Two apps in one year, built by someone who isn't a developer
In May of 2025 I had dinner with a friend who told me that anyone can make an app now. His point was that AI had made it possible for any of us to build something and put it in front of the world, and I'll be honest, I didn't believe him. I've always loved the internet and the idea that you can have a creative idea, put it together, and let people use it, but I'm not a developer. I'm retired Air Force, I've spent the last eight years as a product executive, and while I know my way around roadmaps, sprints and user stories, I had never written code, let alone a full application or shipped one to an app store. It seemed far too good to be true.
A year and a half later I have two apps live in the App Store and Google Play. This is how that happened, what worked, what didn't, and what I'd tell anyone thinking about trying it.
The first attempt
I started with Replit, which at the time was leading the way in letting people build apps with AI. I brainstormed a few ideas and kept coming back to a word game, partly because I like games and partly because my wife plays a lot of them. The idea was simple: what if you had to build a chain of words, where each word connects to the one before it? Talking it through with her and with friends, it turned into a daily puzzle that's the same for everyone, with a starting word and a target word, and each new word has to begin with the last two letters of the word before it. That game became TailSpin.
The first prototype was rough. It was amazing in the sense that I'd never seen anything like it, but I got stuck constantly, or more accurately the AI did. It would go down its own rabbit holes, build things halfway without thinking through how they connected to the rest of the app, and fix one thing by breaking another. By January of 2026 I had a web prototype I could send friends and family a link to, which was far beyond anything I thought I could build, but it barely held together. Oddly, the admin side, where I could see who was playing and how, was more solid than the game itself.
A detour that became the second app
In March of 2026 I read about a popular app for people trying to quit pornography that had left the intimate data of more than 600,000 users exposed, after security researchers had warned the developers for months (404 Media has the full story). These were people trying to better themselves with the help of technology, and that trust was broken in the worst possible way.
By then, tools like Claude Code had come a long way, and I thought I could build a better alternative quickly and properly, with security as the starting point rather than an afterthought. That became Hexys, a privacy-first app for breaking compulsive habits. Everything a user writes is encrypted on their own phone, the key stays on that phone, and nobody on our side can ever read it. If someone broke into our database, all they'd find is scrambled text. I started it in March and had it live by the end of May, and today it helps hundreds of people deal with their habits in a healthy, private way.
I went fast on Hexys and there's polish I still owe it, but it taught me so many things along the way. Writing the code is only part of the job. Getting through app store reviews, setting up the supporting infrastructure, building your website, the analytics, the marketing and everything that ties them together is just as hard, and I honestly don't know how anyone does all of that alone without AI. Most of all, it taught me that I needed a better plan if TailSpin was going to be successful.
Starting TailSpin over
In April, while Hexys was waiting for App Store approval, I set up a brand new project for TailSpin and started from scratch with everything I'd learned. I kept the Replit prototype only as a reference, something to learn from rather than reuse. Work was slow through the summer (working full time and raising two kids will do that), and then at the start of August I committed to it properly. TailSpin launched on Google Play on October 11 and on the App Store on November 16.
What made the difference the second time was how I set up the work. There were four of us, in a sense: me, Claude in a desktop project acting as my architect and reviewer, Claude Code doing the actual building, and Claude Design for the look and feel. From the very first day I asked the desktop project to help write its own instructions, and I was clear that I didn't want something that simply did what I said, or something that went off and did what it wanted. I wanted a partner that would research, push back and argue with me until we agreed.
A typical feature went like this. I'd describe what I wanted, and the desktop project would read the code and our documents, research how others had solved the same problem, and come back with a plan and a list of decisions for me. We'd go back and forth until we agreed, and then it would write the instructions for Claude Code. Claude Code would check those instructions against the code itself and come back with its own plan, which I'd read, add my thoughts to, and hand back to the desktop project to verify. Only once all three of us in that loop had our chance to weigh in and agreed did anything get built, and when it was done, the report went back for one more review before it was finished.
The hardest part was the English language
The game sounds simple, but making the English language fit inside it was by far the hardest part, and I don't believe it would have been possible without AI. TailSpin works from around 80,000 common words that it uses to build puzzles and another 360,000 or so that it accepts when you play. Every one of them had to be checked, filtered and mapped by how its last two letters connect to the next word. Some two-letter endings open the door to more than 20,000 words, fourteen of them lead to exactly one, and twelve lead nowhere at all. Knowing where those dead ends and bottlenecks sit is what makes a puzzle fair or impossible.
On top of that, the game shows definitions for almost every word, which meant combining three separate dictionaries into nearly 345,000 entries, because no single source covered the unusual words players are most likely to want explained. And because a word game should never suggest something offensive, there's a list of words the game will accept if you type them but will never offer you itself. Every one of those decisions was argued through to reach the best experience we could give players.
Where AI fell short, and what fixed it
The longer we worked, the harder it got for the AI to keep up. It would forget decisions we had already made, drift from the plan, or confidently describe something that wasn't true, and sometimes a brand new conversation would struggle more than the one before it for no reason I could see. I'd have to stop and remind it, or catch a mistake it had made with total confidence. It really drove home that AI is a tool, whether you're a founder, a developer or a marketer. It's a powerful one that can help with almost everything, but it's still just a tool, not a replacement for you and not a brain that does your thinking for you.
What fixed it wasn't a smarter model. It was discipline. I kept a short document describing exactly where the project stood, a running log of every round of work, a written record of every real decision and why we made it, and a structure for each piece of work, with its plan, its context and its task list, so that nothing depended on the AI remembering. Keeping those documents current was sometimes tedious, and there were whole days spent just cleaning them up, but it was always faster than untangling a mess.
By the numbers
Since the first commit on April 23:
- 2,339 commits, about 2,000 of them between August and October
- roughly 185,000 lines of code and 160,000 more lines of tests
- more than 16,000 automated tests
- 217 database changes and 19 server functions
- 74 written decision records and more than 600 logged rounds of work
- more than 300 Daily puzzles, plus a separate pool of 500 for head-to-head play
What I'd tell someone starting out
Take the time to set up the structure before you build: your instructions, your documents and the habit of keeping them up to date. It feels slow at first, and you'll be tempted to skip it, but without it things get messy quickly, and a mess is much harder to clean up than it is to prevent.
Don't let the AI simply agree with you, either. The best decisions in both apps came from being challenged. I'd take that a step further and have it bring in the outside world, because you don't just have an AI at your fingertips, you have a research assistant with the whole internet at its fingertips. Before you commit to a feature, a fix or a new idea, have it look at how others have approached the same problem. You and the AI might still disagree, but your decision will be grounded in more than either of your opinions.
A year and a half ago I didn't believe my friend. It turns out he was right, though it wasn't easy, and it wasn't the AI alone. It was the two of us working together.
If you want to see what came out of it, TailSpin is a daily word puzzle you can find at tailspin.gg, and Hexys is at hexys.app.