Generative AI can provide timely, individualized explanations, but it remains unclear when learners perceive them as high-quality and engaging, and how they can effectively support textbook-like practice and clarification. We investigate this in CS1 by pairing explainable-AI (XAI) feedback with adaptive questioning: an IRT-based computerized adaptive testing (CAT) engine maintains a productive difficulty range, and post-response explanations clarify concepts. In a mixed-methods, between-subjects study ( \(N{=}50\) ; one 10-item session), we compared adaptive with random sequencing. Surveys, system logs, and interviews indicated high usability in both groups (SUS \(\approx 90\) ) and valid alignment of difficulty with ability estimates in the adaptive condition (correlated with prior CS1 achievement). Relative to random sequencing, adaptivity increased perceived usefulness and satisfaction and showed an upward trend in feedback quality; intention to reuse did not reliably differ. Targeted difficulty was associated with longer time-on-task and more thorough reading. Learners described explanations as clear and actionable and requested adjustable depth. Overall, results suggest that an XAI-driven, adaptively sequenced question–explanation loop supports textbook-like practice and immediate clarification by aligning difficulty and delivering concise, concept-referenced explanations. Design implications include preserving difficulty alignment, offering progressive disclosure of explanation depth, and signaling coverage of learning outcomes.