I Wanted Claude Motion, So We Built Our Own

This morning was a playground session. No client, no deadline, just one question: how far can we actually push animated video with AI right now?

The obvious place to start was Claude Motion, which Anthropic launched in beta at the start of October. You type /motion in a Claude chat, Claude writes code that animates your text, charts, shapes and images, and you export an MP4. No video model, so no wobbly hands or melting faces. Just code.

That is very close to how we already make our videos, so I wanted to see it.

Where Motion lives, and why that mattered

I do most of my building in Claude Code, in a terminal, inside the actual project. Motion lives in the claude.ai chat app, and it is switched on for Team and Enterprise plans. From where I was working there was simply no button to press.

So I asked Claude to build the same kind of thing itself, with the tools we already had: a web page that can draw any moment of the film, a headless browser that steps through it 30 frames a second, ffmpeg to stitch the frames together, and Python to synthesise the sound. Then we added three.js, a 3D library for the browser, and that changed everything.

It is the same idea I wrote about in let AI do the hard work, use determinism for consistency. AI writes the thing that makes the video. Nothing in the pipeline guesses. Render it twice and you get the same film twice.

What Claude nailed

The finished piece is one unbroken zoom out from inside an atom to the large scale structure of the universe:

• an atom with glowing electron trails
• a DNA helix built from thousands of spheres
• a coiled chromosome and a translucent cell
• a hologram of a person, a city at night with moving traffic
• Earth with continents, clouds and city lights on the night side
• the solar system, the nearest stars, a 110,000 star galaxy, and the cosmic web

The scale is real. The planets sit at their true distances from the Sun. Alpha Centauri is 4.37 light years out. Our Sun sits about 26,000 light years from the middle of the galaxy, so the galaxy is drawn off centre on purpose. A counter at the top rolls through every power of ten in time with the camera.

It renders 750 frames in about four minutes on a normal laptop, with no cloud service and nothing to pay per render. Change one word, render again, and only that word changes.

Where it went wrong, and what I had to correct

This is the bit that matters, because it did not get there on its own.

It looked like Minecraft. The first 3D test, a golf hole, was built from flat shaded triangles with cones for trees. Technically impressive, visually a video game from 2011. When I said so, it rebuilt the course with smooth shading, a painted texture with mowing stripes and sand bunkers, rounded trees, a golden hour sky and proper shadows. The difference was enormous, but I had to ask.

It went ahead instead of answering me. I asked a string of questions about Motion and Remotion, and it got on with building things instead. Useful work, wrong order. I want my questions answered first, then the work.

It kept telling me to go and do things. More than once it handed me a prompt and asked me to go and run Motion and send it the result. I wanted it to build the thing, not set me homework.

It was lazy with the look. The first version of the zoom used exactly the colours, font and sound effects from my golf app videos. It was flat 2D too, after we had just proved 3D worked. I called it out. The rebuild is real 3D, near black with a different accent colour at every scale, a new typeface, and a completely new soundtrack: a drone that opens up as you zoom out, a riser into every level, a sub bass drop on each arrival.

Every big jump in quality came from me saying "that is not good enough", not from the first attempt.

What I would tell anyone trying this

1. You may not need the shiny new feature. Motion is a nice idea. The same approach, built yourself, can go a lot further: 3D, sound, any format.
2. Name the problem bluntly. "This looks like Minecraft" got a better result than any amount of polite feedback.
3. Ask for its own identity. AI will happily reuse whatever worked last time. If a piece is meant to stand alone, say so.
4. Make it do the work. If it starts handing you tasks, hand them straight back.

If you want animated explainers for your own product, made this way, get in touch. You can also see more of what we build on our work page, or read how we keep AI video from looking fake in this post.