Something else that’s interesting is how consistent different prompts are. Like, these are three different requests with different contexts, and the lizard character is extremely consistent between prompts.
Either they have a very small training set or it’s purposefully low variation.