GenCAD

There are mechanical engineers out there who can literally model objects nearly as fast and they can 'think' about the layout of said object. If you look at the complexity of, say, a CAD model from a real, highly complex aluminum casting section of an automotive subframe, or the living-room-sized cross-fuselage spar forging of a fighter aircraft, with hundreds of ribs and fillets and features- and compare that to the simple model you are trying to make in OpenSCAD, you should quickly realize the parallels in difficulty you are trying to express (similar to the person without knowledge of C++ or Python watching someone be able to build applications by typing code from their fingertips as if they already knew what to type...)

You are struggling for a few reasons- 1., it is a knowledge hurdle of an entire field you are trying to surmount- again, go watch someone actually model a real, complex part and watch the speed, they can do so in a tool like Solidworks, CATIA, NX, etc... at a rate that is far different because they have experience that it can honestly take even good people years to accumulate - and 2. they are using professional tools - you mention OpenSCAD, like it is CAD, but it isn't. It is programmatic mesh generation, and it turns out that programmatically typing out how to generate complex things is much more difficult than a combination of a graphical GUI and graph-based generator that all big CAD programs figured out starting in the 1980s. If those tools you use were really the best way to make complex models 'paramaterized', then why do we design our fighter jets, Formula 1 cars, or Space X rockets in Dassault's CATIA or Siemen's NX ?

You want a LLM to take a sketch to your CAD, but what I'm saying is, there are people out there that can skip the sketch and build the CAD as fast as you can likely hand draw the first sketch, and these are skills you can actually learn, but you may just be using the wrong tools and have not had the practice necessary.

I think this is possible, but the ‘trick’ would be translating your instructions in English into some kind of language that the CAD software understands.

I’m on a bunch of 3D printing forums, and everyone tries to describe what the finished product would LOOK like. They end up making PICTURES when what they really want is a STL file.

Two dimensions are easier to visualize then three, so let’s put it this way:

If you wanted to turn “English” into “a 2D image that’s dimensionally accurate”, you’d want to translate from “English” to “SVG.”

SVG is dimensionally accurate. JPG isn’t. The file format itself has no concept of “dimensions” only “pixels.”

I haven't tried telling AI to "make a thing," but I'm able to get Co-pilot to refactor code. It's just the geometry that makes my head spin.

It will take much longer than a day for AI to get to this level, so there's not much to lose by just learning how to use the software now :)

From the outside, the hard part of designing a chair is making a blueprint. At least making a blueprint looks hard to people who've never made one. According to outsiders, the next layer of the onion is perhaps inserting reasonable constraint dimensions for similar reasons.

From the inside, as a guy who's recreationally made furniture, the hard part is judgment about joint selection and design, experience with wood warping (all wood changes shape with the seasons, a good woodworker makes it look easy to work around and a bad one makes expensive firewood that rapidly falls apart). Another insider PoV is judgment about wood selection to get the correct balance of final finish durability and appearance. Finally working toward outer layers of the onion, its time to do parametric joint design decisions... What's the ideal number and size of dovetail joints for, perhaps, a drawer.

I've seen prints of chairs before I don't need a LLM to make one similar to the ones I've seen before and could probably make from memory (at least ones I built myself), the library has loanable books and woodworking magazines. I do see the attraction from the outside.

Consider something like a Windsor chair. The larger the wedge in the spindles the tighter and longer lasting the chair until you break something trying to force them in; there's a lot of judgment and experience in designing, selecting, and installing spindles, but none of it is written down so it'll be hard to train a LLM... Tighten it until it breaks then don't tighten it that much next time. Most super detailed plans for Windsors are for inferior machine produced replicas which are not necessarily useful for a fine woodworker and are not exactly what craftsmen would aspire to. People who want "a cheap chair" will buy a 4-pak of folding chairs from walmart anyway, not make a homemade Windsor-style chair.

