Vibe CADing 2025 Q1™️, e.g. Create models via AI generated Fusion360 Python Scripts

Random or Off Topic
1

Curious if anyone’s figured out a VIBE CADing toolset and AI assisted process that’s working for them in 2025 Q1? Obviously, everything will change by Q2+.

Guessing a few folks here are familiar with Vide coding. Competitive driven innovation in LLM space only seems to be accelerating. Guessing more utility/$ will continue for some time. In just the past week there’s been a bunch of new interesting AI models (Gemini 2.5, DeepSeek v3, QWEN 2.5 OMNI, etc…).

In the past, I’d use LLM based AI to assist with troubleshooting and learning how to do a specific task in OpenSCAD. Lately, given the current Vibe Coding hype, I’ve been leaning on LLMs more to generate a significant portion of an OpenSCAD based thing.

I like OpenSCAD, but it’s a PITA to do somethings that are fast/easy in CAD software like Fusion360, so, since tech CEOs/CFOs assure us that LLMs can quickly generate masses of code…

Lately, have been exploring whether LLMs can help generate maintainable code easily, and quickly enough to model a parametric thing via code (stored in github repo, for collaboratively working, forking, versioning, etc…).

Getting mixed results… Mostly unusable, but I’m learning more about how to use Fusion’s Python based APIs to model things.

image
image1920×992 309 KB

Am steering model to refactor constants/parameters, but also serialize manually edited state to .json to help the model inspect/reflect and reason on how to extend/fix. At least, that’s the idea…

# ------------------------------------------------
# Global “magic numbers” or defaults
# You can later load these from a JSON config.
# ------------------------------------------------
PARAMS = {
    "EMT": {
        "outerDiameter_mm": 17.0,  # approximate OD for 1/2" EMT
        "innerDiameter_mm": 15.5,  # approximate ID for 1/2" EMT
        "length_mm": 1219.2,       # ~48 inches in mm
    },
    "Gripper": {
        "jawSpan_mm": 50.0,        # ~2" or smaller
        "armThickness_mm": 3.0,
        "armLength_mm": 80.0,
        "tpuPadThickness_mm": 2.5,
    },
    "Lever": {
        "leverPivotHeight_mm": 1016.0,  # ~40" from bottom
        "handleLength_mm": 100.0,
        "triggerOffset_mm": 30.0
    },
    "Hardware": {
        "commonBoltDiameter_mm": 5.0,
        "commonBoltLength_mm": 30.0
    },
    "Belts": {
        "useGT2Belt": True,
        "beltWidth_mm": 10.0,
        "pulleyDiameter_mm": 16.0
    },

Maybe VIBE CADing with OnShape or something else would be better? Didn’t try because LLM model tells me it’s not as well tuned for OnShape. Also, OnShape’s Free plan is too restrictive (design privacy), and their cheapest not-Free plan is too much.

2

“Vibe coding” is a fit way to describe it and so cringe.

I haveyseen anyone deacribe it that way unless they are making fun of it. But it definitely is a short, descript way to cut to the chase of what is happening.

I have the same concerns with CAD as I do with code. If you ask for something that fits a particular standard and it instead makes something close, how do you catch that error? How can you be sure it is confident in the dimensions or just halucinating?

I’m surprised you care about the privacy. It wasn’t that long ago that github forced free users to public only repos. Honestly, their “public” designs are kind of hard to navigate through because most projects aren’t designed for public consumption. So no one goes there looking for public projects. At least, I haven’t been successful trying. There is no security through obscurity. But at least you won’t be bothered by casual nuisances because the designs are so hard to find.