Free starter prompt
This is the first prompt we give a coding agent. It builds the skeleton the whole show runs on: a 3D scene drawn from the time alone, and a command that renders it to a vertical MP4.
Build the skeleton of a deterministic 3D reel renderer in this folder. RULE: every pixel is a function of time. One function renderAt(t) draws the whole frame from t (seconds) and constant data only. No timers, no requestAnimationFrame as a source of truth, no Date.now(), no unseeded Math.random(), no state carried between frames. 1. index.html: a 1080x1920 canvas rendered with Three.js (vendor three.module.js locally, no CDN at render time). Use an import map for "three" and "three/addons/". 2. engine/main.js: builds the scene once (async, after fonts load), then exposes renderAt(t). Pin renderer size and pixel ratio (1). ACES filmic tone mapping, sRGB output, PCF soft shadows. 3. Post-processing with EffectComposer: RenderPass -> BokehPass (focus distance set per frame) -> a small ShaderPass for vignette + a STATIC film grain (the same noise every frame, so the video still compresses) -> OutputPass. Use a multisampled render target (samples: 4) so edges stay clean. 4. Rendering: use HyperFrames (npx hyperframes init / preview / render). Set data-duration on the root element, mark it data-no-timeline, and listen for the "hf-seek" event to call renderAt(event.detail.time). If you don't use HyperFrames, write a Puppeteer script that calls renderAt(frame/30), screenshots each frame, and pipes them to ffmpeg (libx264, yuv420p, CRF 18, 30 fps). 5. A test scene: a box on a floor, one key light, one fill light, camera slowly pushing in over 5 seconds. Render it and show me a frame from the start, the middle and the end.
The guide has the other six prompts (characters, the set and lighting, AI voices with lip-sync, performance, the auto-director and solid props), the episode formula, and fixes for every problem we hit. The Studio is the finished project, ready to render.
See the guide and the Studio →