Someone plans to uninstall Blender tonight because one image can make Claude write a disassemblable 3D machine

Elvar left a line on X/Twitter: he plans to uninstall Blender tonight. The reason has nothing to do with Blender itself. A developer released a free Claude skill that can turn an image into a 3D machine that can be taken apart.

Let's make the background clear first. Blender is a common open-source 3D modeling program. In the traditional workflow, you build meshes, rig skeletons, adjust materials, set up lighting, and then render animation. This skill takes another path: it does not build a model; it has Claude write every shell panel, every gear, and every hose as code using three.js. three.js is a JavaScript library that runs 3D in the browser; the model is generated by code, not imported from a file.

The person behind this is Nyxera. He did not write a single prompt. He wrote a job description. The file specifies the appearance, camera movement, and checklist, plus one hard rule: do not shoot the video until I give the nod.

He used a render generated by ChatGPT as a reference and gave only one word, "octopus," to see where it would break. It did not break.

Claude treated that image only as a target; all shells, wheels, and lenses were written by itself in three.js. The eye opens into a 10-blade aperture hidden behind a 62 mm light ring. The top dome is 4 mm borosilicate glass and can be lifted and flipped in the air. Each arm has two rows of suction-cup wheels, running all the way to a 16-segment spiral tip. Eight arms plug into a 48-tooth hub, and the moment it opens, 16 hoses spread out. One part falls off every 3 seconds, and by the 19th second, nothing is still attached.

More surprisingly, the octopus is not one of the 3 animals included with the skill. The built-in ones are tarantula, piranha, and wasp.

In the article, Nyxera breaks the workflow into six parts: reference image, job description, checklist, still frame, page, and video. It sounds like project management, and it is. The reference image only defines the appearance; Claude does not import it, only reads it. The job description is read once before each start of work. The checklist is specific to this one image, 10 to 12 items, and every item can be checked by the reader.

The key is in the wording. Item 5 on his tarantula checklist says: the abdomen is a white curved frame holding transparent glass panels, with green coils glowing from inside; it must absolutely not be one solid white shell. "Must absolutely not" is written on purpose, because a white shell is the default value, and if you do not lock it down, it slides back. Item 7 said "fine black spines on the legs," and the bristles came back like porcupine quills, twice as long. Later it was changed to a number: 0.15 to 0.25 times the thickness of the leg.

He says that every correction after that was a sentence with a number in it.

There is another counterintuitive point: do not rely on your eyes to judge whether it is "dense enough." The moth, hydra, and fly all self-assessed their density as passing, yet they looked like clean white images. The moth went through 9 rounds and was thrown out. Now Claude counts: crop the reference image and the still frame to the object's bounding box, count the proportion of pixels with brightness below 80, and anything below 0.8 of the reference is considered failing. The hydra measured 0.57, and one round targeting that number brought it to 0.80. The fly reached 0.96 after real gaps were cut into the white panels.

Proportions are the same. The piranha passed density in the second round but looked like a crocodile: the lower jaw took up 0.52 of body length, while in the reference image it was 0.37. In the third round, four numbers were measured from the image, and only then did it land at 0.35.

Nyxera tried running four at once and only looked at the results in round 6. The moth alone burned about 430,000 tokens before it was thrown out. One still frame is about 400,000 tokens, and the video adds another 150,000 to 400,000. Looking late is expensive. Now he runs two at a time and checks at round 2, while one note can still change the result.

The skill is free and open source on GitHub. The installation command is one line:

npx skills add Nyxera0x/cutaway -g -a claude-code

It requires Claude Code on Opus 5.5, ChatGPT for generating reference images, Node.js 20 or later, ffmpeg, and Python 3 for video and sound. Claude Code can help you install it.

What is it good for? Nyxera says the same file can be used on anything with a dense mechanical core: engines, drones, cameras, watches. It is not suitable for objects with smooth surfaces and few joints; long white shells look like toys no matter how you fix them. Creatures should ideally have only one set of main limbs; tentacles and wings competing for the same second become noise.

Blender is of course still useful. If you really need production-grade assets, complex rigging, or film-quality rendering, the traditional workflow is still there. But this skill pushes the cost of "explanatory cutaway animation" down to another order of magnitude. For product demos, science videos, and concept explanations, you do not need to build a real model; one image plus a clearly written job description is enough to let Claude write the mechanical logic as code and take it apart for you.

The octopus is cool, but what is really worth looking at is that job description. It breaks "getting AI to do it right" into checkable items, measurable numbers, and corrections that can be written back. When the model changes generations, this method is still there.

GitHub repository: https://github.com/Nyxera0x/cutaway