Generating a full 3D torus motion piece with Claude-written code, no video model
The author had Claude break down a reference video into a spec and write it as code, so the torus, camera, panels and subtitles are all adjustable and a one-line change re-renders, with no After Effects involved.
no video model made this every frame is code. Claude wrote the 3D torus, the camera flying through it, all the panels, the code typing itself, the log, the timing. I didn’t touch After Effects once here’s how it actually got made: > I showed Claude a reference video instead of describing a style. it broke down why it works: the pacing, where the first motion lands, how the camera moves, what the panels do > that turned into a spec, not a vibe. “torus in the center, camera pushes in, dives through the hole, pulls back out, panels stay alive the whole time” > Claude wrote it as code. the particles, the camera path, the easing, the 128 BPM beat the cuts land on then I directed it in plain words: “title is a beat late” titleIn 0.2 0.0 “draw the line slower” chartDraw 1.0 1.5 “hold one more second” holdEnd 0.0 1.0 > each one was a one-line change. not a new render roll. not another prompt and a prayer > export. same input, same frames, every time that’s the part people miss about AI video. > a video model gives you a clip you can’t edit. wrong number on screen? re-roll. timing off? re-roll. you’re negotiating with a slot machine > when the video is code, every number is a variable, every timing is editable, and a correction takes seconds instead of an hour you’re not generating video anymore. you’re compiling it I wrote up the whole studio in the article: the 3 rooms: data, motion, edit the 10 open-source skills that teach Claude motion design the workflow from reference to finished clip the mistakes that waste a day a first-week plan if you’re starting from zero
This post quotes another post: x.com/francescoinweb3/status/2108882188624121896