Development of an LLM-first 2D Videogame

Traditional role-playing video games (RPGs), where players assume the role of a character in a fictional
world, often rely on static dialogues and storylines. This can make replaying the game feel repetitive and
less engaging, as the story and conversations remain the same. In contrast, tabletop RPGs thrive on the
creativity of the human game master, who continuously invents new settings and stories and allows play-
ers to interact naturally with the world. With the rise of large language models (LLMs), new opportunities
are emerging for digital games. A key strength of modern LLMs is their ability to generate human-like
text in response to natural language input. Integrating LLMs into games enables live, context-sensitive
responses to player actions and dialogue, bringing some of the freedom and spontaneity of tabletop
RPGs into video games. This project aims to develop a 2D RPG with a top-down view that combines
traditional gameplay with the creative power of LLMs. At the start of each game, an LLM generates a
unique world. During gameplay, characters respond to the player in real time, meaning no playthrough is
the same. For developers, this moves focus away from scripting fixed content to designing systems and
prompts that enable the AI to take over parts of the storytelling.The game is developed in Unity using a 2D pixel art style and includes typical RPG elements such as story,
exploration, non- playable characters (NPC), battles, items and quests. Game content is partly generated
using LLMs and partly built from predefined assets. For example, character parts like heads, bodies and
clothes are defined in advance, and the LLM selects from these to create NPCs. Similarly, the map is
built from predefined 30x30 tiles depicting elements like forests, plains and villages, which the LLM as-
sembles into a complete world. The generation process starts with a prompt defining the map, followed
by prompts for quests, NPCs and items. NPC appearance and battle skills are generated in separate
prompts to keep outputs manageable. For simpler tasks like NPC visuals, skills and interactions, we use
OpenAI’s cost- efficient GPT-4o-mini. For tasks requiring more consistency and depth, such as map and
quest generation, we use the o3-mini model. We built a functional top-down 2D RPG with LLM-supported elements that generates new content on
each playthrough. Our key finding is that the quality of generated content depends heavily on the context
provided. Without sufficient details, quests and characters often become repetitive or inconsistent. For
example, NPCs might block quest progression, or items can appear in incorrect locations. While these is-
sues occasionally caused confusion, the eight people we selected to test our game generally responded
positively, appreciating the dynamic nature of the experience. A major challenge we encountered was
balancing the amount and complexity of prompts with system performance. Longer or more detailed
prompts improve output coherence but increase the risk of missing details. Splitting input into smaller
prompts helps but raises processing time. A promising approach is sending multiple prompts simulta-
neously for parallel processing, which could improve content accuracy and reduce wait times. Overall,
this project indicates that a thoughtful combination of AI and traditional techniques can open exciting
new possibilities for creating more dynamic and imaginative games in the future.