Quick note from this week's build: I caught myself doing the exact thing I keep warning clients about.

Was moving fast on a feature with AI doing most of the scaffolding - and it worked, first try, looked clean. I almost merged it as-is. Then I actually traced through the edge cases by hand and found duplicated validation logic in two places doing slightly different things, plus an error path that would've silently swallowed a failed API call in production.

None of that showed up in the demo. All of it would've shown up in a support ticket three weeks later.

This is basically the pattern behind the "AI ships faster but costs more to maintain" data going around right now - PRs are up, but so are incidents per PR. The tools are genuinely good at producing code that looks finished. They're not yet good at knowing what "finished" actually means for your specific system.

My process now: AI writes the first pass, I review every path a user could actually take, not just the happy one. It's slower than pure vibe-coding, but it's the difference between a feature that ships and a feature that survives.