I find explicit time bounds are useful for tasks like this, otherwise the LLM will almost certainly return too early.