Towards Generalist Prompting for Large Language Models by Mental Models