Side note, it doesn't always fuck up, but most of the time it'll give just completely wrong matplot instructions. The recent tattoo post comes to mind.
It's good for python stuff, specifically, potentially because python as a language is the closest we have to a natural, descriptive programming language, and as an LLM that might make connections between functional behavior and language easier. That said, it sometimes tells you to do things that won't work because the libraries you're using have some specific incompatibility issues between them and the only way you can find out is via github issue discussions.