What if you could model with words?
What happens when you stop modelling every adjustment yourself and start describing what you want to change?
I've been working on an eyewear concept inspired by traditional Inuit and Yupik snow goggles. The project started with a simple observation: instead of relying on conventional lenses, these objects use narrow openings to control how much reflected light reaches the eyes.
I wanted to explore how that principle could be translated into a contemporary object. After several rounds of research, sketching and form exploration, I arrived at a monolithic eyewear design with a continuous front surface and short side supports.
The next step was turning the sketch into a workable 3D model. I've been using Rhino for years, so normally I would approach this through the familiar process of building surfaces, adjusting control points, checking transitions and gradually refining the geometry. But this time, I wanted to try something different.
Could I develop the model by having a conversation with AI?
For more behind-the-scenes updates on my Inuit snow goggles design process, follow me on Instagram :) @silatulayzeytin
01 / Connecting the tools
I came across Rhino's Model Context Protocol (MCP) integration, which makes it possible for an AI assistant to interact with Rhino through a connected set of tools. I connected ChatGPT to Rhino and started experimenting. The idea was fairly straightforward. Instead of writing scripts myself or manually executing every modelling command, I wanted to describe what I was trying to achieve and see how the AI could translate those instructions into geometry.
For example, I could ask it to create a surface, adjust a connection or modify an existing part of the model. Through the connection, it could work with Rhino rather than simply explain how I should do something. Of course, knowing what to ask for was another matter. I wasn't interested in generating an entirely new design. I already had a sketch, proportions and a particular visual character in mind. The challenge was whether I could communicate those intentions clearly enough for the model to develop without losing its original identity.
02 / From sketch to first geometry
I started by sharing my existing eyewear sketches and describing the basic structure. The object needed to be a single, opaque, monolithic piece. The front surface had to follow the curvature of the face, with narrow horizontal viewing openings and short sides that continued naturally from the main volume. I also wanted to avoid the conventional appearance of sunglasses. No separate lenses or typical frame construction. The form itself was supposed to control the light.
The first model gave me something to work with. It captured the general direction, but the proportions weren't quite right. Some connections felt disconnected, and the result still looked more like a collection of geometric elements than one continuous object. Still, seeing an initial model appear from a written description was interesting. Instead of starting from curves and surfaces, I was starting from the way I explained the object.
03 / One continuous form
One of the most important characteristics of this design is its continuity. I wanted the front and sides to feel like parts of the same volume. Not a front plate with two arms attached to it, but a continuous shape that wraps around the face. This became the focus of the next few conversations. I asked for smoother connections, more consistent curvature and better transitions between the different areas of the eyewear.
The interesting part was that a change which looked correct from one angle could create a new problem somewhere else. Sometimes the model became smoother but lost too much of its original volume. Other times the connection improved, but the proportions shifted.
So I started becoming more specific. Keep this part. Change only that transition. Preserve the front silhouette. Don't reduce the volume too much. I realised I was spending almost as much time describing what should remain unchanged as what needed to change.
04 / Small changes, different results
Once the overall form was closer to what I wanted, I moved on to the surface details. This included the transitions around the front, the curvature near the sides and the overall thickness of the object. Some adjustments seemed minor when I described them, but they had a significant effect on the model.
For example, I wanted the front to remain relatively thin while the sides retained enough volume to feel structurally convincing. My initial targets were around 3 mm at the front and 5 mm at the sides, with further refinement around the connections. Working through conversation made it easier to propose changes, but not necessarily easier to predict the results.
I still had to inspect the model, rotate it, compare it with the original sketch and decide whether the changes actually improved the design. And sometimes they didn't. A smoother surface wasn't automatically a better surface. Occasionally, a slightly sharper transition was important to preserve the character of the object.
That was probably one of the most useful parts of the experiment: being reminded that technically cleaner geometry and better design aren't always the same thing. :)
05 / But how does it actually fit?
At this point, the model was beginning to look closer to the original concept. But there was an obvious question. Would it actually sit on someone's face?
Until then, most of my attention had been on the overall shape and its visual continuity. The eyewear still needed proper contact points, particularly around the nose. I introduced a nose support and started refining its position and geometry. Then I brought in a digital head model. This changed the conversation quite a bit.
Looking at the eyewear on its own, I could make decisions about proportions and surface quality. But once I placed it on a head, I started noticing different things: the distance between the front and the face, the relationship between the side supports and the temples, and whether the nose piece was positioned where it needed to be.
The first placement didn't resolve everything. I went back to the model, adjusted its position and continued refining the nose support. Some areas that looked convincing in isolation needed another round of work once I saw them in context.
It's a familiar part of industrial design, really. An object can look perfectly reasonable on a screen until you start considering how someone will use it.
06 / What I learned from modelling through conversation
After several rounds of revisions, I started to notice how this workflow was changing the way I approached modelling. Normally, there's a fairly direct relationship between what I want to change and the commands I use to change it. I know which surfaces I'm editing and how I want to construct them.
With this approach, there was an additional layer. I had to communicate the intention, let the connected tools execute the changes, inspect the result and decide what to do next. Sometimes that felt surprisingly efficient. Particularly when I knew exactly what I wanted but didn't want to spend time manually reconstructing a part of the model. Other times, it became frustrating. A simple instruction could produce an unexpected result, and correcting it required another conversation. In some cases, it might have been faster to make the adjustment myself.
I don't think the interesting question is whether this workflow is faster than conventional CAD modelling in every situation. It clearly depends on the task. What interested me more was the shift in how I interacted with the model. I was still making design decisions, but the execution was becoming more conversational. It also made me think about how important design vocabulary becomes in this process. Words like smooth, continuous, thin or balanced sound clear enough when discussing a sketch.
But when those words need to become actual geometry, their ambiguity becomes obvious. What does smoother mean in this particular area? How much curvature should change? Which parts of the original surface should remain untouched? The more specific I became about these things, the easier it was to evaluate the results.
Still, the AI didn't always interpret the instructions the way I intended. And the responsibility for judging the result remained with me. That distinction matters. :)
07 / From screen to physical testing
Eventually, I reached a point where continuing to adjust the digital model wasn't necessarily going to answer my remaining questions. I needed to see it in the real world. I selected three versions and sent them for 3D printing. :)
The next step is to evaluate the proportions, how the eyewear sits on the face, the contact points around the nose and temples, and how practical the design feels once it becomes a physical object. I'm particularly curious about how the front volume translates from screen to reality, and whether the geometry that looks convincing digitally will feel equally convincing when worn. For now, the physical tests are still ahead.
“A different way of working, not a finished answer”
I started this experiment out of curiosity. I wanted to know whether I could take an existing industrial design sketch and develop its geometry through conversation, rather than modelling every adjustment manually.
The short answer is yes, to a certain extent. I was able to create and refine geometry, explore different transitions, adjust proportions and work through initial ergonomic considerations. But it wasn't an automatic process, and it certainly wasn't a matter of writing one prompt and getting a finished product. There was a lot of checking, correcting and trying again. I still needed to understand the form, recognise when the model wasn't working and decide which changes were worth keeping.
I don't see this as a replacement for knowing how to model. If anything, having that experience helped me recognise when the results were going in the wrong direction. What I find promising is the possibility of spending less time on certain repetitive modelling operations and more time exploring alternatives, evaluating them and making decisions. Whether that proves more efficient over a complete product-development process is something I still want to understand.
For now, I'm treating it as another tool in my design process. One that I'm still learning how to use. And, as with any other modelling method, the real test begins when the object leaves the screen.
To be continued with the physical prototypes. :)