prompt

It helps to know what the model is reading, because that tells you what it can and cannot get right.

PartWhat it gives the model
A ready-made instructionTells the model what to build and to use the captured values
The capture JSONEach animated property, with its keyframes, durations, delays and easing
Reference markup (gallery cards)The card's own markup, as a visual reference for structure and look

The JSON is the motion. The reference markup is the look. A capture from the extension carries the motion data, so the model rebuilds the movement faithfully but has to invent the visual design around it. A gallery card prompt carries both, so the model can rebuild it closely.

If you customized a gallery card first, your choices carry into the prompt: colour re-tints, text swaps, speed and background. Set those before you copy and you skip a round of follow-ups.

Step by step

  1. Get the cleanest source you can. For a capture, pick the smallest container that holds the whole motion and close any banners first. Stray tracks in the JSON become stray animations in the code.
  2. Copy the prompt. Use Copy > LLM prompt on a gallery card, or Copy for LLM in the extension.
  3. Paste it into Claude or ChatGPT as the first message of a new chat. Do not trim it; the instruction and the data belong together.
  4. Add one line about your stack before you send, if it matters. For example: "Build this as a React component with Tailwind."
  5. Run what comes back. The model returns a self-contained HTML file, or a component if you asked for one. Save the HTML and open it in a browser, or drop the component into your project.
  6. Check the timings against the JSON (see below) before you start changing anything.
  7. Then make it yours. Ask for your own content and colours in a follow-up.
Order matters: get a faithful copy first, then change the content. If you ask for a faithful copy and a redesign in one message, the model tends to drift on the timings while it focuses on the redesign.

Good follow-up prompts

The first answer is a starting point. These follow-ups tend to work well, and each one keeps the motion intact while changing something else.

Notice the repeated 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 the timings

This is the step most people skip, and it is the one that decides whether the result feels like the original. It takes two minutes.

  1. Open the JSON you pasted next to the model's code.
  2. Match each track. For every animated element, find its duration, delay and easing in the code and compare.
  3. Look hardest at the easing. A cubic-bezier() or linear() value should come through unchanged. If the model swapped it for a keyword, the motion will feel flatter. Our guide on finding the easing curve of an animation explains why the curve matters so much.
  4. Check the delays. Staggers are easy to break. If ten items should start 40 ms apart, count them.
  5. Watch it side by side. Play the original (the gallery card, or the extension preview) next to the new file. Differences in feel usually point straight to a changed number.
  6. Correct in one message. List what differs and ask the model to use the given values exactly.

A small table makes the check quick:

TrackIn the JSONIn the codeOK?
Card opacity400 ms, 0 ms delay400 ms, 0 ms delayYes
Card translateY520 ms, cubic-bezier(.22, 1, .36, 1)500 ms, ease-outNo

The values above are an example of the format, not taken from a real capture.

Common problems

Where to go next

If React is your target, gallery animation to React component goes further. If a designer also needs the motion, the Figma plugin guide uses the same capture. And to keep prompts you want to reuse, save the capture in My Gallery.

Questions

What does the LLM prompt contain?

The capture JSON, with each animated property, its keyframes, durations, delays and easing. From a gallery card it also includes the card's own markup as a visual reference, so the model can rebuild it closely.

Where do I get the prompt?

From a gallery card, open View and use Copy > LLM prompt. From the Chrome extension, use Copy for LLM after a capture. Copying needs a free account.

Does it work with ChatGPT as well as Claude?

Yes. The prompt is plain text with a ready-made instruction, so you can paste it into either. The model returns a self-contained HTML file, or a component if you ask for one.

How do I know the model kept the timings?

Compare the durations, delays and easing in its code with the values in the JSON you pasted. If anything differs, ask it to use the given values exactly.