Pharos Night: Crown Pursuit: An AI-Native Deck-Building and Tactical Arena Game Design

Pharos Night: Crown Pursuit explores the integration of generative AI as a foundational gameplay mechanic rather than a mere content-generation tool. The project aims to transform traditional deck-building into an open-ended synthesis process and elevate non-player characters (NPCs) into autonomous social and strategic agents. By utilizing large language models (LLMs) to drive both card creation and NPC behavior, the design seeks to foster emergent gameplay where player intent, negotiation, and deception become central to the experience.

The game is developed in Unity, utilizing the ChatGLM series models to manage NPC interactions and card synthesis. The system architecture employs a request queue to handle concurrent LLM calls, ensuring stable performance during gameplay. Card generation is achieved by mapping qualitative natural language inputs and material combinations into structured JSON data, which Unity then parses into functional game mechanics. This allows for a dynamic, non-linear experience where players influence the game world through natural language spells and social bargaining.

A small-sample study involving 13 participants evaluated the effectiveness of these AI-native mechanisms. Quantitative results indicated high player satisfaction, with perceived strategic meaningfulness scoring 4.15 out of 5 and overall experience scoring 4.13. Qualitative feedback highlighted that 92.3% of players found AI-driven NPCs more dynamic than traditional counterparts, and 84.6% viewed the synthesis system as a novel addition. However, the study also identified a core design tension: while the openness of AI-driven systems creates novelty, it can simultaneously reduce player predictability and sense of control. The findings suggest that while generative AI significantly enhances engagement, future iterations must prioritize improved feedback and guidance mechanisms to balance emergent complexity with player agency.