On a night in late October, as a new foldable iPhone reached shoppers, a writer with a legal education and only a passing familiarity with code began building something of their own. Thirteen days later, a calendar app shaped for the inner screen of that device had cleared Apple's review, its lines written almost entirely by an AI and its judgments still anchored in a human's hands. It is a small story, but it marks a shifting boundary: the making of software is loosening its hold on specialized training, while the weight of choosing, testing, and answering for what gets released remains stubb
Non-Engineer Ships iPhone App in 13 Days Using Claude Code, With AI Writing Nearly All the Code
AI writes the code; humans bear the responsibility
Mark
Thirteen days from a lawyer-by-training to the App Store. What's the real headline?
Mimi
That the code was the easy part. The AI wrote nearly all of it, but Apple's review bounced the app twice, and the developer still had to decide what to build and test it on a real phone.
Mark
The developer set an approval rule, right?
Mimi
Yes. Claude Code had to report the scope of impact and its plan before changing anything, and wait for a go-ahead. Auto mode also blocked things like rewriting GitHub history.
Luke
Worth noting this is one developer's account of their own project. We have no independent audit of the code or the app.
Mark
The safety findings sound serious, though.
Mimi
Two stood out: shortcuts:// addresses slipping into URLs from notification emails, and events saving to the wrong calendar if the destination was deleted mid-edit. Both were fixed by changing behavior.
Luke
But those were found by Claude and Codex checking code. The developer themselves says AI review is no substitute for an expert audit, and AI-written tests can carry the AI's own misunderstandings. So we don't know what was missed.
Mark
And the cost?
Mimi
Existing subscriptions, $200 and $20 a month, plus the $99 Apple fee. The token use would have been about ¥86,000 at API rates, but nothing extra was charged.
Luke
That ¥86,000 is a conversion, not money spent. And whether the ¥300 theme earns anything is not reported.
Mark
So what does it point to?
Mimi
Lower barriers for individuals, with responsibility staying human.
The Pulse
- A non-engineer with no native iPhone experience races to ship an app in under two weeks, using an AI to write nearly every line of code.
- The AI is kept on a short leash: every proposed fix must be reported and approved before it touches the project, and a separate safety model halts risky operations.
- Building for iPhone from a Windows machine forces a workaround through a remote Mac mini, connecting tools across SSH and Tailscale to reach the simulator.
- Apple's review sends the app back twice, first over a thin developer history and then over a single button label on the calendar permission screen.
- Accessibility checks expose failing red contrast for color-blind users and a year display stuck on the Japanese era code, both corrected before approval.
- The app is approved shortly after midnight on day 13, but the developer's own caveats make clear that post-launch bugs and final decisions still belong to people.
On a night in late October, as a new foldable iPhone reached shoppers, a writer with a legal education and only a passing familiarity with code began building something of their own. Thirteen days later, a calendar app shaped for the inner screen of that device had cleared Apple's review, its lines written almost entirely by an AI and its judgments still anchored in a human's hands. It is a small story, but it marks a shifting boundary: the making of software is loosening its hold on specialized training, while the weight of choosing, testing, and answering for what gets released remains stubbornly, and rightly, with people.
On the night of October 23, the day the foldable iPhone Duo went on sale, a writer who covers smartphones and IT sat down to build an app. The writer had studied law and had only learned the basics of programming, and had made web apps but never a native iPhone one. Thirteen days later, the app went live on the App Store.
The app, "Tsuranari Calendar," is designed for the inner screen of the foldable. Nearly all of its code was written by Claude Code, working through what is called vibe coding: the developer described what was wanted, looked at the app on a real device, and asked for fixes without reading the code. The concept took shape in conversation with Claude, and within a day there was a document of screen designs and requirements. The first build ran on a device by the next night, and the submission to Apple followed three nights later.
The developer kept a firm hand on the process. An instruction file required the AI to report its plan and wait for approval before any change. A separate safety layer stopped risky actions, such as rewriting the project's history on GitHub. Since the developer works on Windows and iPhone apps need a Mac, the setup was improvised: a remote Mac mini, connected through SSH and Tailscale, let the AI build and launch the app in a simulator. The AI also wrote over 300 tests that ran with every fix.
Apple's review was less smooth. The first submission drew a request for more information, and after a long wait the second rejection targeted a permission button labeled "Allow Access." Changing it to "Continue" was enough, and the app was approved on day 13.
