Build 3D
3D is earned, never a default. First ask whether a flat technique does the job
(~/agent-skills/shared/creative/adaptation.md); 3D fits product configurators, spatial
data, games, art and brand heroes where depth carries meaning.
Steps
- Engine: follow
~/agent-skills/shared/creative/selector.md. One WebGL engine per page: OGL for a single shader plane, three.js (or R3F in React) for real scenes, PixiJS for 2D. Reuse what the project has. Assets (models, HDRIs, textures) must be owned or licensed; note the source. - Technique files (load only what you use) from
~/agent-skills/shared/creative/README.md:shared/creative/techniques/shader-distortion.mdshared/creative/techniques/fluid-particles.mdshared/creative/techniques/environmental-lighting.mdshared/creative/techniques/depth-parallax.mdshared/creative/techniques/scene-transitions.md
- Budget from the start: DPR capped at 2, canvas sized to its element, render loop
paused by IntersectionObserver and
visibilitychange, stopped when the scene is at rest,webglcontextlosthandled, geometry instanced, textures compressed. A still poster renders first; the live scene upgrades after load. The heavy-hero-performance workflow inbuild-animationhas the frame governor. - Accessible: the canvas is
aria-hiddenor labelled; every action in the scene has a DOM control; content never lives only inside the canvas. Reduced motion shows a still frame or a static camera. - Delegate heavy scene code to Codex in a worktree
(
~/agent-skills/shared/creative/codex-orchestration.md); review the result yourself. - Verify with
~/agent-skills/shared/creative/visual-loop.md: screenshots in both themes and three viewports, WebGL-disabled fallback, and theaudit-designrunner clean on offscreen-render, idle-render, long frames and reduced motion. Use a seeded or virtual clock so screenshots are comparable.
Report
Engine and why, asset sources, frame and idle measurements, screenshot paths, and fallbacks tested.