Typically, there's two ways to do this.

The first is to just use the fact the GPT-4 will have seen a lot of blender code so just knows how to do it.

The second way is to tell GPT-4 in the prompt what the API surface looks like and have it script against that.

You can actually just look in this file and see that the first method is being used. He feeds in a prompt with a basic example.

https://github.com/gd3kr/BlenderGPT/blob/main/__init__.py

Oh, thanks. I thought somebody got a fine-tuning version of GPT-4. However, the author used the exact same prompts to demonstrate the ability of his interface, so I don't think it says much about its fitness for the purpose.