meet stovetop

The study and creation of a Python tool for Autodesk Maya using Claude Fable 5.

the idea

The goal was to create a tool that could take objects in Autodesk Maya and separate them into multiple shatter fragments. The goal was to use this for breaking artifacts for a VR archaeology project. I wrote certain minimum criteria the tool had to meet before uploading it to Claude Fable 5, before having it do the rest:

UI Overview:
+ Add object button - Load in the object that is going to be cut
+ Fragmentation - Number of fragments that need to be made
+ Refresh Button - Refreshes the edge cuts if the segments are updated
+ Initialise button - Creates the shattered asset

Since I was diving more into learning how LLMs work, I figured it would be a good experiment to use Anthropics new Fable 5 model to help make this tool.

I am very impressed by the improvisation that Fable 5 made to my original tool. In terms of the efficiency of the code and how effectively it was written, I had just enough Python knowledge to understand how the code actually works and remove some unnecessary components (Such as plugin imports that weren’t needed).

The approach I took was to take the object and use a VoronoiNoise map to set cuts along the UV plane, where the generated darkness would reflect where the cut marks would be on the actual object, and use the UV coordinates to reflect where the separation is.

While I was able to ideate how I wanted to approach the tool and the steps were possible in theory, I couldn’t figure out how to apply the VoronoiNoise map segments to the geo-cutting. I used Opus 4.8 to ask how it would approach making a tool like this, but it came with a lot of errors in the process, where the cut values wouldn’t connect with the vertices made by the noise map.

With Fable 5, I gave it the same prompt and questioning on how to approach the tool, and it not only completed the task, but it also took it a step further by making it a non-destructive tool.

That being said, Claude Code, and especially Fable 5, is a very strong tool. Performing this process reminds me of the “Vibe Code” terminology, which I’m not a big fan of. It’s like ordering your own bowl at Chipotle, but we don’t call it “Vibe cooking”. If you wrote the script yourself, say so. If you used AI, then say you used AI.

I plan to push other models further and use them for more code-based scenarios, and I am excited to see what developments Anthropic makes to the iterations of its future LLMs.