A one-minute cartoon with no video AI — every frame is code

Here's a fun one. Usually an "AI cartoon" means a video model: you write a prompt, wait, get a clip. Then in shot three your hero is suddenly wearing a different jacket. And you can't fix it — only regenerate.

The idea in one line: type "make a short cartoon about the bridge that collapsed under marching soldiers" — and get a minute of vertical animation where every frame is code. Not a picture from a black box, but a scene you can open and edit.

Why this only just got easy

On November 1, 2026 a free skill called Claude Animated Shorts (MIT) came out. It makes 55–70 second shorts in a "flat 2D characters in a colourful 3D town" style: ink outlines, blurred backgrounds, smooth camera glides, narration, music and sound effects.

The key part: there's no video model inside. The world and characters are built in three.js, and the clip renders through Remotion on your own machine. The agent writes the script, times shots to the narrator's words, renders stills, compares them to a reference and fixes them. Even the sound is synthesized in code: effects plus a march ducked under the voice.

The example in the repo is a 67-second short about the 1831 Broughton bridge: 74 soldiers marched in step, and the bridge gave way. The skill passed 360 stars in its first day.

What you'll learn

• That video can be described in code: scene, camera, light are just variables.
• How Remotion works: a React component rendered to mp4 frame by frame.
• How to sync picture to voice: every on-screen reaction lands 0–3 frames before its word.
• How to work with an agent in a "still → compare → fix" loop instead of one prompt.

Starter prompt

Copy the skill folder into ~/.claude/skills/cartoon-3d-short/ and install Node, Python 3, ffmpeg and Remotion with @remotion/three. Then precision decides everything. "Make a cartoon" gets you something average. Give the agent a story, a length and a voice rule:

Don't like a shot? Say it in words: "the builder's hat blends into the sky — make it dark", "the camera leaves the cat too fast".

What you'll get

A 1080×1920 (or 2K) vertical clip: a painted little town, bean-shaped characters with outlines, the camera gliding around the tower, a narrator telling the story over quiet music. It looks like one of those science-cartoon channels.

And unlike a video model — you own the source. The hero is the same in every shot because it's literally the same object. Want a different sky colour — change a line and render again.

Honest limits. You need your own voice: record or generate voice.wav, plus word timings from any transcription. The agent loops burn a lot of subscription tokens. Rendering 67 seconds in 2K takes about 10 minutes. And disk: every build copies ~1.3 GB into a temp folder, so point REMOTION_TMP at a big drive right away.

A year ago an "AI cartoon" meant a lottery with a video model. Now AI writes the cartoon in code — and you can edit it like text.

Source: Claude Animated Shorts — a free skill for cartoons in three.js and Remotion (GitHub, mromidev)