Um estudo empírico sobre o CHATGPT aplicado a escrita e desenvolvimento de requisitos no formato de história de usuário

Abstract

Artificial intelligence tools have been used to automate activities in various areas of software engineering. One such tool that has gained popularity is ChatGPT, especially because it allows interactions in natural language. Studies show that ChatGPT possesses capabilities in natural language processing and code generation, but it also provides incorrect responses in several situations. The capacity of ChatGPT in software engineering activities, such as code generation, bug correction, user story generation, or other related tasks, is still not clear. Therefore, additional studies are necessary to further understand the potential and utilities of ChatGPT in software engineering. In this study, two stages are conducted: firstly, 50 user stories are generated, including acceptance criteria, complexity, effort estimation, and related requirements. Secondly, 8 user stories are selected from the list of 50 requirements in step 1 and coded using NodeJS with Typescript and Prisma. For code generation, prompts at three different levels of description are used based on ChatGPT’s output. Although ChatGPT successfully generated the requirements, the code generation process exhibited bugs and syntax errors, with some stories requiring manual corrections. Nevertheless, the generated requirements prove useful as a starting point for developing software engineering documents. The study’s results include the success percentage of ChatGPT in coding user stories, the NodeJS project developed by ChatGPT, the list of 50 generated requirements, and an overall analysis of the outcome.