Part 2: When Nobody Grades Their Own Homework

Short summary

The author describes using two separate Claude Code agents—a builder with write access and a read-only reviewer—to evaluate subjective UI quality that can't be captured by automated tests. The reviewer rejected the builder three times, catching issues like a fake crossfade and a missing loading state, and even flagged a blind spot in the test evidence rather than the code itself. The key insight is that an agent reviewing its own work nearly always approves, so independence and good evidence are essential for meaningful AI-driven quality gates.

• •Two separate Claude Code agents (builder + read-only reviewer) enforce subjective UX quality via a written rubric
• •Reviewer rejected builder 3 times, catching a fake crossfade, missing loading state, and flawed test evidence
• •An agent reviewing its own work almost always approves— independence and evidence quality are critical

Generated with AI, which can make mistakes.