Developer finds AI code reviewer kept solving the wrong problem despite repeated fixes

A developer building an AI-assisted editorial pipeline for technical writing discovered that repeated improvements to their reviewer tool were addressing symptoms rather than root causes. The core issue was that a score-first review process judged drafts too early, producing QA-style feedback instead of genuine editorial critique. Adding an adversarial review stage — which assumed the primary assessment was wrong until evidence proved otherwise — proved more effective than expanding rubrics or lengthening prompts. Further failures revealed the AI consistently gave additive feedback without flagging redundancy, causing drafts to grow longer without improving. The developer ultimately found that isolating distinct reasoning tasks into separate pipeline stages outperformed prompt tuning as a reliability strategy.