Automated Generation and Refinement of Agile User Stories from Game Design Documents Using Large Language Models

Game Design Documents (GDDs) are widely used in the video game industry to capture gameplay mechanics, narratives, and user experiences. Often compared to a movie script, the GDD serves as the central planning document for game development. In modern game production, which typically follows the Agile framework, the primary planning artifact is the Product Backlog, where features are expressed as user stories. However, GDDs are often rich in narrative and loosely structured, which makes it challenging to systematically extract Agile artifacts such as user stories. This paper presents an approach that leverages Large Language Models (LLMs) and prompt engineering to automate the generation and refinement of user stories from GDDs. We apply three prompting strategies Chain-of-Thought (CoT), Refine-and-Thought (RaT), and a Hybrid variant using two LLMs (ChatGPT and DeepSeek), resulting in six generation pipelines per GDD. The quality of the generated user stories is evaluated using AQUSA, an automated assessment framework based on Agile criteria. Prompt refinement strategies are then applied to improve story quality based on AQUSA feedback. Experiments conducted on eight diverse GDDs (commercial and academic) show that prompt refinement eliminates all previously observed defects, including minimality, atomicity, and uniformity. Our findings highlight the feasibility of automating Agile documentation from narrative design sources, offering a scalable solution to bridge the gap between creative ideation and engineering requirements.