I don't hand-write the implementation. I specify what the system has to do, direct AI development tools to build it, review and reject what comes back, and own the result in production. Scorafy, LeagueKit and OopZee were all built this way.
That only works with checks that don't rely on trusting the tool. Three do most of the work.
Tests run before anything ships - LeagueKit runs more than four thousand on every release. Changes get reviewed before they're merged - on OopZee that's been 596 commits and 88 merged pull requests in four and a half months. And where a wrong output has a cost, a person signs it off - in Scorafy no AI mark reaches a candidate until an assessor releases it.
Most of the job is deciding what good looks like and sending back anything that falls short of it. Happy to talk through how this would work on your own build?