Teardown: Claude Code builds 3D machines you can take apart in your browser

What it's about

How does a four-stroke engine actually work? How does a mechanical watch keep time? The Teardown skill answers questions like these with a 3D model instead of a block of text. A skill is a bundle of instructions, scripts and checks that teaches an AI agent to do a specific task well, every time. Teardown is built for Claude Code, Anthropic's command-line coding agent. Ask it to build a machine, and you get an interactive simulator in your browser that you can operate and take apart.

The project lives on GitHub at elvarofficial/teardown. It was created on October 30, 2026 and is released under the permissive MIT license. It caught on fast. On November 1, the Japanese blog note.com counted 688 stars and 110 forks (copies of the project made by other developers). By November 6, explainx.ai put the count at around 2,176 stars. Yet the repository contains only two commits, meaning two saved snapshots of changes. In other words, the project is still very young.

What the skill produces

According to the README, the project's description file, the output is a single HTML page that runs without an internet connection. The machine runs in the middle of the page. Gauges around it show real values from the simulation model, so they aren't just decoration. Press a button and the part that's currently working opens up, a bit like a cutaway drawing in a textbook. If nobody touches the page, a guided tour starts on its own.

Teardown has two output modes:

• Live version (-live): the interactive simulator you control yourself.
• Film version (-film): a recorded tour. A script renders it frame by frame into a 60 fps MP4 video, or into contact sheets of stills.

Sound is optional. It's generated in code and mixed to a loudness of −14 LUFS, a common target for video platforms. Recorded sounds come from Freesound and are CC0-licensed, so they're free to use. For 3D rendering the skill uses Three.js, a popular JavaScript library for 3D graphics in the browser.

A post on X cited by note.com says a finished file weighs in at about 800 KB. Teardown is also supposed to be able to build a model from a single reference photo of a machine. According to note.com, neither claim appears in the README, so both are unconfirmed. The slash command mentioned there doesn't show up in the documentation either. note.com says the safer bet is to simply ask Claude in plain language.

Three sample machines and getting started

The sims/ folder contains three ready-made machines: a V8 engine, a Cessna aircraft and a Strandbeest, artist Theo Jansen's wind-powered walking machine. The demo GIFs in the README, which show a robot bee, a scorpion and an octopus, are generated examples, not machines that ship with the project.

To install it globally for Claude Code, run:

npx skills add elvarofficial/teardown -g -a claude-code

Then run npm install && npx playwright install chromium once in the install folder. Playwright is a tool for controlling a browser remotely, and Chromium is the open-source base of Google Chrome. You need Node.js 20 or later. Video requires ffmpeg, and sound requires Python 3. According to note.com, if you only want the HTML page, Node.js is all you need.

Want to see what comes out before involving Claude? You can build the examples directly, for instance with node sims/build.mjs v8-live. The file lands at teardown-out/dist/v8-live.html, and you can open it in your browser. The README suggests prompts like: "Build me a simulator of a four-stroke motorcycle engine that I can take apart," "How does a mechanical watch keep time? Show me as a simulator," or "Do a teardown of a hydraulic excavator arm and record a 20-second video of it."

How the skill is built

explainx.ai says the structure makes a good template for your own skills:

• SKILL.md is the core. It tells Claude how to build a machine, which rules keep the look consistent and how to check the result.
• sims/kit/ holds the runtime: the stage, the renderer and a step that pre-warms the shaders. Shaders are small graphics programs. Loading them ahead of time keeps the page from stuttering when it opens.
• Each machine consists of three files. A shared file defines shape and behavior, and both the live page and the film tour draw on it. That way the two versions can't drift apart.
• sims/build.mjs bundles Three.js and the fonts straight into the HTML file. That's why it runs offline and you can email it or drop it into a presentation.
• sims/capture.mjs renders images in a headless browser. This lets Claude look at its own output and fix mistakes. According to explainx.ai, this self-checking is what separates a skill from a one-off prompt.

Under the hood, Teardown builds on claude-motion, an open-source toolbox by whaleyxbt that is also MIT-licensed. According to the README, the runtime, frame capture, audio engine and validation rules all come from there. Teardown itself contributes the machine kit, the skill text and the examples.

Tips and weak spots

explainx.ai recommends naming the machine precisely ("four-stroke engine" rather than "engine") and stating your audience, say a ten-year-old or mechanical engineering students. Ask specifically for the part you care about ("Let me open up the valvetrain"), and request numbers you can check, such as RPM or gear ratios. Finish the interactive page first and record the video afterwards, because recording is the slowest and most expensive step. Your best starting point is a variation on one of the sample machines.

explainx.ai also spells out the limits:

• Accuracy is unproven. "Real numbers" only means numbers from the model Claude wrote. They can look convincing and still be wrong.
• Cost and run time are unknown. Every machine takes a long Claude Code session, including the review loop. The README doesn't say what a machine costs.
• Setup takes work. The full video pipeline needs four tools: Node, Chromium, ffmpeg and Python.
• There's no official release, and only two commits.

Both sources also urge caution on security. A skill is made of instructions the agent follows and scripts it runs. On top of that, this one involves npm install and a downloaded browser. Read SKILL.md and the scripts first, and look out for network access, writes outside the output folder and shell commands. Ideally, run the skill in a throwaway folder or a container, not in a project that holds credentials. explainx.ai stresses that it hasn't tested the skill itself.

The bigger picture

Teardown shows what agent skills are good for: a complex workflow with fixed rules and built-in self-checks becomes a package anyone can reproduce. You used to have to learn 3D software first. Now you describe the machine and let it be built. According to explainx.ai, much of the excitement comes from the fact that you end up with a single HTML file that runs anywhere. That's far easier to share than a screenshot. Related tools such as Archify generate architecture diagrams. Teardown, by contrast, targets moving, mechanical processes.

It's most interesting for teachers, explainer-video makers and tinkerers who want to understand a camshaft or a differential. It's not fit for real engineering work: the result is a teaching aid, not a validated model. And GitHub stars measure attention, not quality. With a project that's a week old, keep that in mind.