Yesterday, I built my first Android app. Then, I made two more — three in one afternoon. For one, I literally typed 148 words into my web browser and walked away. Ten minutes later, I had an entire new app on my actual Android phone. I did have to prep that phone by enabling a USB debugging mode and plugging it into my PC, but as advertised, Google’s AI Studio did literally everything else for me.
I typed in words, I hit install, and voilà: an entire working program. I was nearly ready to agree with David, Allison, and Jen: The personal software revolution is here, it’s coming to your phone, there’s a future where the average person can make complicated smart home gadget messes work even with no programming skills. Then, I tried actually using my three apps: a calorie counter and two games. They were kind of bad. And just when I started to enjoy iterating on them, trying to make them better, AI Studio informed me I’d reached my daily limit. I’d have to pay or wait for more.
So yes, there’s still friction, but it’s impressive how much you can do. In one morning, my colleague Stevie Bonifield made a personal workout tracker they found good enough to actually use. Confronted with Gemini’s upsell, my first reaction was: “What if I try paying for a couple months?” I didn’t expect that from Google.
How Google’s AI Studio builds an Android app
On Tuesday, when Google showed off AI coding on a Doom-like game, we joked that I should make MOOD. It would be a Doom-like text adventure game: Modern Online Oratory Dungeon. That was all Google needed to start. When I typed “Make me a Doom-like text adventure game called MOOD, where MOOD stands for Modern Online Oratory Dungeon” into AI Studio, Gemini began typing additional ideas itself, attempting to autocomplete my thought. To start, it typed the phrase “It should feature procedural generation of levels and challenging, turn based combat.”
I didn’t want randomized levels that all feel different — I wanted a classic text adventure where you’re exploring a curated place with a real map. But sure, turn-based combat, and maybe the game could auto-generate the map for me too? Then Gemini suggested it should have “secrets hidden in its rooms,” and “a satisfying progression system,” and more. I mostly nodded along.
Then, it was off to the races. Unlike Claude Code, my colleague Jake points out, Gemini doesn’t make a plan and ask you if you want to proceed. It sprints ahead automatically — though you can inspect the code if you want. One minute later, it already had five design mockups for me. 20 minutes later, I pressed the “Install” button to transfer the game to a Pixel 9 phone.
The writing was terrible, as expected. There were no demons in sight. The entire dungeon consists of just 11 rooms, and you can “win” just by spamming the attack button every single time. You can beat the game in a single minute if you try. Or at least you can now that Gemini helped me fix two showstopper bugs.
Bug fixes can be remarkably seamless, so long as the bug is one Gemini can correctly identify. When I told it that the game breaks during a conversation with “The Whistleblower” because the button that ends the conversation is missing, it spit out a new version of the app right away. I pressed “Install,” the app on my phone restarted itself, and I found myself exactly where I’d left off — only now with the button I needed.
The calorie counter that couldn’t count
My other apps may need more work. The calorie counter decided the best way to estimate calories in a given quantity of food was to ask the paid Gemini API, and I don’t have a paid Gemini API key. When I told it to search for that information in other databases instead, I discovered it vastly understating the number of calories in various kinds of food. But when I told Gemini there’s no way a 16-ounce boba milk tea is just 190 calories, it seemingly did discover the silly error in its own code. It had decided “milk” was a good enough match for “boba milk tea,” and chose low-calorie 1 percent milk to make matters worse. Gemini claims it’ll match more reliably now. Still, my three-ounce serving of Taiwanese popcorn chicken just rang up at 140 calories, and I’m pretty sure it should be double that, so I’ve got work to do.
Super Peach Rescue: Nintendo knockoff or learning tool?
Last and least, I thought I’d better check if Google is still letting people make bad Nintendo knockoffs like my colleague Jay Peters did with Project Genie earlier this year, or whether it’d learned its lesson. With great shame, I present to you Super Peach Rescue. It is a terrible program that crashes as soon as its horrific, one-eyed-floating-alien-of-a-Princess-Peach dares to touch a single power-up block, every single time, and Gemini has not yet been able to figure out why. Also, it’s impossible to clear the game’s second pipe, as Peach simply can’t jump that high.
