Turn your worst AI-build nightmare into two or three rules you will actually follow. Not a pledge, not a values statement: tripwires specific enough that a tired, behind-schedule version of you cannot argue past them.
Your AI-assisted build has gone badly wrong three months from now. Not “something broke”: the specific, ugly version. The deleted table. The hallucinated fact that shipped in a changelog. The one-line fix that quietly became a rewrite.
Now walk it through. Short answers are fine.
Who tells you, and how?
What specific action, by the agent or by you approving its output, actually caused the damage?
What made it look fine? What check got skipped or waved through?
Hours, dollars, data, trust. Be concrete.
Was there a warning sign you noticed and brushed past?
Where was the latest point this could still have been stopped?