The paper introduces a novel graph condensation method named CraftIng RationaL trajectory (CTRL), which aims to address the issues of high costs and storage in training graph neural networks (GNNs) on large-scale graphs. CTRL focuses on improving the gradient matching process to reduce accumulated errors between the training trajectories of GNNs on the original and synthetic graphs. The method combines cosine similarity and Euclidean distance to refine the gradient matching criterion, ensuring both the direction and magnitude of gradients are aligned. Additionally, CTRL uses a clustering algorithm to initialize the synthetic graph, ensuring a more even feature distribution compared to traditional methods. Extensive experiments on various graph datasets and downstream tasks demonstrate the effectiveness of CTRL, showing superior performance in node classification, graph classification, cross-architecture applications, and neural architecture search. The method achieves state-of-the-art results on several datasets and reduces the impact of accumulated errors, making it a promising approach for efficient graph condensation.The paper introduces a novel graph condensation method named CraftIng RationaL trajectory (CTRL), which aims to address the issues of high costs and storage in training graph neural networks (GNNs) on large-scale graphs. CTRL focuses on improving the gradient matching process to reduce accumulated errors between the training trajectories of GNNs on the original and synthetic graphs. The method combines cosine similarity and Euclidean distance to refine the gradient matching criterion, ensuring both the direction and magnitude of gradients are aligned. Additionally, CTRL uses a clustering algorithm to initialize the synthetic graph, ensuring a more even feature distribution compared to traditional methods. Extensive experiments on various graph datasets and downstream tasks demonstrate the effectiveness of CTRL, showing superior performance in node classification, graph classification, cross-architecture applications, and neural architecture search. The method achieves state-of-the-art results on several datasets and reduces the impact of accumulated errors, making it a promising approach for efficient graph condensation.