In scenarios where GPU resources are constrained or code privacy requirements are stringent, fine-tuning models with fewer parameters to achieve performance comparable to large-scale models remains a common and effective approach. In the field of Au- tomated Code Review (ACR), existing fine-tuning methods have shown promising results by constructing a Chain-of-Thought (CoT) to simulate the general process used by human reviewers. However, although the performance gain of a CoT-based ACR has been evidently supported by current studies, a “Code Analysis” phase—where human experts establish an overview understanding of the code prior to the actual review—has not been explicitly reflected in existing ACR CoT frameworks. To address this issue, we propose CoRe, which explicitly incorporates a “Code Analysis” phase at the beginning of the CoT. This phase includes summarizing the code, analyzing core flows, and examining the content of diff changes. We then employ a sequential fine-tuning strategy to train CoRe, i.e, the model is first trained to perform code analysis and subsequently trained to generate review comments. Empirical evaluations on the public CodeReviewer dataset and our curated CoRe dataset demon-strate that a 14B parameter model fine-tuned with the CoRe method outperforms state-of-the-art large models—such as DeepSeek-R1 and GPT-5—despite their parameter scales being several orders of magnitude larger. This highlights a significant efficiency advantage. Ablation experiments further verify that explicitly incorporating this preliminary analysis into the ACR-oriented CoT is the fundamental driver of this performance improvement.