In increasingly complex manufacturing systems, enhancing and managing productivity is crucial. To achieve this, many companies have adopted digital twin technology. Digital twins collect and manage real-time data from manufacturing systems. Based on this data, production reports are generated, and various data analyses are performed. The time required for these analyses varies depending on the system's complexity. In highly complex and sophisticated systems like semi- conductor fabs, the amount of data is substantial, and the structure is intricate. Therefore, these analyses take a significant amount of time. This paper proposes a methodology for designing Large Language Model (LLM) prompts for semiconductor fab digital twins. LLMs are charac- terized by their flexibility and scalability. Additionally, the performance of an LLM varies based on the quality of the prompts that serve as input data, making it essential to craft prompts suit- able for digital twin applications. This paper describes the components of prompts for semicon- ductor fab digital twins and conducts LLM performance experiments in a simulation environment.
