← All use cases

中文

Building software · Claude Code

Design a 3D-printed robot arm by talking to Claude Code

Idea + starter prompt, not run by us yet

For: Hobbyists who want to build hardware but have never used CAD

Claude Code wrote one macro that FreeCAD re-rendered live, so the whole arm came from conversation. The builder printed and assembled it, then added voice and webcam control: speech and an image go to Claude, which returns JSON actions that local code turns into servo angles on an ESP32. About 30 hours produced four working servos.

Steps

  1. Install FreeCAD and ask Claude Code to write one macro that builds your design. Tell it the parts, sizes and servos you have.
  2. Iterate by chat and look at each render before you print.
  3. Plan for reprints. The builder taped servos, glued a lid, redid a base and replaced a failed hinge.
  4. Remember FreeCAD's macro runner needs ASCII-only scripts, so ask Claude to avoid special characters.
  5. Test with power limited and your hands clear of the arm. There is no feedback loop and calibration is rough.

Starter prompt

Starter prompt. Change the [brackets] first.

Trial first: build only one test part, a simple bracket sized for [servo model], in a new folder called [arm-trial]. Do not design the whole arm until I have printed and checked that part.

I want to design a small 5-joint robot arm in FreeCAD using only a Python macro. My servos are [model, size], my printer bed is [size] and my material is [PLA or other]. List the parts you plan to build with their dimensions, and ask me anything unclear. Then write one macro file I can re-run after every change, with ASCII characters only.

Build one part at a time and stop for my feedback. Warn me about parts that are hard to print, like print-in-place hinges. Stop and wait for my yes before writing any motor control code or anything that moves hardware. A finished design is not permission to power it up.

Original post: hiitzroman.com · found via agentusecases.ai. Written up in our own words. We haven't run it, so test it on made-up data first.