Attention Deficit Hyperactivity Disorder (ADHD), a prevalent neurodevelopmental condition affecting cognitive and behavioral functioning, presents significant challenges in clinical management. While Large Language Models (LLMs) show promise in mental health, their application to ADHD-specific interventions is impeded by two critical barriers: (1) insufficient alignment with evidence-based Cognitive Behavioral Therapy (CBT) protocols, and (2) a lack of structured clinical reasoning pathways for generating contextually appropriate recommendations. To address these, we curated a specialized ADHD clinical case dataset comprising annotated CBT sessions, enabling LLMs to capture nuanced therapist-patient interactions. Building on this, we propose a mind map-guided meta-prompting framework. This framework integrates LLM capabilities with expert-designed ADHD mind maps to structure and guide the model’s reasoning process. Our hybrid evaluation, which combined blinded assessments by certified psychiatrists with computational metrics, demonstrated marked advancements in clinical appropriateness compared to baseline models. The framework achieved a mean expert rating of 4.47 on a 5-point Likert scale for therapeutic relevance, outperforming leading models such as HuatuoGPT (3.96), DeepSeek-V3 (4.12), DeepSeek-R1 (4.24), and GPT-4 (4.21).