FeatureBoardby Lewis Valentine

On large software projects, their design, and their users

It can build a V12 engine from a text brief and still fail a cylinder-on-cylinder union. That gap says less about the model than it does about how CAD was designed — around human limitations that no longer exist.

A lonely mental space

I think it's important to realize when you're actually out in a lonely mental space. I met up with an uncle recently and it took me five minutes (seven years in conversational time dilation) to really introduce what an AI-built CAD kernel was. Do they know what CAD is? If not, how about drafting or design? Do they use AI? Do they understand what an MCP plugin is? Can I at least approximate it for them in a way they get?

It's a lonely space, building a large-scale software project. The scope is enormous — the kind that would normally take hundreds if not thousands of developers years to write. In doing so I can't possibly cover everything all at once, and am constantly comparing what I have built to a battle-tested product. As are my customers and evaluators.

What it can do, and what it can't

I was pleasantly shocked by what it could do in text-to-CAD. I had only hoped for test data to drive the development, but I got a full-fledged solution to the problem. On the other hand, as I have leaned into that more and more for development and bugs, I can see the human-driven modeler lagging far behind. Torus intersections and cylinder-on-cylinder unions fail easily. Primitive intersections just fail.

Yet it can make a V12 combustion engine via MCP. A dovetail desk. Turbine blades and an entire 42 MW natural gas plant in a scene. Fasteners and all.

Why the basics fail and the hard scenes don't

Why does it fail at the basics and yet still work for complicated scenes? It works differently than a human developer and has different needs. We've built out applications that are some hundreds of times more code than mine to handle the wide variety of human needs, and the inability to force users through specific workflows to aid their design. It's possible for me to simply include hints in the MCP tooling surface that drastically increase the success of the model using it and avoid buggy areas.

The fragility of the human worker

In some ways this highlights some of the fragility of the human worker. We don't do ourselves many favors in the contest against an AI worker. We skim signs, ignore warnings, and minimize our response to problems to conserve energy. We fear responsibility and extra tasks. We hold meetings to argue that others should do the work, and leave them too drained to do what we need.

Operators built for limits that no longer exist

To go back to software design for a moment: we've chased the demands of the human users toward software that is more complicated, and that drags the AI down loops it doesn't need to follow. How important are booleans if you can just directly edit the geometry appropriately and quickly validate collisions and mating conditions? We've poured millions of hours of human effort into operators that are designed around human limitations that no longer exist. If you can conjure up a grate or a handrail from scratch, why reach for the way you know to create that geometry?

Design is what you want, and whether it holds

It's time to think more abstractly than the previous years of CAD. Design is about what you want and how you can validate that it holds up to your design constraints and goals. The design loop is closing, from many people validating many constraints to one AI instance iterating through simulation, design, and manufacturing information to produce optimized designs at incredibly low cost.

I used to think about classes, arguments, and semicolons. No longer! Today I purely think about how to design systems to converge on data-driven goals.


It's time for CAD designers, coders, and others to start doing the same. If you're ready to make the switch, try our NativeCAD MCP plugin in Claude today — the one-click Claude Desktop bundle, the npm line and the install guide are all on the NativeCAD page.

← Previous Post
What a CAD Kernel Is Actually Made Of