Aerial-ground person re-identification (AGPReID) aims to retrieve target person images across both ground and aerial views. Compared to traditional ReID tasks, AGPReID faces significant appearance variations due to heterogeneous views, which limit the effectiveness of existing AGPReID methods. To address this, we propose a Global-Local Prompts-driven Semantic Guidance (GLPSG) framework that leverages global prompts for global visual-semantic alignment and local prompts for fine-grained local alignment. To mitigate interference from irrelevant patches while preserving semantic consistency during local prompts learning and alignment, we propose a person-relevant semantically consistent patch selection. Additionally, we apply two strategies to diversify the global and local prompts, respectively. Extensive experiments on multiple benchmarks demonstrate the effectiveness of our approach on AGPReID and its satisfactory performance on traditional ReID tasks.