<p>Few-shot text classification involves transferring knowledge from a limited dataset to perform classification tasks in unseen domains. Existing metric-based meta-learning models, such as prototypical networks, have shown to be effective. However, these models fail to separate embeddings of similar classes, resulting in performance stagnation. Meanwhile, traditional text classification recently has succeeded with the support of extensive prior knowledge from large language models. However, in few-shot scenarios, it is challenging to generate discriminative sentence embeddings using pre-trained models, often leading to overfitting issues. Furthermore, random support set sampling adversely affects prototype quality, resulting in inaccurate category representations. In this paper, we propose the label-guided contrastive capsule network (LGCCN) to address these issues. LGCCN introduces label information as external knowledge for each category, comprising a label-guided contrastive clustering module (LGCCM) and a label-guided prototype enhancement module (LGPEM). LGCCM employs label information to guide a contrastive clustering process, using labels as anchors to enhance the process, yielding discriminative sentence embeddings from the pre-trained model. LGPEM iteratively generates a robust prototype with label guidance, significantly reducing the impact of random support set sampling. The proposed LGCCN was thoroughly evaluated on seven benchmark datasets from diverse application domains. Experimental results demonstrate that LGCCN outperforms the state-of-the-art (SOTA) method by 9.7% in 1-shot tasks and 1.1% in 5-shot tasks on news datasets. Additionally, on intent recognition datasets, LGCCN surpasses the SOTA with a 3.3% improvement in 1-shot tasks and a 1.4% improvement in 5-shot tasks, highlighting its superior performance in few-shot text classification.</p>