#I Stopped Reading Every AI Diff

I used to treat every agent PR like untrusted code.

Read the full diff on GitHub. Build in Xcode. Run the simulator or a real iPhone. Tap through the flow myself. Check tests. Only then approve.

That made sense when the models still missed UIKit lifecycle, SwiftUI focus, or half the scheme setup. On mobile you cannot fake the device. If it does not run, it does not ship.

What I do now

Models got better, and I got pickier about the loop around them.

I still care that it builds, runs on a device or simulator, and that the PR shows proof. Before and after screenshots. A short note on what changed. Tests green. Claude Code babysits the PR: AI review comments come in, it fixes them, pushes again.

Most days I glance the diff, look at the screenshots, skim anything touchy (permissions, Keychain, networking), and approve.

I still dig in when it can hurt users: auth, payments, migrations, anything with private API smell. The doubt did not vanish. It just stopped living on every line of Swift.