iOS is an interesting domain for vibe coding because the SwiftUI/UIKit conventions are well-represented in training data, but the edge cases — App Store submission requirements, entitlements, push notification provisioning — tend to be where AI-generated code breaks down badly. The "last 10%" problem.
What's your experience with the parts where the AI generated code you couldn't review or understand? That's often where the real costs show up — not in the initial working prototype but in the maintenance cycle months later.
This tension between fast generation and long-term maintainability is central to the Agile Vibe Coding Manifesto (https://agilevibecoding.org), which tries to formalize how Agile teams should approach AI-assisted development. Its ninth principle: "Generated systems remain readable, maintainable, and evolvable by humans, regardless of how they were produced." On native mobile especially, that's harder than it sounds.
Would love to hear what worked and what didn't.
