Vibe coding vs vibe-CADing: what changes when the output has to fit
Vibe coding works because code has tests; physical parts do not. Five differences between vibe coding and vibe-CADing, and what a good tool does about them.
By BlacksmithAI · · 3 min read
Key takeaways
- Vibe coding and vibe-CADing share an idea: describe what you want and let AI write it.
- They differ where it hurts: a part has no test suite, a miss costs material, and "looks right" is not a check.
- Good vibe-CADing tools make up for it: real measurements in, live parameters, and a stated match out.
If you have vibe coded, vibe-CADing will feel familiar: you describe, the AI writes, you nudge. Here is where the two part ways, and what that means for the tools.
1. Code has tests. Parts have tolerances.
The secret of vibe coding is that software checks itself: run it and see. A CAD model has no equivalent on screen. It renders beautifully whether the bore is 22.00 mm or 22.40 mm, and only one of those takes the bearing.
What the tool must do: hold the model to something real, a caliper reading, a scan, or the original part, and say how close it is in millimetres.
2. Undo is free in code. Not in steel.
A bad commit costs a revert. A bad part costs material, machine time and, for moulded parts, tooling. The feedback loop that makes vibe coding fast is the expensive part of manufacturing.
What the tool must do: make the cheap loop on screen as complete as possible: flag every inferred dimension, rebuild instantly when you change one, and export a file a shop can check before cutting.
3. You review code. You review dimensions.
Nobody reads every line of vibe-coded software, and often nobody needs to. In CAD the equivalent of reading the code is checking the critical dimensions and the features they belong to, and that cannot be skipped.
What the tool must do: present the model as an engineer thinks about it: a feature tree in plain words and a slider per dimension, not a wall of generated code.
4. In code, the words are the spec. In CAD, the part is.
A program is fully described by what it should do. A replacement part is described by the part it replaces: its holes line up with something, its outline clears something. Words lose that. This is why the strongest vibe-CADing starts from evidence (photos, a scan or a STEP file) rather than a prompt. See what AI reverse engineering is.
5. The output is not the code.
Many vibe-CADing tools, including ours, have the AI write the model as code; ours is build123d on the OpenCascade kernel. But the deliverable is not the code. It is a B-rep solid, a STEP file a machine shop can quote, an STL a printer can slice, or a feature tree in Fusion 360 an engineer can keep editing.
Side by side
| Vibe coding | Vibe-CADing | |
|---|---|---|
| Input | A prompt | A prompt, photos, a scan or a STEP file |
| Check | Tests, running the program | Measurements, match against the scan |
| Iteration cost | Seconds | Seconds on screen; days and material once made |
| What you review | Behaviour | Critical dimensions and features |
| Deliverable | Source code | STEP, STL, a feature tree |
Read the full definition in What is vibe-CADing?, or try it on a real part with AI Reverse Engineering.
Frequently asked questions
Can I vibe-CAD with a general chatbot?
You can ask a chatbot to write CAD code, and for simple parts it works. What a general chatbot lacks is the loop: building the solid, measuring it against real evidence, showing you live parameters and exporting manufacturable files.
Is vibe-CADing only for people who cannot use CAD?
No. Engineers who use CAD every day use it to skip the hours of sketching and constraining a part they already understand, and spend their time on the dimensions that matter.
Try it on a real part
AI Reverse Engineering turns photos, an STL scan or a STEP file into editable CAD. It is in private beta, with plans from $24.99 a month.