Predict-Observe-Explain in AI Tutoring: Design Principles for Socratic Dialogue in Engineering Simulation Games

Socratic AI tutoring promises to transform educational technology from answer-giving systems to guided discovery companions. However, operationalizing Socratic dialogue in AI systems remains challenging, particularly for open-ended engineering design problems with multiple valid solutions. This paper proposes design principles for Socratic AI tutors grounded in the Predict-Observe-Explain (POE) pedagogical framework, demonstrating how this inquiry-based strategy naturally integrates with physics simulation for engineering education. Drawing from theoretical foundations (Vygotskian scaffolding, cognitive load theory, self-determination theory, constructivist productive failure) and empirical evidence on tutoring effectiveness (effect sizes 0.4–0.8), we synthesize ten actionable design principles for AI tutor development. The principles address learner agency preservation, adaptive scaffolding, productive struggle balance, question type selection, tell/ask decision-making, and simulation-integrated inquiry. We operationalize Socratic dialogue through POE cycles where AI tutors (1) prompt predictions with reasoning before simulation runs, (2) guide observations during interaction, and (3) scaffold explanations of discrepancies post-run. This structure leverages cognitive dissonance for conceptual change while maintaining learner autonomy. Detailed examples demonstrate application to water infrastructure engineering design in the AMAIG simulation platform. The novel contribution lies in bridging constructivist learning theory with implementable AI prompt engineering, specifically addressing the gap between generic Socratic principles and practical deployment in simulation-integrated, open-ended design contexts. The principles enable cross-age adaptation (13–65+) and integration with retrieval-augmented generation for domain-grounded questioning.