Learning on homophilic and heterophilic graphs has gradually gained attention. Compared to homophilic graphs, where neighborhood information is reliable, nodes in heterophilic graphs may face neighborhoods that are dissimilar to their own, containing inappropriate information. However, existing multi-view graph clustering methods address this issue in heterophilic graphs solely by masking or weighting structural information. Such solutions make it harder for these nodes to access valid information–their overall information content becomes impoverished. While reducing the aggregation of inappropriate information, it also discards much suitable information. To address this challenge, we propose a Cluster-Guided Contrastive Encoding method called H2MGC, which unifies the modeling of homophilic and heterophilic graphs and aims to enhance data by generating positive samples for nodes using prior knowledge. In particular, building upon contrastive learning with adaptive adjustment of topological weights, we introduce cluster center-guided data augmentation. By leveraging information entropy to measure the latent relationship between cluster centers and nodes, we construct positive samples for each node, thereby achieving unified data augmentation for both homophilic and heterophilic graphs. Extensive experiments demonstrate that H2MGC achieves state-of-the-art performance across diverse homophilic regimes while exhibiting stable convergence.