PRL: Prompts from Reinforcement Learning