Temporal Information Graphs (TIGs), which aim to capture temporal relationships among various entities and events, are widely utilized to analyze dynamic systems and track changes over time. Therefore, many TIG models have been proposed for modeling the dynamic representation of TIGs. However, these models continue to adhere to the “pre-train, fine-tune” paradigm, disregarding the inherent discrepancies in training targets across diverse tasks. Additionally, these works predominantly emphasize the evolution of individual nodes, while neglecting higher-order collective relationships. Moreover, the necessity of fine-tuning the entire model parameters for each downstream task restricts flexibility. In light of this, we develop a community-based prompt learning method for TIG models (CPTIG). First, we introduce a “pre-train, prompt” paradigm and utilize it on TIGs through a paired prompt function. Then, a community-based prompt function is designed to prompt hierarchical information and bridge the semantic gap between different tasks. Specifically, two prompt tokens with temporal projection are generated within the detected community, aligning the downstream dynamic node classification task with the pre-training link prediction task. The token pair consists of a candidate dynamic class label and a node entity. Extensive experiments on three real-world datasets validate the proposed method’s superior performance over baseline models, along with significant efficiency improvements.