Uso de LLMs na automatização da geração de histórias de usuário de qualidade

Abstract

The requirements specification phase is the process of recording system requirements in a clear, complete, and consistent manner, using accessible language and simple elements. However, it can be challenging in practice due to differing interpretations and stakeholder expectations. In agile methodologies, this challenge is amplified as it demands continuous communication, rapid adaptation to change, and intense collaboration among stakeholders. In this scenario, user stories have become a widely adopted technique due to their simplicity, user-centric focus, and alignment with iterative development. Nevertheless, the manual writing of these stories often suffers from quality issues, hindering a shared understanding of requirements and negatively impacting the final product. With the advancement of Large Language Models (LLMs), such as ChatGPT, new opportunities arise to automate the generation of user stories. Still, the literature lacks experimental studies investigating the effectiveness of these approaches, the quality of the generated texts, professional perception of utility, and their application in different contexts. Given this, this work aims to support requirements engineers responsible for specification in the automated generation of high-quality user stories. To achieve this, US-Prompt was iteratively created—an artifact based on the combination of Meta-Prompting and Few-Shot Prompting techniques. As an evolution of this artifact, the web tool Story AI Generator was implemented, which operationalizes US-Prompt in an interactive environment for the automated generation and refinement of user stories. To conduct this research, the Design Science Research (DSR) methodology was adopted, focusing on developing artifacts to solve practical problems through experimental studies articulated in relevance, design, and rigor cycles. The developed artifact was evaluated based on the Quality User Story (QUS) framework, which defines quality criteria for user stories, as well as an analysis of user perception regarding usefulness, ease of use, and intention to use, grounded in the TAM3 model. The results indicate that using LLMs with prompting techniques is promising for supporting the generation of clearer, more consistent user stories that adhere to requirements engineering best practices. The developed tool demonstrated potential for adoption in real-world contexts, expanding the applicability of the proposed approach. It is concluded that the application of LLMs with techniques such as Meta-Prompting and Few-Shot Prompting, combined with their operationalization in support tools, can effectively contribute to the automated generation of quality user stories in agile contexts, increasing productivity, consistency, and requirements quality while reducing the cognitive effort of the professionals involved.