User stories have become the most widely adopted technique for requirements elicitation and documentation in agile teams. However, ensuring high-quality requirements through user stories remains a significant challenge. Natural Language Processing (NLP) techniques have been proposed to enhance the quality of these artifacts, but technological constraints have limited their effectiveness. More recently, the emergence of Large Language Models (LLMs) has prompted new approaches. Nonetheless, existing proposals remain preliminary and constrained. This research presents a conceptual framework for an LLM-based approach designed to support the generation of high-quality user stories from real-time conversations with end-users, thereby reducing the reliance on traditional requirements documents. The approach integrates context-specific information and incorporates automated and manual validation steps to ensure the quality of LLM-generated content and mitigate hallucinations. A preliminary evaluation was conducted with software engineering experts to evaluate the proposed approach's adherence to real-world problems. The proposal was perceived as easy to use, effective, and capable of producing high-quality outcomes. Furthermore, participants expressed a positive experience with both the introduction to the approach and their involvement in the evaluation. This engagement prompted reflections on novel opportunities for integrating LLMs into their professional practices.
Pereira et al. (Mon,) studied this question.