One Image, 300 Parts: A Free Claude Skill Turns a Static Image Into a Mechanical Scorpion That Can Pry Off Its Shell
No modeling software, no manual rigging. Kitsui uses a single image to generate a three.js mechanical scorpion with 8 legs, 5 tail segments, and 300 parts; the pincers pry off the shell, and the stinger rotates 360°. Someone in the comments points out the key: a single-input constraint.
X user Kitsui posted a video: an image of a mechanical scorpion was turned by a free Claude skill into a model that can move in three.js. In the video, the scorpion is not looping a showcase animation; it has legs, tail segments, and back plates, and the near-side pincer also pries the outer shell open along a gear train.
The numbers given in the post: 8 legs, 5 tail segments, 7 back plates, with 300 small parts packed in between. The appearance is set by a single image, while the model itself is in three.js. Opus first posed its own still frame to match the pose in the image, then iterated until the two images placed side by side showed no visible seam.
The motion details are denser. The 8 legs run in, and the near-side pincer pries the shell open along a 6:1 gear train and a 180 N grip-force pushrod. The tail connects segment by segment, the stinger blade unlocks and rotates a full 360°. The back plates fan out, and the glass cover swings upward 60°, enclosing a core driven in rotation by 4 gears. At 0:11, it spreads out on a turntable, resets while still rotating, and walks straight toward the camera for a strike.
Kitsui says the skill and the complete workflow are in Elvar's article, for free. Someone in the comments replied, 'Already using it, no problem,' and another said, 'Free skills beating paid courses is the theme of this year.' That remark is a bit snarky, but against the post's line that 'some paid courses still don't do as much as this free thing that was posted,' it isn't entirely unfounded.
What's really worth watching may not be whether it 'looks like it.' X user ReignSwords points out that the key is the single-input constraint: one function drives the machine, the gauges, and the timeline at the same time, so the parts don't each rotate on their own. In conventional AI 3D demos, every component can be animated separately, and when assembled they fall apart. Here the linkage relationships are locked down first, and then the question is how many moving parts the model can support. Leo asked in the quote thread whether it still holds when there are many moving parts. That question is more useful than 'wow.'
The quoted tweet also includes an exploded-view demo of a revolver, with the data for the cylinder, trigger, and hammer following the motion. Someone commented that this would work well as teaching material; seeing how a machine moves is more direct than reading text. That judgment is not exaggerated. The hard part of mechanical principles is often timing and constraints, not static appearance.
Kitsui's exact words were that they have not yet encountered an image that cannot be turned into a working machine. For now, this conclusion can only be treated as a personal experience. Only with more samples—especially machines with more than 50 moving parts—will we know where the limits of the single-input constraint lie. The free skill has already lowered the barrier by a notch; what comes next is whether it can reproduce results reliably, rather than just looking good in one scorpion video.
发布时间: 2026-11-03 16:48