elvarofficial/teardown

teardown is a Claude Code skill for people who want to understand how a machine works by taking it apart in a browser rather than reading about it. You describe a machine in plain language, and Claude builds a single offline HTML page in which the machine runs in the middle, surrounding panels display its real numeric values, a button opens the working part, and an unattended page tours itself.

How a request becomes a simulator

The skill ships SKILL.md, which the README describes as the material Claude reads: how a machine is built, visual rules, and a review loop. Supporting assets include a runtime/stage/shader module and a build script that bundles a machine into one HTML file with three.js and fonts inlined, so the result needs no network connection README.

A concrete workflow is to ask Claude Code for a described machine, for example a machine that can be taken apart, or to request a teardown and a short recorded video of it. The repository also contains three complete machines — a V8, a Cessna 337 landing gear, and a Strandbeest — each split into a shared machine definition, a live page, and a recorded film. You can build the packaged examples directly with node sims/build.mjs and the example name README.

Requirements differ by output

Using the skill requires Claude Code plus an install step through npx skills add. A one-time npm install and a Chromium install through Playwright are then needed in the installed folder. Node.js 20 or newer is the stated baseline. Only two extra paths are optional: ffmpeg for the recorded video, and Python 3 for sound README. Frame-by-frame 60 fps capture and still review sheets are handled by sims/capture.mjs, and procedural sound is run through run_sfx.py, mastered to −14 LUFS README.

Scope and provenance

The project is MIT-licensed. Its runtime, frame capture, sound engine and review rules come from the MIT-licensed claude-motion project, and the recorded sound samples are CC0 from Freesound, prepared by a separate project with sources listed in the repository README.

The evidence describes the mechanism and the example machines, but does not evaluate simulation accuracy against real engineering data, nor does it state model-coverage limits for arbitrary machines. Treat the “real numbers” panels as part of the generated simulator's design rather than as verified measurements. The stated fit is exploratory: turning a question about a mechanism into a manipulable visual model.