A viability prototype for prompt-driven 3D storytelling: a world is a JSON document, cameras are objects, and takes are real video files — all in the browser with three.js. No build step, no backend, no AI (yet).
The whole world lives in the SCENE object at the top of main.js. Later,
an LLM writes that object; today it's hand-written.
npm install
npm start # serves on http://localhost:8642
| input | action |
|---|---|
| WASD (+ shift) | walk (run) |
| right-drag | look around |
| left-drag empty space | look around |
| left-drag object | move it (red = invalid spot, snaps back) |
| middle-click | placement menu (works on ground and walls) |
| Q / E (shift) | rotate selected/held object (15° / 5°) |
| scroll wheel | raise/lower selected/held object (stacking works) |
| G | toggle grid snapping |
| T | third person |
| Del | remove selected object |
| 1–9 | switch camera |
| R | record a take (.webm, listed bottom-left) |
Headless end-to-end check (placement, red-deny, cancel, move, rotate, elevate, wall mounting, camera add, take recording):
npm run verify
Real: world-as-data, collision registry shared by placement + player + actor steering, wall mounting with surface normals, height-aware validity (stacking), in-browser take recording.
Faked: the actor is a capsule with a bob (no rig), no physics engine (clamps and push-out), no real reflections, no audio.