This echoes my experience. I often use ChatGPT to help with D&D module design and I found that O1 did best when I told it exactly what k required, dumped in a large amount of info and did not expect to use it to iterate multiple times.