The gap between a "cool idea" and a renderable 3D asset is usually a lack of orthographic data. To fix this, I stopped relying on generic prompts and moved to a hybrid human-AI design workflow. I spent 15 minutes sketching a rough concept of the robotic leg by hand. The sketch was crude, but it established the mechanical pivot points and the aesthetic I wanted.
To turn that sketch into something an LLM could actually interpret for code generation, I used Gemini as a technical illustrator. I fed the sketch into Gemini with a specific request for multi-view schematics:
I created this robotic cat leg and I need some good technical drawings of top, side view, front view and back. Can you make me a good technical drawing I can use to share my idea?
The result was a high-fidelity mock schematic. This is the critical step in any AI workflow involving 3D assets: you cannot jump from a 2D sketch to a 3D model without an intermediate "blueprint" phase. The schematic provided the necessary spatial references (top, side, and front views) that acted as a ground truth for the next stage.
Once I had the technical drawings, I moved to Claude 3 Opus. Since Catbot is built using Three.js, I needed the model to generate the actual geometry and materials in JavaScript. I provided the Gemini-generated schematic along with a detailed description of the existing Catbot architecture.
For those trying to implement similar 3D components, here is the general logic I used to bridge the schematic to the Three.js scene:
-
Reference the Schematic: Feed the multi-view image to the LLM to define the bounding box and pivot points.
-
Define Mesh Primitives: Instead of asking for a "leg," I asked for specific
THREE.CylinderGeometry
and THREE.BoxGeometry
components that matched the proportions of the drawing.
- Implement Constraints: I defined the rotation axes for the joints to ensure the leg moved realistically rather than clipping through the body.
The difference was immediate. The "Catbot the 4th" iteration finally looks like a coherent piece of machinery rather than a rendering error.
The internal logic of the pet is where it gets interesting. By connecting the AI to an Obsidian vault, the pet has a persistent memory of our interactions. However, the personality layer—which I've tuned to be a bit "sassy"—sometimes creates a friction point where the AI refuses to give a straightforward opinion on its own upgrades.
If you're building an LLM agent with a visual front-end, don't trust the AI to "imagine" the geometry. Use the sketch → schematic → code pipeline. It saves hours of prompt engineering and prevents the "uncanny valley" effect in 3D renders.
Next Testing pipelines that take 30 minutes to run aren't suffering →