A step-by-step guide for how to kill the line-by-line code review

The article presents a step-by-step guide for implementing a more effective code review process by separating verification tasks from judgment. The key points are:

1. Verification should be handled by automated agents that check code against established rules and criteria, while judgment about the code's intent and appropriateness remains with human reviewers.

2. Two main layers of review are created:

- Acceptance criteria describe what the change should achieve.

- Invariants are the codified standards and rules that must always be true, regardless of the change.

3. Invariants define the fundamental rules of the engineering organization, while intent and acceptance criteria capture the specific purpose and expected behavior of the code change. Both need to be captured before the change is made.

4. The review process then builds a test plan based on the criteria and invariants, gathers evidence to support the verdict, and finally allows a human reviewer to make a decision on the change.

5. The key to success is that invariants are refreshed regularly after merges, ensuring that the team's rules remain up-to-date. Ownership, wording, and scope are important considerations when drafting invariants.

The goal is to create a review process that combines the efficiency of automated verification with the nuanced judgment of experienced human engineers, resulting in higher quality code while avoiding the limitations of relying solely on AI-generated reviews.