Designing a framework for ethnography-driven prompt engineering in social work

This study proposes a framework for integrating ethnographic principles into prompt engineering (PE) for large language models (LLMs) in social work. While PE has emerged as a key methodology for optimizing LLM outputs, it often lacks grounding in the cultural contexts of end users. By bridging theoretical approaches and methodologies from social and computer sciences, the proposed framework addresses the limitations of a strictly semantic approach in PE. The framework’s theoretical foundations are grounded in empirical data gathered during an ethnographic field study conducted within a social service organization. Ten interviews were analysed following the key stages of the ethnographic analysis method. Key cultural themes composed of multiple semantic relationships were uncovered and then connected to core prompt components. These components were further elaborated into various prompting techniques and developed into a set of prompt templates that can be applied to LLM evaluation and further customization.