I used AI to build an app in 6 hours, but it still took me a full year to get it listed.

A British engineer spent 6 hours using AI to build a habit-tracking app, yet took more than another year before he finally dared to list it for real on the App Store. The code was a mess, hidden implicit errors were lurking in the cloud sync, Apple’s new operating system kept failing in sequence—making him realize that artificial intelligence can take you to 80%, and the remaining 20% is what real engineering is.

(Background recap: an engineer’s confession: AI made me almost forget how to write code, and the industry skills feel hollow—raising quiet concerns.)

(Extra context: Is the human brain better than machines? Tens of thousands of people on websites pretend they’re ChatGPT)

Table of contents

Toggle

• Why do it yourself—what AI delivered in six hours
• Collapsing details, iOS 26 calls out the AI
• Hidden worries of the industry becoming hollow

Six hours, one weekend. In March 2025, British software engineer Alex Hyett built an app that looked usable, but it took him more than an additional year before he truly felt brave enough to put it up on the App Store.

Alex Hyett revisited that experience in a video on his personal account: what was originally supposed to be “AI helping you do the work while sipping a cocktail on the beach” ended up becoming the software engineering exercise he poured the most effort into. The gap was so big that it highlighted the part of the line “AI can already replace developers” that is most often ignored.

Why do it yourself—what AI delivered in six hours

Habit-tracking apps are generally viewed as one of the easiest kinds of apps to build, second only to to-do lists. As a result, there are hundreds of options on the market. Alex and his wife tried dozens in succession, but none truly worked.

The requirements he listed weren’t complicated: the list needs to be long enough that you don’t have to flip through pages; data shouldn’t go to the cloud or exist on anyone else’s servers; you can use iOS 18’s generated emoji as icons; the target can be set to daily, weekly, or monthly—or set no target at all and simply count occurrences; and it also needs to let you write notes and add reminders.

The free Streaks uses pagination instead of a long list, the icons aren’t good enough, and it can’t count only occurrences; HabitKit costs £1.99/month, £11.99/year, or £29.99 as a one-time purchase; Grit is more expensive—£9.99/month, £29.99/year, and the buyout price is £44.99. In his view, if you stack these subscription fees, by the end of a year you’re almost at the cost of a new console game.

So he decided to do it himself. At the time, Cursor still had free quotas and Claude Code hadn’t been released yet, so it wasn’t the later kind of proxy-like fully automated development. Instead, he broke the work into a bunch of small tasks, then manually went back and forth with the AI for dialogue and testing each item. He had never written Swift before; everything had to be learned from scratch. The app later got named HabitTed, and his wife even designed a little bear logo.

He started with a bare-bones “hello world” screen, then added features step by step: habit lists, adding habits, editing habits, and so on.

In about a weekend—roughly 6 hours—he had a working app: it could add habits, set custom icons, tap to mark as done, and it also supported different goal options and iCloud sync. It sounded like the perfect demo of an AI myth—until he opened the code and took a look.

Usually, once a project is finished—even if it gets stuck in the middle and takes longer than expected—afterward he still feels a sense of achievement, thinking he learned a little more. But for these 6 hours, he admitted afterward that he barely learned anything about Swift. Every time an error popped up, all he did was copy-paste the error message to the AI and ask the AI to figure out a solution. He complained that the AI wasn’t smart enough, while also complaining that he couldn’t do anything.

Collapsing details, iOS 26 calls out the AI

Having a working app doesn’t mean the code is good. A single view easily pushed toward 1,000 lines. The Xcode compiler couldn’t handle it, constantly throwing warnings telling him to split things up. Buttons that looked identical on different screens used completely different styling code. He didn’t rely on AI to clean it up anymore; instead, he manually refactored everything, cutting each view down to under 100 lines.

But only after refactoring did he discover there were more problems. iCloud sync “looked like it was moving,” but in practice, as soon as you deleted the app and reinstalled it, the data disappeared entirely. For the statistical numbers in the detail page, every time he entered the page he had to rerun loops over all habits; later he changed it to calculate at the moment of check-in and store it as properties. The data model itself was overly complex, and the naming was weird, forcing him to make a whole separate migration plan—otherwise, when he changed the schema later, the old data would break.

What truly exposed the AI was iOS 26. When he enabled the accessibility option for “reduce transparency,” the app immediately broke. In dark mode, when he opened a habit, the title bar turned into a white background with white text, and you couldn’t see anything. The AI couldn’t help at all, because iOS 26 had just been released. He told the AI it was the latest version, but the AI answered that he must have remembered it wrong. In the end, he solved it by reading other developers’ blog posts, and the transparency bug wasn’t fixed until Apple patched it in 26.1.

This experience quietly lines up with an earlier METR study. From February to June 2025, METR found 16 senior open-source developers and ran a randomized controlled experiment on 246 real tasks. The main tools were Cursor Pro paired with Claude 3.5 and 3.7 Sonnet. The result: using AI actually made them 19% slower, but afterward the developers themselves rated that AI made them 20% faster. It feels like there’s a sizable gap between reality and perception.

Hidden worries as the industry empties out

Alex Hyett’s conclusion is blunt: AI can take you to 80%, but the remaining 20% will eat 80% of the time—and if you can’t understand what the AI wrote, it’ll take even longer.

He added one more point: AI is stronger than it was a year ago, but it’s also more expensive—and it doesn’t solve a fundamental problem. Using AI won’t make you learn any skill yourself. If you can’t get anything done when Claude crashes, that means you’re relying on AI too deeply.

He also observed that the whole industry is changing, and external data points to the same direction. Stanford Digital Economy Lab data shows that by July 2025, the number of software developers aged 22 to 25 was nearly 20% lower than the peak at the end of 2022. No companies want to hire newcomers—between senior engineers and AI, they’re enough.

Senior engineers themselves are split into two camps. One camp hands everything over to AI; the “feel” of writing code slowly disappears without them noticing—like boiling a frog in warm water, until they realize it’s already too late to go back. The other camp writes prompts every day until work burnout, not wanting to touch it anymore. He worries that once the AI bubble truly bursts, there will be very few developers left who are willing—and able—to clean up the mess.

He also admits that without AI, he wouldn’t have started building this app at all. But in the other direction, without relying on AI completely, he could have learned more—and wouldn’t have had to spend months cleaning up the code…