I'm dealing with this now because I'm working on something that is very far outside my familiar domains so even if I had the time to hand-code it, I wouldn't be able to do it.

My answer so far is to dedicate a substantial token budget to continual polishing, refactoring, debugging, cleanup and explanation. Some practical steps:

1. output has to be perfect. Anything that looks even slightly off get's logged and a session is started to investigate, this usually leads to a bunch of other issues.

2. Ask the AI if the code works the way you think it does -- you should still have some model of how the software works in your head

3. Ask for documentation and explanation around interfaces and make sure you understand the interfaces. It's ok if something's a black box, as long as you have a good understanding of what goes in and what goes out.

4. continually ask for performance improvements and assessments.

5. In addition to agents and readmes, maintain an ADR so all decisions and past constraints are readily available to be enforced on new code.

6. Put in as many hard rules as you can about things that need your explicit approval.

Fundamentally, my current job as a "AI-native builder" (cringe) is way more similar to my previous position as EM than anything resembling an IC.

well said, and it's the position I'm taking, using AI to be more considered, more willing to refactor, more prepared to be scrupulous about clean architecture .