Generative artificial intelligence (AI) in vehicles enables multimodal, context-aware driver interactions, yet practical design strategies remain underexplored. This study derives information prioritization and output modality strategies to minimize driver distraction and enhance user experience, proposing multimodal design guidelines. Using a mixed-methods approach, we analyzed 12 near-futuristic driving scenarios and 44 information types. Surveys with 63 drivers and focus group interviews with 20 early adopters revealed that vehicle warnings and unexpected events are high-priority, while notifications like email alerts are less critical. Pearson correlation analysis identified 23 critical information types in the first quadrant (Q1), confirming that information necessity and preference vary by driving context. Users favored concise, proactive information delivery, minimal voice output, and privacy protection, emphasizing AI intervention timing aligned with driver attention and aversion to information overload. A quadrant-based output strategy was proposed: Q1 information is delivered immediately via voice and haptics, Q2 and Q4 via passive visual displays, and Q3 upon user request. Interaction policies were refined to address output timing, modality prioritization, and user engagement, mitigating overload and excessive intervention. Additionally, a Large Language Model (LLM)-based architecture was proposed to dynamically adjust multimodal outputs using real-time data and user context, supported by user experience (UX) flow examples. This architecture optimizes outputs, enhancing Human-Machine Interface (HMI) processes for Original Equipment Manufacturers (OEMs). This study offers a theoretical framework and practical strategies for generative AI-based in-vehicle multimodal interfaces.