I've had pretty good success with using LLMs to generate basic shapes using python and cadquery (which generates real parametric step files you can edit in fusion, not glorified triangulated STLs). Yesterday I had GPT5.5 build a python script to generate fender style Mandolin with separate neck and body, with correct bolt holes for the neck, gibson style bridge, and stop-tail, even fretboard with the little dots and cutouts for the fret wire (I didn't ask for these, but it added them anyways). Everything looks correct and to scale. I can pastebin the .py later if anyone is interested.
Text to CAD doesn't need papers. You can literally just try it and see that it works well with the frontier models. If you want reification/meshing I recommend [1] which is what Godot uses. You can throw the results in a physics engine in an afternoon and see for yourself.
This wasn't obvious a year ago, but today CAD literally reduces to Simon Wilson's pelican test, since CAD is largely a matter of functional CSG, and CSG is really not that different from SVG. It's just one more dimension, which it turns out is not a problem.
LLMs consistently one-shot CSG based video game levels with interesting physics puzzles (citing myself). Given this I'm willing to conclude that the frontier models are good at automated CAD if given the correct harness. But I guess a lot of people don't know this yet.
I got a freecad API script for a catfood colander / dust sieve out of gemini / cline in a single prompt supplied with 10 measurements and a good description of the vision. Printed first try, the two halves mated without issue, and the fit was what I pictured. Under a certain threshold of complexity and given sufficiently specific instructions it works. I've also spent many hours on more complicated tasks with little success.
Personally I can complain to no end about arXiv papers being insufficiently rigorous, especially in the text classification area that I've been interested in since 2005 or so.
Show me the blog posts where people talk about the results they got "vibe coding" and those arXiv papers look great in comparison!
There is the insidious thing with LLMs is that they can get the general shape of something right but that thing will not be useful if the last 1% is wrong. It might be that the operator sees the problem and fixes it, but it may also be that the LLM hypnotizes the operator into not seeing the errors and gaps.
I know there are many sorts of problems where I've had good experiences with LLMs but I know other people have had bad experiences and some of it might be my skill but some of it is just plain luck.
I don't think I've ever heard a mechanical engineer say "torus" in my life unless they're talking about the car. When you are doing feedback with a human operator you use terms like "make this thicker" or "rotate that this way" while pointing at them. Text does not have this.
The only time text-to-SCAD has ever worked for me was when Claude (the app, not even Claude Code) decided spontaneously to spin up an environment and check the renderings of its SCAD code. That session lead to something 90-95% of the way to the finished product, and in some ways even surpassing my expectations by looking up measurements for relevant products instead of using placeholders.
Modifying the prompt and then trying it again did not lead to that self-verification loop and the output was unusable garbage.
I've had pretty good success with using LLMs to generate basic shapes using python and cadquery (which generates real parametric step files you can edit in fusion, not glorified triangulated STLs). Yesterday I had GPT5.5 build a python script to generate fender style Mandolin with separate neck and body, with correct bolt holes for the neck, gibson style bridge, and stop-tail, even fretboard with the little dots and cutouts for the fret wire (I didn't ask for these, but it added them anyways). Everything looks correct and to scale. I can pastebin the .py later if anyone is interested.
Text to CAD doesn't need papers. You can literally just try it and see that it works well with the frontier models. If you want reification/meshing I recommend [1] which is what Godot uses. You can throw the results in a physics engine in an afternoon and see for yourself.
This wasn't obvious a year ago, but today CAD literally reduces to Simon Wilson's pelican test, since CAD is largely a matter of functional CSG, and CSG is really not that different from SVG. It's just one more dimension, which it turns out is not a problem.
LLMs consistently one-shot CSG based video game levels with interesting physics puzzles (citing myself). Given this I'm willing to conclude that the frontier models are good at automated CAD if given the correct harness. But I guess a lot of people don't know this yet.
[1] https://github.com/elalish/manifold
I got a freecad API script for a catfood colander / dust sieve out of gemini / cline in a single prompt supplied with 10 measurements and a good description of the vision. Printed first try, the two halves mated without issue, and the fit was what I pictured. Under a certain threshold of complexity and given sufficiently specific instructions it works. I've also spent many hours on more complicated tasks with little success.
Personally I can complain to no end about arXiv papers being insufficiently rigorous, especially in the text classification area that I've been interested in since 2005 or so.
Show me the blog posts where people talk about the results they got "vibe coding" and those arXiv papers look great in comparison!
There is the insidious thing with LLMs is that they can get the general shape of something right but that thing will not be useful if the last 1% is wrong. It might be that the operator sees the problem and fixes it, but it may also be that the LLM hypnotizes the operator into not seeing the errors and gaps.
I know there are many sorts of problems where I've had good experiences with LLMs but I know other people have had bad experiences and some of it might be my skill but some of it is just plain luck.
I don't think I've ever heard a mechanical engineer say "torus" in my life unless they're talking about the car. When you are doing feedback with a human operator you use terms like "make this thicker" or "rotate that this way" while pointing at them. Text does not have this.
The only time text-to-SCAD has ever worked for me was when Claude (the app, not even Claude Code) decided spontaneously to spin up an environment and check the renderings of its SCAD code. That session lead to something 90-95% of the way to the finished product, and in some ways even surpassing my expectations by looking up measurements for relevant products instead of using placeholders.
Modifying the prompt and then trying it again did not lead to that self-verification loop and the output was unusable garbage.