Production lessons from building local-first, AI-powered iOS apps.
It connected to Claude Code on the first try, so I shipped it. Then I read the official 13,749-line SDK and found five bugs in my 170 lines — four reproducible in one shell pipe, one that makes a client hang forever.
No SDK, no framework — the raw HTTP traffic behind an agent loop, five pitfalls that will bite you, and a prompt-caching measurement that took uncached input from 7,534 tokens to 6.
What transfers from mobile development, what's genuinely new, and a four-phase learning path to agent engineering — documented in public, with real code and real costs.
Why we removed every third-party SDK from our apps, what we use instead, and how it affects App Store review and user trust.
How we route between Haiku and Sonnet, stream responses in SwiftUI, and keep API costs under control in a consumer iOS app.
Hard-won lessons on conflict resolution, migration gotchas, and sync reliability from shipping two production apps on SwiftData + CloudKit.
A week-by-week breakdown of what it actually takes to go from concept to a live App Store listing — design, development, review, and the surprises in between.
Practical guidance on finding, vetting, and working with remote iOS developers in Europe — contracts, time zones, rates, and red flags to watch for.
A clear-eyed look at the real costs of integrating AI into an iOS app — API fees, on-device vs cloud trade-offs, development time, and ongoing maintenance.