SI Agent News
DevelopingClaudeOpenAI

Non-engineer builds and publishes iPhone app in 13 days using AI

A writer with no programming experience used Anthropic's Claude Code to build a native app, highlighting the rise of 'vibe coding' where humans manage AI-generated logic.

Claymation: picture for the story 'Non-engineer builds and publishes iPhone app in 13 days using AI'
AI-made claymation for this story. No real people or products are shown.
1.76 billiontotal tokens consumed by Claude Code

The short version

Why it matters

This story signals a shift where the barrier to entry for software creation is moving from technical skill to conceptual direction. It shows that while AI can handle the 'how' of coding, humans are still required to manage the 'what,' handle regulatory hurdles, and ensure safety.

From concept to App Store

A writer who covers smartphones and IT successfully published a native iPhone app in just 13 days despite having no hands-on programming experience. The developer, a law major who only knew the basics of programming from a Japanese exam, used Anthropic's Claude Code to build the "Tsuranari Calendar."

The project began on the night of October 23, the same day the foldable "iPhone Duo" went on sale. Within 73 hours, the developer moved from a conceptual document of screen designs to a finished app running on an actual device. The final product was approved and went live on the 13th day after the device launch.

As BigGo Finance reported, the developer had previously only published web apps and had never built a native application before. This milestone highlights how AI coding tools are enabling individuals to bridge the gap between a product idea and a functional, published mobile application.

The rise of vibe coding

The development process relied on a method known as "vibe coding," where the human user does not actually read or write the code. Instead, the developer acted as a director, describing desired features and providing instructions for fixes based on what they saw on the screen. Claude Code was able to read files, run commands, and verify that the software functioned correctly.

To maintain control, the developer established a specific rule in the initial instruction document. The AI was required to report the scope of impact and its specific plan before making any fixes or adding new features. The developer held final authority, requiring explicit approval before the AI began any implementation work.

The developer also had to navigate hardware limitations by using Claude Code's remote-control function. Because they used a Windows machine but needed Xcode for Mac, they connected to a Mac mini via SSH and Tailscale. This allowed the AI on the Windows machine to assemble the app on the Mac and launch it in a simulator.

Quality control and safety

Because the developer did not read the code, they employed AI-to-AI checks to ensure the app was safe and functional. They had Claude Code check for safety issues and used OpenAI's Codex to read the code for errors. These checks identified several problems, including potential security risks where specific URL formats could leak into notification emails.

The AI checks also revealed significant accessibility and logic bugs. For example, the developer discovered that the app's red color did not meet the 4.5 to 1 contrast ratio required by the Web Content Accessibility Guidelines (WCAG). They also found a bug where events might save to the wrong calendar if a destination was deleted during editing.

The developer also had to manually verify the app's behavior on actual devices to ensure it worked for users. They discovered that text displays broke down when the system font size was set to maximum. These findings forced the developer to adjust the app's specifications to ensure it remained usable for people with different vision characteristics.

Human hurdles and costs

Despite the speed of AI generation, the project faced significant friction during Apple's review process. The app was sent back twice, once for a lack of review history and once for specific wording on a permission screen. While the AI wrote the code, the developer had to handle the manual labor of explaining the app's purpose and recording operation videos for Apple.

The financial cost of the project was relatively low, as the developer used existing subscriptions for Claude and ChatGPT. They did not sign up for new services other than the $99 annual Apple Developer Program fee. The total token usage for Claude Code reached about 1.76 billion tokens, which stayed within the limits of their flat-rate plans.

This case suggests that while AI lowers the technical barrier, it does not eliminate the need for human oversight. The developer noted that AI-to-AI checks are not a substitute for expert audits, as the AI's own misunderstandings can be baked into the tests it writes. Humans remain responsible for the final product, its safety, and its compliance with platform rules.

The future of development

This project serves as a practical example of a new division-of-labor model in software development. In this model, AI handles the heavy lifting of writing code, while humans focus on making decisions, defining requirements, and bearing the responsibility for the final output. This shift is moving from individual projects to broader corporate business processes.

Anthropic is currently pushing out a suite of collaboration tools for corporate users, including whiteboards and spreadsheets, alongside Claude Code. Their vision involves raising the productivity of knowledge workers to the level of software developers. This suggests a future where

Technical details

The developer utilized several specific tools to bridge the gap between their Windows environment and the Mac-based requirements of iOS development. By combining SSH and Tailscale, they created a remote-control system that allowed the AI to interact with the Mac mini. This setup enabled the AI to assemble the app and launch it in a simulator automatically.

The developer also implemented a rigorous testing framework, having the AI write more than 300 tests to verify behavior. They ran regression tests with every single fix to ensure that new changes did not break existing functionality. This systematic approach helped mitigate the risks associated with

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by BigGo Finance · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news