If AI can do the work, maybe the test should be more focused on what AI can’t do? This is like anyone still doing a traditional coding interview with leetcode problems just because they haven’t yet done the work to figure out what to test for in a world where Claude Code exists.