With no one reading the code, outside checks mattered. Other AI tools found a loophole that could let links open unintended apps and a path that could save events to the wrong calendar; both were closed. Accessibility testing found text that broke at large sizes, red and blue that failed contrast standards for color-blind users, and a year display that showed only an era digit. Each was fixed.
The cost was modest because existing subscriptions were used, plus Apple's $99 annual fee. The app is free and carries no ads, and a paid color theme came later.
The developer is clear about the limits. Choosing what to build, testing on devices, handling Apple's review, and owning bugs after launch remain human work, and an AI checking an AI can carry the same blind spots into its tests.
On the night of October 23, the day the foldable iPhone Duo went on sale, a writer who covers smartphones and IT sat down and began building an app. By the time it went live on the App Store, 13 days had passed. The writer majored in law, had learned only the basics of programming while studying for Japan's Fundamental Information Technology Engineer Examination, and had published web apps before but never a native one.
The result is "Tsuranari Calendar," a calendar app optimized for the inner screen of the iPhone Duo. Nearly all of the code was written by Claude Code, Anthropic's AI coding tool, using a method known as vibe coding, in which the human does not read the code. The developer described the features wanted and, after looking at the app on a real device, gave instructions for fixes. The concept was worked out in conversation with Claude, and that same first day produced a document laying out screen designs and requirements. By the next night an app was running on an actual device. Three nights after that it was submitted to Apple. From concept to first submission took about 73 hours.
The developer did not give the AI free rein. The instruction document loaded at the start of work required Claude Code to report the scope of impact and its plan before any fix or new feature, and to wait for approval before implementing. Claude Code's auto mode, in which a separate AI model halts operations it judges dangerous, also stepped in; it stopped actions such as rewriting the change history on GitHub.
The setup took some ingenuity. Building iPhone apps requires Xcode on a Mac, but the developer normally works on Windows. Using Claude Code's remote-control function, along with SSH and Tailscale, the developer connected to a Mac mini. By the second day the AI on Windows could build the app on the Mac and launch it in the simulator. The AI also wrote more than 300 tests, and regression tests ran with every fix.
Apple's review sent the app back twice. The morning after the first submission, Apple asked for more information because the submission came from a developer account with little review history. The developer explained the app's purpose and use in English, attached a screen recording from a real device, and resubmitted, then waited nearly five days for a reply. The second rejection concerned the wording on the screen that asks permission to access the calendar. The developer changed a button from "Allow Access" to "Continue," and the app was approved shortly after midnight on day 13. Apple aims to review more than 50% of apps within 24 hours and more than 90% within 48 hours, but cases needing extra confirmation take longer.
Because nobody read the code, the developer leaned on outside checks: Claude Code reviewed its own work for safety, and OpenAI's Codex read the code too. They found that "shortcuts://" addresses, which launch another app, could slip into URLs pulled from notification emails, so the permitted formats were restricted. They also found that if a destination calendar was deleted mid-edit, an event could be saved to a different calendar without permission; the app now declines to save and asks the user to choose again.
Accessibility testing turned up more. Event displays broke at maximum text size, and date visibility fell for some color vision types. Against the WCAG contrast requirement of 4.5 to 1, iOS's standard red measured 3.55 to 1 under normal vision and 2.93 to 1 for people with deuteranopia, so the developer adjusted the red and blue to pass. A further bug showed only "8" as the year under the Japanese era calendar setting; the display was standardized on the Western calendar.
The developer signed up for nothing new. Existing work subscriptions were used as they were: Claude's $200-a-month Ultra plan and ChatGPT's $20-a-month Plus plan. The only added cost was Apple's $99 annual developer fee. By November 8, Claude Code logs showed about 1.76 billion tokens consumed, of which about 1.73 billion were cache reads and roughly 23 million were new input and output. At Claude Opus 5.5 API pricing that would equal about ¥86,000, but the flat-rate plans meant no extra charges. The app is free in its basic form, with no ads and no server of its own; after launch the developer added a ¥300 color theme called "Wa no Iro" (Japanese Colors).
The developer's own caveats are plain: deciding what to build, testing on devices, handling Apple's review, and owning post-launch bugs all stay with people. AI-to-AI checks are no substitute for expert audits, and since the AI writes the tests, its misunderstandings can seep into them. Anthropic, meanwhile, is extending the same division of labor to corporate users through collaboration tools for meetings, whiteboards, and spreadsheets.
Notable Quotes
AI-to-AI checks are no substitute for expert audits, and AI-written tests can carry the AI's misunderstandings.— The developer (paraphrased)
Want the full story?
Read the original at BigGo Finance ↗