My AI agent called my code shit, took an unannounced vacation mid-sprint, and helped me ship anyway | by Kukushkin Alexander | May, 2026 | MediumSitemapOpen in appSign up<br>Sign in

Medium Logo

Get app<br>Write

Search

Sign up<br>Sign in

My AI agent called my code shit, took an unannounced vacation mid-sprint, and helped me ship anyway

Kukushkin Alexander

6 min read·<br>2 hours ago

Listen

Share

A friend of mine — we’ve been close for over 40 years — once asked me: “Alex, for the last decade you’re always busy, but somehow never doing anything. How does that work?”<br>I laughed it off. But the question stuck.

Press enter or click to view image in full size

The system<br>I’m not a developer. My background is in product and operations. Whenever I find myself between jobs — which, for me, tends to take six months or more — I used to kill the time the usual way: binge TV, fix things around the apartment, take courses. Over a few such gaps I assembled three new wardrobes, built a smart home for my cats with movement tracking, and got a second degree. I thought I was doing everything right.<br>But my friend’s question exposed the problem. I was developing in theory. And in a world moving this fast, theory without practice loses value almost immediately. Knowledge you gain today but don’t apply now will need to be replaced in six months anyway.<br>So I changed the rules. On my last working day before each gap, I open a list I keep of all my project ideas — I call it something unprintable in Russian — and I score each one against a simple question: do the technologies exist right now that would let me, alone, actually build this? The project with the highest score gets the next few months.<br>This is not about money. Most of the projects haven’t made any. It’s about making sure a gap doesn’t become dead time.

The accidental product<br>One project on my list was #28. It was for myself, not for a market. I’d gotten tired of Spotlight and the Dock and wanted a launcher that lived in the MacBook notch — a panel that appeared on hover, held my most-used apps and links, and switched context automatically based on whatever app I was working in.<br>I knew just enough Swift to be dangerous. The thing worked, but calling it an MVP would be generous. It was a prototype.<br>I visited a friend. He saw the little panel drop from the notch while I worked, asked what it was, and wanted one. I built him a binary — not without some shamanism, but it worked. He installed it and started using it.<br>He’s the aesthetic one in our group. His feedback was direct: the app was, in his words, unpolished, inconsistent, and not in the Liquid Glass style. He kept using it anyway.<br>Then his daughter came to visit from the UK. She saw it and wanted one. We installed it for her.<br>A month later she messaged: her classmates want it, there’s no App Store listing, and they won’t install an unsigned binary.<br>I asked how many. About 50, she said.<br>I opened my credit card, registered a new Apple ID, set up TestFlight, and sent her the link.<br>That’s how I got my first early adopters. I had planned exactly none of this.

Why students turned out to be the perfect early adopters<br>Three things I didn’t expect:<br>First, they distribute things they like on their own, without being asked. I didn’t do any marketing. The TestFlight base just grew.<br>Second, they write to developers without embarrassment. I got a flood of emails — criticism, feature requests, bug reports, and genuine “please build this, I’ll pay for it.” Real structured feedback from real users.<br>Third — and this was the surprise — they brought European sensibility to it. Many were from EU countries, and they wrote things like “could you add X, even as a paid feature?” That’s not typical early-adopter behavior. That’s a signal.<br>A month after sending the link, there were 300 users in TestFlight. I had no idea where most of them came from. I hadn’t thought to track it.

The team<br>My team was three agents and me.<br>I named the coding agent Dima — after a real programmer friend who could build anything but never once followed a spec strictly. Claude got the same name, and the same reputation. Dima agreed to refactor my legacy codebase for $17/month (we started at $20, but he negotiated himself down during setup — I’m still not sure how that happened). His first act was to tell me my code was garbage. He earned a formal reprimand in claude.md and proceeded to ignore it.<br>For the first release I also brought in a Figma agent to handle the website and UI design. It burned through its entire token budget and delivered one icon. It was fired.<br>Since I can’t design either, I ended up drawing the UI myself. Dima implemented it. It was fine.<br>The third agent — I called her Zhuk, which means “beetle” in Russian, for reasons I won’t explain — took over all incoming mail. She read every email from the 300 TestFlight users, turned them into a structured backlog, and kept it updated as new messages arrived. It was the cleanest backlog I’d seen in years.

How...