Hm could be but I think it's likely just the finickiness of LLMs in general and requires experimenting with prompts and models