Author: Sean Hollister, the Verge; Compiler: Peggy, BlockBeats

Google AI Studio is bringing AI programming to a more intuitive stage: users no longer just let the model "write code", but can directly generate an Android application in natural language and install it on a real phone in a few minutes. From entering prompt words in the browser, to Gemini automatically generating code, designing interfaces, fixing bugs, and then to the application appearing on the device, the threshold for software development is being further lowered.

The Verge author Sean Hollister recently experienced the "prompt to phone" capability of Google AI Studio. He built three applications in one afternoon, including a text adventure game, a calorie calculator and a Mario-like game. There was almost no need to write code himself, and some bugs could be quickly fixed by continuing the conversation. This experience shows that AI programming tools are moving from development environments to consumer scenarios closer to ordinary users.

This is also the most imaginative part of the "personal software revolution". In the past, ordinary users could only wait for developers to make general products; now, they may be able to temporarily generate a fitness tracker, calorie calculator, or even a simple game based on their specific needs. For Google, this is not only a demonstration of AI programming capabilities, but also a new entrance for Gemini into the mobile phone, developer ecosystem and subscription revenue.

But this experience also shows that there is still a clear gap between AI-generated applications and true maturity. It can quickly create "running" programs, but it may not be able to create reliable, accurate, and easy-to-use products: the game narrative is rough, the mechanism is weak, calorie data will be seriously misjudged, and Mario-like mini-games will even crash repeatedly. More complex issues also include copyright boundaries, data sources, product judgment and long-term maintenance capabilities.

What really deserves attention is not whether AI can replace developers, but that the starting point of software production is changing. Google has proven that it is becoming a reality for ordinary people to use prompt words to create mobile applications; but from "generating an application" to "making a good application", human professional experience, aesthetic judgment and continuous iteration are still required. AI can significantly speed up development, but the last mile of software quality cannot yet be delivered.

The following is the original text:

Yesterday, I made my first Android app. Then I made two more styles – three in total in one afternoon.

In one of the apps, I actually just typed 148 English words into a web browser and then walked away. Ten minutes later, there was a new, complete app on my real Android phone. Of course, before that, I do need to make some preparations for the phone: turn on USB debugging mode and connect the phone to the computer. But other than that, AI Studio pretty much literally does all the work for me, just as Google advertises.

I entered the text, clicked install, and voilà, a complete program appeared that worked. At that moment, I was almost ready to agree with David, Allison, and Jen: the personal software revolution has arrived, and it's coming to your phone. In the future, ordinary people may be able to make complex smart home device systems run even if they do not have programming skills.

Then I started actually using these three apps: a calorie counter, and two games. As a result, they didn't perform very well. And just when I started enjoying iterating and trying to make them better, AI Studio reminded me that I had reached my daily usage limit. Next, I either pay or wait for the credit to be restored.

So, the friction still exists. But it’s undeniable that what individuals can accomplish today is pretty amazing. That same morning, my colleague Stevie Bonifield also built a personal fitness tracking app that he thought was good enough to actually be used. Faced with the paid upgrade prompt that popped up from Gemini, my first reaction was: "Do I want to pay for a trial for a few months?" This was not the reaction I originally expected to have on Google products.

How Google’s AI Studio builds an Android app

On Tuesday, when Google showed off using AI to write a Doom-like game, we joked that I should make a game called MOOD. It will be a text adventure game similar to Doom, and MOOD stands for "Modern Online Oratory Dungeon".

This information alone is enough for Google to get started. When I typed into AI Studio: "Help me make a Doom-style text adventure game called MOOD, MOOD stands for Modern Online Oratory Dungeon", Gemini began to automatically add more ideas in an attempt to continue my creativity. It first entered a sentence: "This game should have procedurally generated levels and challenging turn-based combat."

Gemini is trying to autocomplete my app ideas. Image source: Google

I didn't want randomly generated levels that were completely different every time - I wanted a classic text adventure game where the player explores a designed environment with a real map structure. However, the turn-based combat is acceptable. Maybe, this game can also let AI automatically generate maps for me?

Subsequently, Gemini suggested that the game should have "secrets hidden in the rooms", "a satisfying growth system" and more settings. Most of the time, I just nodded along with its train of thought.

