in my experience TDD is a very powerful paradigm for use with LLMs.

it does a good enough job of wrangling behavior via implied context of the test-space that it seems to really reduce the amount of explanation needed and surprise garbage output.