Claude Code teardown skill: 3D disassembly from one image

The Claude Code teardown skill turns one reference image into a 3D disassembly simulator written in three.js code. What it builds, what it needs, and where it…
Hand one reference image to Claude Opus 5.5 and you get a 3D simulator of a robot bee that comes apart piece by piece. The whole procedure is packed into a single skill called teardown, which runs on Claude Code and is free under the MIT license. This article covers what the teardown skill actually produces, what you need to run it, and how far you can take product explainers and training material in-house.
What the Claude Code teardown skill produces
The example is a robot bee. From one reference image it becomes a 3D simulator you can disassemble part by part. The shell, the gears, the hoses are all emitted as three.js code rather than assembled from off-the-shelf 3D models. Under the large gear sit 260 smaller components. When it plays, some part detaches every two or three seconds. The author built seven of these in two days and condensed the process into one skill.
Two things stood out to me.
- The output is code. It is not a binary 3D asset, so you can read the diff and review it.
- The disassembly has a timeline. It is not a frozen exploded diagram but an explanation with an order to it.
Together those two change what a 3D asset is. It stops being a deliverable someone hands you and becomes a document you can edit. Reorder the explanation, fix a part name, recolor one piece — if those are code edits, the asset is something you maintain rather than something you commission once.
One reference image decides the result
The author’s own note is the most useful detail: without a reference image, the output ends up looking like a generic toy. That means the lever on quality is not a long prompt. It is a single image.
So the working order is simple.
- Pick one image where the structure of the object is visible.
- When the result is off, swap the image before you add more words.
- Then read the generated code and correct part names and disassembly order.
A weak image gives a weak result. The flip side is that if you already have product photos or diagrams sitting in a shared drive, that archive is your input quality. Whether you need to reshoot anything is answered by trying one image. Since the cost of an attempt is low, spending your time on more attempts beats agonizing over which photo is perfect.
What 260 parts and a part every few seconds tell you
260 small components under one large gear says this is not a shell that merely resembles the object. It also raises a question you have to answer yourself: how fine does your own teardown need to be?
- To teach assembly or maintenance, break it down to the units a person actually touches.
- To convey structure, stop at functional clusters — more granularity usually means less understanding.
- For an exhibition or a sales deck, pacing and visual appeal matter more than part count.
The two-to-three-second rhythm matters too. It is a pace you can keep watching, and the order of disassembly becomes the order of the explanation. Which means the real decision in this workflow is not modeling — it is what comes off, and in what sequence. Only someone who knows the product can decide that, and once it is decided, the rest is producible.
Seven builds in two days is the other number worth reading. The shorter one build takes, the more freely you can throw away an attempt. Commissioned work pushes you toward refining a single version; cheap rebuilds let you put several versions side by side and pick. That changes how asset production is run, not just how it looks.
Running the teardown skill: requirements
The requirements are clear.
- Claude Code
- Node.js 20 or later
- ffmpeg and Python 3 if you want video output
If video can wait, Claude Code and Node.js are enough to start. Get something running on screen, decide whether it is useful, and add the video toolchain afterward.
The skill ships with three worked examples: a V8 engine, a Cessna landing gear, and a Strandbeest. Running one of those first is safer than starting with your own product. A single run tells you three things at once — whether your environment is set up correctly, how dense the output gets, and whether this style of teardown suits your use case. And because it is a free MIT-licensed skill, the internal paperwork before an evaluation is light.
Why this lands in Japan: product explainers and training material
Japan has a lot of products that need explaining — machinery, factory equipment, components, medical devices. 3D explainer assets have traditionally been commissioned from specialist studios, and products without that budget made do with flat diagrams and text. A workflow like teardown reaches exactly the side that never got the budget.
Uses that come to mind:
- Trade shows and online meetings where you cannot bring the physical unit
- Training material for new hires and partner contractors
- Product explanations handed to distributors and resellers
- Briefing candidates who need to grasp the structure before they join
It applies to overseas AI and tech companies selling into Japan as well. In my experience Japanese buyers want to understand how a thing works before they approve it. An explanation that keeps the mechanism hidden may sound impressive and still fail to clear internal review. A teardown is a format that answers “show me the inside” head-on, and paired with Japanese-language narration it carries further than a translated deck.
There is a caveat. A 3D model generated from code is not the physical object. Where dimensional or specification accuracy is demanded, keep explainer assets and engineering documentation separate. Decide upfront which one you are making, and the conversation does not stall on precision.
How to put this to work
Start small. Run one of the bundled examples and look at the density of the output. Then build one model from a single image of your own product. Show it to someone who knows that product and check one thing only: whether the part names and the order they come off hold up as an explanation. If they do not, change the image and build another.
The test is not how beautiful it looks. It is whether the disassembly order is the explanation order. Get that right and it is usable material. Get it wrong and you rebuild. That cheapness of rebuilding is, to me, the most important thing this example demonstrates.
FAQ
Is the teardown skill free to use?
Yes. It is published as a free skill under the MIT license and runs on Claude Code.
What do I need to run the teardown skill?
Claude Code and Node.js 20 or later. If you want to turn the result into video, you also need ffmpeg and Python 3.
Does it work without a reference image?
It still produces something, but the author says the result looks like a generic toy without one. A single reference image is what sets the quality.
Based on a post on XOriginal post ↗