We beat GPT-4o's baseline with a simple re-prompting loop