Abstract: Power electronics design serves as the foundation for power electronic equipment development. The implementation of intelligent power electronics design can significantly improve research and development efficiency while reducing costs. However, power electronics design tasks inherently involve multi-step decision-making characteristics, which puts forward higher requirements for the design of algorithms. A collaborative design framework that integrates deep reinforcement learning（DRL） with large language models（LLMs） is proposed. The methodology first employs LLMs to interpret natural language design requirements and generate initial states for DRL. Subsequently, through DRL’s reward-driven mechanism, the system performs trial-and-error iterations on these initial states to ultimately obtain design solutions that maximize rewards, which fully satisfies design requirements and constraints. This synergistic framework achieves complementary advantages by addressing the limitations of LLMs in vertical domain knowledge and generation accuracy, while simultaneously mitigating the challenges of large search spaces, low training efficiency, and convergence difficulties in DRL. The framework is validated through two typical applications: power electronics topology generation and layout/routing design, with experimental results demonstrating its effectiveness. Further research should explore algorithmic optimization, task-algorithm alignment, and inherent limitations to enhance the framework’s robustness and scalability.