Claude Code writes great Swift, then loses the plot around phase four. Eight opinionated skills + a 347-line playbook keep it honest: scoped PRDs, session-sized phases, fingers-on-simulator verification, App Review survival. No vibes. Rails.
launch week price, $29 after ✍️ P.S. Claude's new Fable 5 built this entire product + site in ~3 hours, as a "$1k in 24h" experiment. The clock is real ⏰## Phase 3: Persistence, entries survive relaunch Verify (in simulator): 1. Cold-launch. Add three entries. 2. Force-quit. Relaunch: all three present, newest first. 3. Swipe-delete one. Relaunch: exactly two remain. 4. Dynamic Type to XXL: nothing truncates. Done when: data survives force-quit, zero fixture data left in the target.8skills inside
Raw idea in, developer-ready PRD out. One-sentence-job scope test, screens map, Swift data model, and a handoff block the agent can't drift from.
PRD in, build plan out. Every phase fits ONE Claude Code session, sized against a real context budget, verified with fingers, not "it compiles."
Architecture before features: @Observable state, route enums, a persistence decision table, mandatory empty + error states. The decisions that cost a rewrite when you make them in phase seven.
JSX, HTML, or a screenshot into idiomatic SwiftUI, plus the list of web patterns you should never port to iOS.
Hunts what AI-generated Swift actually breaks: strict-concurrency violations, phantom features, state duplication, crash-on-empty paths.
Every App Store Connect field, ready to paste. How to pack the 100-character keyword field, and an above-the-fold description that earns the tap.
The 5-8 shot narrative with six-word overlay headlines. The first three screenshots decide the install; plan them like it.
Signing, TestFlight, privacy labels mapped from what your code touches, and the App Review rejection traps with pre-submission fixes.
FridgeFriend is a real mini-app taken from a three-sentence idea through PRD, build plan, and full SwiftUI source to an App Store listing. The whole thing ships in the free repo, so the quality bar is public.
## Claude Code Handoff iOS target: 17.0 Persistence: SwiftData, one container Dependencies: none (hard constraint) Constraints: no force-unwraps, every screen handles empty data, Swift 6 concurrency clean Consumed by build-phases. Sized: 4 phases.
Phases are sized against a real context budget: file counts, framework introductions, project-level changes. Mid-phase compaction is treated as a sizing bug, not weather.
Every phase ends with a numbered script of finger actions in the simulator and the exact observations you must see. "Tests pass" is banned as a milestone.
Doesn't earn its keep? One message inside 14 days, full refund.
It's a system with opinions that cost real debugging time to earn: context budgets per phase, a persistence decision table, an audit list built from how AI-generated Swift actually fails. The free repo exists so you can check whether those opinions hold up before paying.
An instant Ko-fi download: a zip of plain markdown. Unzip it, drop the skill folders into ~/.claude/skills/, and they trigger automatically in your next Claude Code session. No license key, no account, no course portal — you own the files. Updates come through the same download link.
It's written for Claude Code's skill format, and the triggers are Claude Code specific. The PRD template, phase budgets, and audit checklists port to any agent as plain markdown, but automatic invocation is Claude Code only.
Two of the eight skills and the entire worked example are free and MIT-licensed. The paid pack funds maintenance and updates, and it's priced under one hour of a developer's time.
The skills are plain markdown with a small YAML header — the least breakable thing in this stack. When something meaningful shifts, 1.x updates ship through the same download link, free. And since you own the files, you can edit them the day anything changes, without waiting on me. There's no runtime here to rot.
iOS 17 and up as the baseline, Swift 6, SwiftUI, SwiftData. Notes are included where newer APIs change the recommendation.
Message me on Ko-fi within 14 days and you get a refund. The files are markdown, so this policy runs on honesty in both directions.