Still, Gemini did not hesitate to create “a working Super Mario game where I play Princess Peach and go rescue Mario, with all the trappings of a traditional Mario sidescrolling game,” and it kind of did! It even suggested I might want to “Give Peach a variety of classic Mario power-ups like the Super Mushroom, Fire Flower, and Super Star” while I was at it, and labeled the controls “NES System” all by itself. I think I’ll delete this one.
The bigger picture: AI coding and the future of development
This experience highlights a transformative moment in software development. Google AI Studio, powered by Gemini, lowers the barrier to entry for creating mobile applications. Anyone with a concept and basic typing skills can produce a functional app in under an hour. But the gap between “functional” and “good” remains wide. The apps generated are often crude, buggy, and lack the polish that human developers bring. The calorie counter’s reliance on the Gemini API for nutritional data is a prime example of the AI’s limitations: it misunderstood context and miscalculated values.
Moreover, the daily usage cap imposed by AI Studio reminds us that even cutting-edge AI is not free. The free tier allows for limited experimentation, but serious development will require a subscription. This tiered pricing model may push hobbyists toward more open-source or local solutions. Yet, for quick prototypes, learning exercises, or personal utilities, the speed is undeniably impressive.
Beyond the trivial apps, AI Studio could democratize software creation in fields like education, healthcare, and small business. A teacher could build a custom quiz app for a classroom; a nutritionist could design a meal tracker for patients; a gaming enthusiast could experiment with game design without learning complex programming languages. The potential is vast, but it comes with caveats. Security, data privacy, and intellectual property concerns arise when an AI generates code that might contain vulnerabilities or replicate copyrighted material.
Google has taken some steps to limit misuse: the Nintendo knockoff was allowed to exist, but if a developer tried to publish such an app on the Play Store, they would likely face legal challenges. The line between inspiration and infringement is blurry in AI-generated content, and makers of intellectual property, like Nintendo, are already vigilant about protecting their franchises.
Another important aspect is the quality of the generated code. While AI Studio can produce a working app, the code may not follow best practices. It might be inefficient, hard to maintain, or include unnecessary dependencies. Professional developers often criticize AI-generated code for being “spaghetti code” that works but is not scalable. For a personal project, that might be acceptable, but for a commercial application, rigorous review and refactoring are essential.
AI Studio also struggles with complex logic. The MOOD text adventure had only 11 rooms, and the game could be beaten in one minute. The calorie counter made fundamental errors that a human quickly spotted. The Super Peach Rescue crashed on every power-up. These failures underscore that AI, at least in its current form, lacks the deep understanding and intuition that human developers possess. It excels at pattern matching and assembling boilerplate but stumbles when faced with nuanced requirements or novel interactions.
Nevertheless, the pace of improvement is staggering. Just a year ago, generating a functional Android app from a natural language prompt was science fiction. Now it’s a reality, albeit a flawed one. Google’s investment in AI coding tools signals that the company sees a huge market in enabling “citizen developers.” Microsoft is betting on similar tools with GitHub Copilot and Azure AI. The competition will likely drive rapid advancements, making AI coding assistants more reliable, affordable, and capable.
For journalists like me, these tools offer a glimpse into the future where our audience can become creators, not just consumers. The day may come when I can tell a reader, “Hey, here’s a prompt that will build you an app to solve your specific problem,” and it works flawlessly. But for now, I’ll settle for the thrill of seeing my terrible MOOD game actually run on a phone, knowing that the revolution is just beginning.
At least one of the two games I vibe coded was playable, right away, with no sweat from me — unless you count all the psychic damage I feel knowing how many game developers are out of work these days. To be clear, I’m glad the games I vibe coded are bad. While I might justify building a completely free personalized calorie counter because no one will do it for me, my game time is better spent supporting human beings.
Source:The Verge News