Idea
Use time pressure as the point. One hour, one ugly interface, and a very direct question: does the AI improve the right things first?
Status
We made meaningful progress fast, then hit the familiar wall where semantic correctness and actual usability part ways.
Problems Faced
- The easy fixes arrived instantly.
- The more important judgement calls still needed a human eye.
Techniques That Helped
- Pair automated checks with manual reading order and keyboard passes.
- Treat suggested fixes as drafts, not answers.
Lessons Learned
Accessibility is a strong test for whether an AI understands intent or just patterns.