Five hours, seven Xcode targets, 45 minutes of hands-on time

4:05 PM to TestFlight, on four platforms

Daniel Hayes Smith made his first commit at 4:05 PM. About five hours later, he says, a visual timer app called ClearTime was running in TestFlight builds on Mac, iPhone, Apple Watch and Apple TV. Not four mockups, not a web app wrapped four times — four native Apple apps written in Swift and SwiftUI. His own hands-on time, by his account, was roughly 45 minutes, and he describes himself as brand new to app development.

These are his numbers and his receipts, not independently verified ones. But the thread is unusually specific about the parts most "AI built this" posts leave out: which model, which permission mode, and where the design decisions actually got made. That specificity is what makes it worth taking apart.

The motivation is worth stating because it shapes the product. Smith says he has ADHD and what he calls time blindness — ten minutes and twenty minutes feel identical until he's late. He wanted a solid disc of colour that shrinks as time passes, readable at a glance on whatever screen was nearby. He'd been putting it off for a while. Then, he writes, he developed SVT at 48, hit 220 beats per minute, spent a week in the ICU and had an ablation. He is now off his ADHD medication until doctors clear it. He stopped putting the app off.

The design step is where the time was saved

The instinct is to credit the coding agent. Smith credits the handoff. Before any code, he used Claude Design to turn decisions that existed only in his head into files a coding agent could follow without stopping to ask what he meant.

What came out of that was not a folder of pictures. According to the thread it included a design brief covering the dial, states and accessibility; colour tokens, typography, sizing and motion rules; separate reference boards for each platform; and — the load-bearing item — the dial geometry expressed as code, with test vectors and "golden" reference files. A golden file is a stored correct output: every platform draws the dial, then compares its result against the reference. If watchOS renders the shape three degrees off, that's a test failure, not a taste argument.

There was also a build guide that named which document wins when two disagree. If a screen didn't match its reference board, the board was right. If one platform drew the dial differently, the shared geometry won. That sounds like bureaucracy; it's actually the mechanism. An agent stuck between two contradictory sources will invent a third answer. This one had somewhere to look.

The proof that the design was doing real work: the original technical handoff called for Tauri on desktop and Expo on mobile — two cross-platform frameworks. Smith asked Claude Code whether that was overkill, and says the agent argued for native Swift and SwiftUI first. The stack changed before the first app target existed. The design survived intact, because it was never tied to a framework.

The settings, in plain language

Smith says the implementation ran on Claude Code with Claude Opus 5.5 at High effort and Bypass permissions — meaning the agent acts without pausing for approval on each file write or command. He connected Xcode's MCP bridge, which lets the agent operate on the real project rather than guessing from snippets pasted into a chat. The result was one Xcode workspace with seven targets — macOS, iOS, watchOS, tvOS, plus widgets, Watch widgets and TV Top Shelf — over a single shared package holding the timer engine, dial geometry, colour math, UI and tests. For planning and pressure-testing the idea, he used ChatGPT Sol 5.6 at Extra High.

One thing to flag: the thread promises to explain what makes the agent stop and ask, and the excerpt cuts off mid-rule after "Solve implementation problems on your own." On bypass permissions, that escalation boundary is the most consequential setting in the whole workflow, and it's the one detail the published text doesn't finish.

Questions You Should Be Asking

• Does "working TestFlight build" mean feature-complete, or compiled, signed and launching? Those are very different five hours.
• Who wrote the tests the golden files check against — and would a wrong reference file have been caught, or faithfully replicated across all four platforms?
• If your team runs an agent on bypass permissions, what exactly is it allowed to touch, and what does it escalate? Write that list down before the first run, not after.
• How much of the 45 minutes is transferable? Smith made the taste calls once, up front. Does anyone on your team currently do that, or do they make them mid-build?
• What happens at version two — when requirements change and the authoritative documents no longer describe the shipped app?

What To Watch Next

The signal is maintenance. A five-hour first build proves the handoff worked once; it says nothing about whether a first-time developer can debug a seven-target workspace he didn't write when Apple ships an OS update. Watch whether ClearTime gets a second release that changes real behaviour — and whether the design brief and golden files get updated alongside it, or quietly go stale while the code moves on.