One-Shot Federated Learning (OSFL) addresses the communication inefficiency of traditional federated learning by limiting client-server interaction to a single round. However, existing OSFL methods often suffer performance degradation in data heterogeneity scenarios due to overfitting of dominant classes on the client side and limited generator fidelity in data-free knowledge distillation. To address these challenges, we propose FedCGP, a new OSFL framework that leverages CLIP-guided, prototype-based distillation to enhance both client-side training and server-side knowledge integration. On the client side, we introduce prototype similarity-weighted knowledge distillation, which uses CLIP embeddings to provide class-wise guidance, mitigating overfitting to locally dominant classes and improving ensemble generalization. On the server side, we design a prototype-constrained fidelity training strategy, combining a sine intra-class consistency loss and a prototype contrastive loss to enhance the semantic fidelity of synthetic data. Extensive experiments on four benchmark datasets show that FedCGP consistently outperforms state-of-the-art OSFL baselines under various data heterogeneity settings. Specifically, on CINIC-10 with Dirichlet \(\alpha =0.1\) , FedCGP achieves an accuracy improvement of up to 14.1% over FedSD2C. Our ablation studies further verify the individual contributions of the prototype-based components.