Fifteen motion-design films, and not a single AI-generated pixel

What it does

This is a showcase of fifteen 10-second motion-design films, each in a distinct style — cel animation with line boil, isometric 2.5D, Bauhaus geometry, synthwave VHS, pixel art, and so on. Every film was produced by Claude Opus 5.5 writing code from a single prompt: the visuals are SVG, Canvas, WebGL and Three.js (Blender for the one 3D film), and each film’s audio.py synthesises its own music and sound effects from 103 recorded instrument samples. Crucially, there’s no image or video generation anywhere — the picture is deterministic code, frame by frame.

The interesting bit

The architecture is the quiet star: each film is a web page exposing window.renderAt(t), so any moment is computable. Headless Chrome captures frames, ffmpeg muxes an exact 10-second MP4, and a second AI reviewed each film against rubric.md for two rounds of revision (final scores 7.57–8.21/10). The prompts themselves are published in full, structured around output spec, technical route, creative seed, and “signature features” — the part that makes a style recognisable.

Key highlights

• 15 styles, each with film, style notes, full prompt, and complete source published
• Sound is code too: audio.py synthesises music and effects on the same cues as the picture
• Reusable harness/render.mjs renderer and lib/audio soundtrack library you can point your own AI at
• template.md lets you write a new style by filling in four sections
• Bilingual showcase page (English/Chinese) with all films playable with sound

Caveats

• Fonts aren’t in the repo — you source them separately from assets/fonts/README.md
• Rendering needs Node 22.12+, Chrome, ffmpeg, and Python 3.11+; the 3D film additionally needs Blender
• The renderer finds Chrome in its default macOS location; other platforms require pointing it at the binary

Verdict

Worth a look for anyone curious about AI-driven creative pipelines, prompt engineering for visual output, or code-as-motion-design. If you wanted a generative video model, this is deliberately the opposite — and that’s the point.