AI and software

One useful question: when a tool can generate more code than I can comfortably read, what practices help keep my understanding real?

For now, the answer is probably still boring in the best way: small changes, clear interfaces, reviewable diffs, and writing down what I think I know.