Before I let it officially start writing code, the final prompt word is this:

Image source: Google

Then it officially started running wildly. My colleague Jake points out that unlike Claude Code, Gemini doesn't formulate a plan and then ask you if you want to proceed. It will advance automatically - of course, you can always view the code it writes if you wish.

One minute later, it has generated five sets of design prototypes for me:

Twenty minutes later, I pressed the "Install" button and the game was transferred to a Pixel 9 phone.

As expected, the copy was poorly written. There is no trace of demons at all in the game. There are only 11 rooms in the entire dungeon, and players can "clear" it directly by pressing the attack button all the way. If you play hard enough, you can finish it in a minute. At least it does now - and before that, Gemini helped me fix two bugs that were serious enough to render the game unplayable.

The following is the actual effect of MOOD:

I wasn't too surprised to find that Gemini's promise of a "compelling narrative with branching dialogue options and multiple endings" was boiled down to a simple branch at the end of the game: I could defeat "Core Orator"—an AI that somehow turned the internet's rage into corporate profits—by attacking it, merging with it, or entering a backdoor password.

In addition, the game will take the initiative to directly expose all the "secrets" originally promised to players: it turns these contents into glowing buttons, and players do not even need to enter any text. When you encounter a glowing treasure chest, the game takes pains to remind you that it's actually a Mimic—the classic Dungeons & Dragons monster disguised as a treasure chest.

Not only did it clearly warn you to "check the treasure chest at your own risk," it even directly marked it as an enemy and refused to let me leave because the system prompted: "A hostile "Clickbait Mimic" is blocking the way!"

Speaking of which, MOOD will even directly tell you the backdoor password to unlock the hidden ending when you need it.

However, the bug repair process can be surprisingly smooth, provided that the bug is a problem that Gemini can correctly identify. When I told it that the game would freeze while talking to "The Whistleblower" because the button to end the conversation was missing, it immediately spawned a new version of the app. I pressed "Install" and the application on my phone automatically restarted. When I re-entered the game, I found that I was staying where I just left - only this time, the button I needed had appeared.

My other apps may need more polishing. The best way for that calorie counter to determine the calories of a certain food is to call the paid Gemini API, and I don't have the paid Gemini API key. When I asked it to search other databases for relevant information instead, I found that its calorie estimates for a variety of foods were grossly underestimated.

However, when I told Gemini that a 16-ounce cup of bubble tea couldn't possibly only have 190 calories, it did seem to find that stupid bug in its own code. It previously believed that "milk" was enough to match "boba milk tea". To make matters worse, it also chose low-calorie 1% low-fat milk as the basis for its estimate. Gemini claims it will now make more reliable matches.

But even so, it just counted 140 calories for my 3-ounce serving of Taiwanese Salt Crispy Chicken, and I'm pretty sure the real calories are at least twice that number. So, the app clearly needs work.

Last, and least importantly, I felt compelled to test whether Google still allows users to make those terrible Nintendo knock-offs, like my colleague Jay Peters did with Project Genie earlier this year, or whether Google has learned its lesson.

With deep shame, I present to you-"Super Peach Rescue":

Three screenshots showing a really bad game. Image source: Google

This is a terrible program. Princess Peach in the game is made into some kind of terrifying, one-eyed floating alien image, and whenever she dares to touch any prop block, the game will immediately crash - every time. So far, Gemini hasn't been able to figure out why.

In addition, the second water pipe in the game is impossible to pass because Princess Peach cannot jump that high at all.

Thank you, Google... Is this what you mean by "help"? Image source: Google

Nonetheless, Gemini had no qualms about producing such a game. My brief was: "Make a working Super Mario game, where I play Princess Peach to rescue Mario, and have all the elements of a traditional Mario side-scroller game." In a sense, it did.

It even proactively suggested that I might be able to "add a series of classic Mario props to Princess Peach, such as super mushrooms, fire flowers, and invincible stars." It also labels the operating method as "NES System" on its own. I think I'll delete this game.

At least one of the two games I made using vibe coding was playable from the get-go with almost no effort on my part - that's if you don't count the psychological trauma I feel when I think about how many game developers are out of work these days.

Let me be clear: I'm actually glad that the quality of the games I vibe coded was poor. I might still be able to defend myself against a completely free calorie counter customized just for me: After all, no one would make such a tool just for me. But if it's a game, I'd rather spend my time supporting real human creators.