Something I think would be useful for LLM coding agents would be "take this pretty complex over engineered code and refactor it to be cleaner, more readable, more maintanable" - but then it keeps refactoring it different ways like 5 times and gives me a PR with each version