Why would one not use TDD when coding with Claude? It's so cheap!
Yes you need to help it with the types of tests: you still need to know what you want from it. But once you have all types of tests (unit, db, fake-services, e2e, etc.) in place and documented; it can basically write tests until you cov-tool says it's 85%. Then you can ask it to find the weakest tests: you review those and make sure they are not weak, or Claude understands why they are not weak. Then let it find the next batch of weakest tests. Etc.
TDD finally makes sense economically for me on the types of projects I usually work on.
But I use TDD when coding with AI.