Motion · reference · 08-08-2026
A prompt that writes prompts. You do not hand it to Claude to get a scene — you hand it a goal, and it writes the full brief that builds the scene. One layer above the recreation prompts we already keep, which are single finished briefs. This one manufactures them.
Three.js r160+, React Three Fiber, Drei, WebGPU/TSL, GSAP.
Scene graph, shaders, lighting, physics, performance, XR.
System role as 3D architect, tool calls for assets, Tripo/GLTF pipelines, hard limits on frame rate, mobile and code quality.
Good for. An interactive viewer, a particle hero, a scroll-driven scene, a game-style interface — anything the browser must render live and let a person turn.
Not for. Pictures. It never produces an image, a diagram or an illustration; it produces code that draws in real time. If the need is a still, the road is
the image models, not this.
One. Name the goal in one sentence, in the product's own words, never in 3D vocabulary. "The three parts of the system, turnable" — not "a rotating scene graph".
Two. Read the brief before running it. The brief is where the house constraints get inserted; a brief without them produces an agency demo.
Three. Insert our constraints, every time. Light theme — the meta-prompt's world is black by default and ours is not. The product's palette. prefers-reduced-motion landing on a still, composed frame rather than a frozen accident. A frame-rate floor and a mobile budget written as numbers. And the page still usable if WebGL fails.
Four. Judge it against the thing it replaces. A 3D scene costs weight, battery and attention. It earns its place only if it says the idea faster than a still image would.
It is heavy. Three.js plus a scene is hundreds of kilobytes before a single asset, on a page whose job is to open fast in a meeting room.
It can read as a demo. Real-time 3D signals technology company. For a public-service tool shown to elected officials that can cost the trust the page exists to build — the same reason we refused the black, video-heavy look in
the motionsites analysis.
There is a middle road. Depth, parallax and a turnable feel are reachable in plain SVG and CSS at a fraction of the weight, and
the recipes already cover them. Take that road first; reach for real 3D when the thing genuinely has to be turned by hand.