Cross-domain recommendation (CDR) aims to alleviate the user cold-start problem of recommender systems by transferring information from auxiliary domains. Existing CDR methods can be divided into two main paradigms: the mapping-based paradigm and the joint learning paradigm. Although these works have achieved promising results, the former often fails to distinguish domain-specific features, while the latter suffers from suboptimal performance under data imbalance when one domain has limited data. Meanwhile, research into enhancing CDR through clustering techniques is also advancing, but they typically adopt fixed cluster centers, which restrict the adaptability of user representations and neglect the correlations between items and user clusters. In this paper, we propose a Cluster-Guided Cross-Domain Recommendation (CGCDR) framework to address these limitations. We introduce learnable cluster centers in each domain to dynamically capture evolving user intents and jointly optimize clustering and recommendation. Then, through soft assignment and weighted aggregation, we decompose each user’s interaction sequence into disentangled cluster-level intent before knowledge transfer. Finally, we design a cluster-assisted contrastive loss to enhance representation robustness in the target domain. Extensive experiments on benchmark datasets demonstrate that CGCDR outperforms state-of-the-art CDR baselines in cold-start scenarios. Our source code is available at https://github.com/yuhuping/CGCDR .