Another somewhat more blunt example is for actual woodworkers the "problem" with hand cut dovetails isn't knowing what they look like or how to make a diagram of one, but gaining the experience behind a hammer and chisel to push your luck while cutting them as far as possible without going too far and turning the part into scrap. One unavoidable part of woodworking is I've turned quite a bit of wood into scrap on the last step; oh well make another. At least I can burn scrap wood to keep warm LOL.

Its kind of like from outside the programming fraternity the non-programmers think the only skills required to program are typing real fast and being very experienced at fizzbuzz during interviews. But that doesn't work IRL, from an inside-out perspective.....

The woodworking world is not exactly lacking for a library of "semi-decent" plans. An automated system to make enormous quantities of low quality unverified and untested plans would not really help the field, no.

Basically leverage the randomness to create many variations, then select the most accurate variation automatically.

Terribly wasteful of time and processing power, but so is using GPU time to make pretty pictures randomly.

I would even argue that for basic modelling majority of tools/features in CAD operate at the abstraction closer to CSG for describing what and B-rep is only treated as how. Just like good chunk of code based CAD use combination of CSG for what and triangle mesh based geometry engine for how. That's assuming you consider standard 2d->3d operations (extrude, revolve, sweep along arbitrary 2d profile) as valid primitives for CSG.

User comes into direct contact of B-rep in very specific situations: 1 doing operations like fillets/chamfers/draft/thickness based on intermediate geometry, 2 attaching sketches or other features to generated geometry or using generated edges (instead of new sketches) for guiding operations like complex sweeps, 3 surface based modelling workflows where you build up the the solid from individual faces typically including complex curved surfaces.

In case of 1 and 2 the the dependency on b-rep based representation is only marginal, in theory you could select edges in triangle mesh based underlying representation but the final result but quality of result wouldn't be as nice and TNP issues for parametric model editing would likely be even bigger than it is for existing CAD. That's not really CSG territory anymore but isn't exclusive to B-rep either, and involves a bunch of work that's outside the scope of B-rep. In non parametric mesh modellers with more destructive editing workflows like blender chamfers and fillets work fine. And if anything for reliable parametric models you often want to limit dependencies on intermediate geometry as that depends on CAD keeping track of where each edge/face originated from outside the b-rep and increases the chance of TNP issues.

3 is critical for industrial design containing large amount of complex curved surfaces like cars and other consumer products, but there are also many more technical parts where it can be completely ignored. Cad tutorials for beginner tutorials almost completely ignore this category of cad modelling. The part about not being exclusive to b-rep also applies for surface modelling part.

What is your workflow for llm integration to openscad?

In my own project I use LLMs very sparingly and hesitantly, but made the observation that they are not very useful on the hard parts of CAD. I expect this is because of a lack of training material. Most professional CAD applications are proprietary and books on the topic are usually sparse on implementation details. The non-BRep CAD applications such as OpenSCAD and family are probably overrepresented in the training data.

This might also explain why people's experiences with LLMs are very varied. If you stay in the happy path of CRUD web development and stuff all is nice and well, but if you start to veer off this path you get more and more challenged.

What wasn't working? I would love real feedback -

I’ve seen several vibe coded attempts at a geometric kernel (including several of my own) and this happens every time.

Vibe coding a geometric kernel is practically impossible because sooner or later* the LLM inevitably takes the tessellation shortcut and if you don’t catch it, the codebase is completely compromised. At the end of the day, there isn’t enough training data in architecture or algorithms (opencascade solvespace and truck being the only real examples, all significantly worse than commercial kernels like Parasolid or ACIS).

* usually as soon as you ask it to do anything non-trivial. If you’re lucky you’ll get a naive Newton marching algorithm on analytical bodies, which is slightly better but has the same problem with degenerate and pathological cases (coincidence, tangents, parallels)

https://arxiv.org/abs/2603.04337 https://arxiv.org/abs/2603.05607 https://arxiv.org/abs/2605.01171

For a more detailed review: https://github.com/lichengzhanguom/LLMs-CAD-Survey-Taxonomy