It has to do with the structure of the information/inputs. Its more a DSL with a structure rather than a jumble of API arguments. So first few times out it may take some getting used to, but as that structure becomes clear it becomes more intuitive. Thats also why even with way more matplotlib examples in the world, LLM agents do better "guessing" how to use plotnine... or at least it was that way a few months ago when we were testing them head to head.