Investigando estratégias de engenharia de prompt para melhorar a qualidade de requisitos gerados por LLMs

Abstract

The quality of software requirements is a determining factor for the success of projects, since specification failures generate high rework costs and inconsistencies in the final product. Large Language Models (LLMs) have been explored to support requirements elicitation and documentation; however, they still present limitations regarding the accuracy of business rules and hallucinations. This study empirically investigates the impact of different Prompt Engineering strategies (Input Only, Chain of Thought, and Tree of Thought) on the quality of functional requirements generated by the GPT-5, Gemini 3, and DeepSeek-R1 models. The adopted methodology followed a hybrid approach, combining a quantitative analysis of structural completeness (actor, action, object, and condition) with a qualitative evaluation based on clarity, completeness, consistency, and relevance, concluding with practical validation involving stakeholders. The results indicate that, although the models demonstrate high performance in basic attributes, the specification of Conditions represents the greatest challenge, with presence rates below 60% under the Input Only strategy. The combination of the GPT-5 model with the Chain of Thought technique produced the highest overall quality requirements, achieving average scores above 4.90 in the criteria considered in the qualitative analysis. However, validation with experts revealed that only 33% of the generated requirements were fully accepted without restrictions, with Information Omission being the most prevalent defect (50%). Furthermore, a strong positive correlation was observed between the structural completeness of conditions and the human perception of quality. It is concluded that the use of advanced reasoning strategies is essential to mitigate LLM deficiencies, serving as a solid foundation for requirements engineering; however, it does not eliminate the need for specialized human review.