Prompt Engineering for LLMs and Its Use in Banking

This thesis explores the field of prompt engineering for large language models (LLMs) and its practical application in digital banking environments. The theoretical part outlines key techniques, prompting strategies, and emerging trends, including role-based prompting, retrieval-augmented generation, and decomposed prompting. Special focus is placed on output quality evaluation, security challenges, and the risks associated with jailbreak prompts. The practical section is based on hands-on experience designing and optimizing prompts for a digital assistant at a major Czech bank. It includes an analysis of the system's architecture, categorization of prompt types, and a summary of findings from user testing and client feedback. The results demonstrate that well-designed prompts significantly enhance the relevance of responses, improve customer satisfaction, and strengthen trust in the bank’s digital channels.