For Chinese Named Entity Recognition (NER) tasks, achieving better performance with fewer training samples remains a challenge. Previous works primarily focus on enhancing model performance in NER by incorporating additional knowledge to construct entity features. These approaches neglect the semantic information of entity labels and the information of entity boundaries. Moreover, conventional methods typically treat NER as a sequence labeling task, which makes them inadequate for addressing the issue of nested entities. We propose a new span-based approach by using contrastive learning and prompt learning to address these problems. By pulling similar entities closer together, pushing dissimilar entities further apart, and leveraging entity label information, we improve model performance in few-shot scenarios effectively. Experimental results demonstrate that our method achieves significant performance improvements on a sampled Chinese nested medical dataset and several other flattened datasets, providing a new insight into addressing challenges in few-shot NER tasks.