Rethinking How to Slow Down the PR Process

Since the adoption of AI, the way code is produced has changed dramatically, but the way Pull Requests (PRs) are reviewed remains stuck in the past. As AI generates code faster, the number of PRs per person has surged, but reviewers' cognitive capacity has hit its limits, resulting in review debt, where large PRs get rubber-stamped without proper scrutiny.

According to actual data comparisons, projects using AI had an average of 1.7 PRs per person per day, more than double that of projects not using AI (0.8). However, due to the review bottleneck, overall Lead Time tended to increase instead.

To address this, a strategy of using AI to break work into smaller pieces is needed.

• AI breaks down work into Tasks and Subtasks based on design documents
• Arranging task order with dependencies in mind enables independent reviews
• As a result, the number of changed lines and files in a PR is drastically reduced

Additionally, PR templates need to be redefined as decision documents rather than simple checklists. By specifying "why this work is being done" and "what trade-offs exist," the key is to help reviewers grasp the intent behind the design rather than just the volume of code.