In response to the challenge that the current intelligent wargaming system exhibits insufficient human-machine adaptability and struggles to meet diverse tactical requirements, this paper presents a personalized adaptation framework for agents underpinned by the Large Language Model (LLM). The framework adheres to a “cloud-based development and design combined with end-side autonomous training” paradigm. Specifically, the development team formulates the fundamental model and the structured prompt engineering module in the cloud environment. Subsequently, users leverage natural language interaction to convert their personalized tactical specifications into executable reward functions, which enables the autonomous training of agents. This study devises two representative preference-based requirements and conducts end-side experiments. The experimental findings reveal that, in comparison with the sparse feedback rewards in the traditional setting, the reward functions generated by LLM can substantially enhance the training efficiency and tactical performance of agents. Simultaneously, through the structured prompting design and the interactive optimization mechanism, the framework effectively mitigates issues such as the command misinterpretation of LLM during wargaming. Consequently, it offers a novel solution for addressing the personalized adaptation problem in intelligent wargaming systems.