Prompting Visual-Language Models for Efficient Video Understanding