Designing Meaningful AI-Generated Dialogue: The Behaviour-Driven Conditional Prompting Framework for Serious Games

Abstract

Many serious games use dialogue between players and non-player characters (NPCs) to enhance learning.
However, designing appropriate dialogue is often time-consuming for game developers. Recent advancements in artificial
intelligence, particularly Large Language Models (LLMs), have made open-ended dialogue with virtual characters feasible,
though managing it in educational contexts remains a significant challenge. This study explores how game designers can
guide open-ended dialogue powered by LLMs to create meaningful educational conversations. Expert interviews and a
review of existing approaches to implementing open-ended conversation in games led to the formulation of design
requirements for a new framework. Based on these insights, the Behaviour-Driven Conversational Prompting (BDCP)
framework was developed. This framework offers practical guidance for designers to create scenarios where behavioural
learning objectives are achieved through structured dialogue. It combines lock-and-key narrative design, dynamic character
prompts that dictate LLM-generated responses, and behaviour analysis prompts that assess player interactions. To validate
the framework, a functional prototype called ‘Detective Duck’ was created. This detective-style 'whodunit' game has players
solve crimes through open-ended conversation with AI-driven characters. Players encounter challenges such as persuading
hesitant witnesses, or verifying alibis. These challenges can only be solved by demonstrating reasoning and conversational
strategies relevant for detectives such as lateral thinking and persuasion. Upon demonstrating these behaviours, character
prompts can be dynamically adjusted, ensuring that - only then - players receive key clues needed to advance the narrative.
The framework and prototype were evaluated against the established design requirements and through expert interviews.
Results were largely positive, indicating that the BDCP framework supports meaningful, open-ended dialogue aligned with
educational objectives. However, some inconsistencies in dialogue coherence and adherence to designer intent were noted.
Future work will focus on refining narrative consistency and enhancing adherence to designer intent by fine-tuning the
language model, integrating AI-driven player feedback, and incorporating other game mechanics into the framework. These
improvements would further strengthen the BDCP framework as a tool for designing serious games centered around openended
conversation.