Forcing Claude Code into TDD

Inspired by Paul Hammond’s approach, I’ve been trying to force Claude Code and Kiro into TDD.

Interestingly, I found prompting and explicit CLAUDE.md instructions alone were not enough to keep it on track on complex features. It bit off more than it could chew in one go - big bang engineering. It wrote tests that wrote the logic under test into the test itself, and tested no units, it got lost in doom loops, I could go on.

I had a lot more success in collaborating up front on a phased plan, that I asked it to write into a .agents/ directory, and telling it to refer back to it as working memory. It did especially well when I baked TDD thinking into the plan itself.

It’s been great fun and incredibly productive, but I can’t help but feel that it’s still more art than science to get great results out of this workflow. It takes quite a lot of discipline to shift your brain into doing almost nothing but review, it’s easy to get lazy.