One image generates 460 parts: This Claude skill turned a jumping spider into a take-apart mechanical pet

Elvar posted a video on X showing the mechanical spider his Claude skill built from an image of a robotic jumping spider. The skill mentioned here can be understood as a reusable task configuration consisting of prompts and an execution workflow.

In the footage, the spider has a white and orange shell, eight mechanical legs, and sensors and optical equipment visible under a transparent head dome. On the left are labeled gait frequency 2.7 Hz, jump height 0 mm, leg servo load, and operating mode; on the right are the optical system, assembled components, display, and subsystem status.

Elvar himself said he is starting to be a little afraid of his Claude skill. He has been testing its limits. Today he gave it an image of a robotic jumping spider, and the skill made it like this: the image only defines the appearance; all 460 parts are generated by code, including the pistons and springs in the glass lower legs. The motion is written as math first, before anything is drawn: the 8 legs are split into two groups of 4, each first sinks down, then extends in 0.12 seconds to launch. The gait chart, servo bars, and jump meter read from the same set of math, so they don't contradict each other. The irises of the two large eyes contract with every jump. With one click it explodes apart on a turntable, then closes back up and rushes toward you.

In the comments, Joachim hit the key point: most AI demos draw the visuals first and then fill in numbers for the gauges; Elvar wrote the motion as math first, so the gauges don't fight each other. Specter pressed further: from one image to 460 parts, how many attempts did it take, and how much was fixed by hand. Auren said the ones who should be afraid are 3D model developers. Bruno liked the explode-and-reassemble part and said it was pure showmanship. Fahrenheit said this no longer feels like a skill; it feels more like a pet.

The highlight here isn't the grand phrase 'AI can generate models', but the fact that appearance, motion, and gauge data are tied to the same set of math. In many demos, gauges are just decoration, and the data has nothing to do with the action; here the gait chart, servo bars, and jump meter are all derived from the same motion parameters, so you don't get the disconnect of legs jumping while the meters drift. For developers, the key to reproducing this is to define state and parameters first and let the visuals follow. For ordinary users, it looks like a toy, but behind it are code generation and system design.

Elvar put the skill and prompts in an article, and the original post also includes the video.