The author shares insights on reviewing pull requests (PRs) in an era where AI aids code generation but complicates the review process. While tools like Claude and Codex can help, they can't replace the nuanced understanding that human reviewers bring. The author emphasizes the importance of high-level abstractions, system integrity, and ensuring that code is intuitive for new team members. A well-structured codebase should allow reviewers to grasp its components without needing to understand every implementation detail. The author outlines specific heuristics for evaluating PRs, focusing on general reviewability, design, code quality, testing, and error handling. Key points include checking if the PR description is detailed, assessing the size of the changes (less than 500 lines is ideal), and evaluating whether the proposed abstractions make sense. The review process should also consider the author's effort, any comments they leave, and the effectiveness of unit tests. The author notes that while AI can assist in identifying basic errors, it struggles with deeper conceptual issues that require a more holistic view of the PR. Despite advancements in AI for code review, the author argues that current tools haven't kept pace with code generation. The review agents can catch obvious mistakes but often fall short when it comes to understanding the overall intent or structure of a PR. The author mentions a humorous anecdote about developers creating code for an ideal scenario rather than the practical realities they face, highlighting the disconnect between AI-generated code and human expectations. Ultimately, the human element remains crucial in maintaining coding standards and taste in a landscape increasingly filled with AI-generated code.