ARPG+: a simulation-based study of real-time coaching for educational LLM prompting

Abstract Despite widespread adoption of large language models (LLMs), most students cannot effectively prompt them. The core challenge is teaching students how to ask: transforming prompting from trial-and-error guessing into a systematic, transferable skill. Existing solutions, such as static templates, rule-based hints, and automated rewriting, either ignore individual learning needs or optimize outputs without building competence, leaving students dependent and unable to generalize. ARPG+ is a real-time coaching system grounded in cognitive load theory and zone of proximal development that senses when learners struggle, delivers calibrated just-in-time interventions, and fades support as skills develop. The system tracks learner capability with uncertainty quantification, estimates cognitive overload from behavioral signals, diagnoses prompt quality across six dimensions, and adapts scaffolding intensity through a dynamic schedule with periodic skill probes. A lightweight-deep dual architecture ensures fast responsiveness for routine interactions while reserving richer analysis for critical moments. Evaluation with simulated learners shows ARPG+ produces improvements: prompt quality increases 143% beyond unguided practice, learners achieve independence in 91% of final interactions versus 59% under fixed support, and the approach generalizes to other domains without retraining. Our work establishes that principled real-time coaching can improve prompt quality, accelerate learning, prevent cognitive overload, and foster durable autonomy. All reported evaluations are conducted on LLM-based simulated learners; the present work does not yet constitute empirical validation with real students, and classroom validation in authentic educational settings is identified as a necessary subsequent step.