One Prompt per Key: He Used Opus 5.5 to Bring His Room into three.js

Meng To posted a video on X: a room with blinds shadows, camera movement, and a close-up of a keyboard. The room in the video is not live-action footage. Meng To had Opus 5.5 rebuild his own room as a three.js scene, to animate a tutorial video.

three.js is a library for running 3D in the browser; Blender is a common 3D modeling program. For readers unfamiliar with 3D, this one sentence is enough to understand the choices that follow.

Someone asked how many prompts the keyboard took. Meng To replied: one per key. But the execution is done by AI, so even he doesn't know exactly how it was implemented. This detail is quite interesting. The control granularity sounds very fine, but the actual process is still a black-box loop.

Someone else asked whether the tutorial animation runs live from the three.js scene or is rendered into clips. In another reply, Meng To explained the tool choice: he deliberately used three.js instead of Blender because Blender is time-consuming, makes the laptop heat up, and can't multitask.

The room isn't a one-off set. Meng To says it's a playground where you can change lighting and move the camera, like having a new basement studio. Tutorials can reuse the same scene over and over, with the camera and lighting serving the explanation, without rebuilding it each time.

Someone noticed the blinds shadows on the wall and asked whether they were raymarching or baked. Meng To replied: not sure about raymarching, but they are part of the lighting setup. That answer is quite telling. He doesn't chase rendering terminology; as long as the lighting setup produces the effect, that's enough.

More important is the iteration method. Someone asked whether Opus compares its own rendered screenshots with photos of the room, or relies only on text feedback. Meng To replied: it loops for hours, and that's how it is for every object.

This is closer to the actual workflow than generating a room from a single sentence. The prompts can be as granular as one per key, yet the execution process is corrected in loops by AI. The author can control the references and the result, but may not be able to explain every step. For people making tutorials, the value isn't in showing off either. A reusable 3D set saves the time spent repeatedly waiting for renders and building scenes, letting the camera, lighting, and keyboard close-ups directly serve the explanation. The video looks like the result, but the replies are where the method is.

Original link: Meng To's tweet