Traditional episodic meta-learning methods have made significant progress in solving the problem of few-shot node classification on graph data. However, they suffer from an inherent limitation: the high randomness in task assignment often traps the model in suboptimal solutions. This randomness leads the model to encounter complex tasks too early during training, when it still lacks the necessary capability to handle these tasks effectively and acquire useful information. This issue hampers the model’s learning efficiency and significantly impacts the overall classification performance, leading to results that fall well below expectations in data-scarce scenarios. Inspired by human learning, we propose a model that starts with simple concepts and gradually advances to more complex tasks, which facilitates more effective learning. We introduce CPT, a novel curriculum learning strategy designed for few-shot node classification. This approach progressively aligns the meta-learner’s capabilities with both node and task difficulty, enhancing the learning efficiency for challenging tasks and further improving overall performance. Specifically, CPT consists of two stages: the first stage begins with basic tasks, aligning the difficulty of nodes with the model’s capability to ensure that the model has a solid foundational ability when facing more complex tasks. The second stage focuses on aligning the difficulty of meta-tasks with the model’s capability and dynamically adjusting task difficulty based on the model’s growing capability to achieve optimal knowledge acquisition. Extensive experiments on widely used node classification datasets demonstrate that our method significantly outperforms existing approaches.