Only said make the model, and Claude delivered a dismantlable mechanical octopus

Fokki posted a demo on X: give Claude Opus 5.5 an image and just say make the model, and the model returns a robot octopus that can be taken apart in the browser. The video is 25 seconds long, and the code is 2,616 lines.

On the left side of the frame is the Pelagos OC-8 data panel, with thruster pressure, battery level, and arm curl status all flickering. On the right, the mechanical octopus is made up of petal-shaped shells, lenses, and linkages. The output is not a pre-rendered video; every part is built with three.js code. three.js is a 3D library for the browser; ordinary readers can think of this passage as a model in a webpage that can be rotated, disassembled, and posed in real time.

Fokki documented the process in great detail. The reference image first establishes the appearance, with petals, lenses, and connectors aligned one by one. The model was not animated directly; first a still frame was rendered side by side with the original image, and three rounds of seam fixes were done before moving on to the later 25-second short. The first version looked like a toy, so it went back and packed 202 linkages and 480 wheels into 8 arms. Each arm submits only one draw call, and each frame calculates its pose along the bending chain. A draw call can be roughly understood as a drawing batch the renderer has to process each frame; fewer of them means more stable real-time interaction.

The animation rhythm is also clear. Every 2 to 3 seconds, a part opens: the lens rotates open, the outer shell's hinges spread, and the pump rises while rotating. At second 16, the entire octopus spreads out on a 330° turntable, then closes back up before the turntable finishes rotating. At the end of 25 seconds, it walks out of frame on its own.

What is worth watching in this demo is that the model first does still-frame validation before touching animation. Many generative 3D demos die at the first step: they have a shape from a distance, but the seams don't line up up close, and they fall apart as soon as they move. In Fokki's record, the three rounds of seam fixes happened before rendering the short video; this order is very much like a human 3D engineering workflow. Someone in the comments noticed this, saying that the still-frame fixing step is slow work no one sees. Someone else closed Blender after watching, and others asked how many hours the rendering took.

Some netizens were reminded of the Sentinel from The Matrix. The video is below.

Fokki put the skill, prompt, and live sim in the original tweet, and Elvar also wrote an article explaining this skill. For developers, they can take it apart to look at the three.js hierarchy, draw call control, and per-frame poses. For non-technical readers, this is a complete example of an image becoming an interactive mechanical model. This time Claude Opus 5.5 did not just provide a beautiful still frame; it delivered the whole process of being dismantlable, movable, and self-checking together.

Original link: https://x.com/0x_fokki/status/2116289885870981296

Elvar article: https://x.com/i/article/2116115649856430080