The post discusses a novel approach termed System Prompt Learning (SPL), where language models (LLMs) utilize learned strategies to improve their problem-solving capabilities. The SPL framework has specific limitations for efficiency: a storage cap of 10 strategies per problem type and a maximum of 3 strategies applied during inference. This selective process ensures robustness by only utilizing strategies with at least 5 attempts and a success rate of 40%. The effectiveness of this system has been particularly noted in reasoning models such as DeepSeek-R1 and QwQ, where successful strategies notably enhance the models' decision-making processes. Additionally, there are expectations to evolve SPL by incorporating collaborative strategy sharing and extending its applications to multimodal problems. The integration of SPL with existing optimization techniques in projects enhances its utility, and future explorations may lead to meta-learning approaches to better formulate strategies efficiently.