A Weekend of Vibe Coding ⭐

At some point, I will move past the magic that is vibe coding new apps, not just with Google AI Studio, but with other tools as well. But today is not that day.

This is all Google’s fault.

As just one of what felt like 1000 announcements at Google I/O last week, the online giant announced one developer-related advance that caught and held my attention more than most:  You can now build native Android apps based on Kotlin and Jetpack Compose using Google AI Studio, a tool that previously let developers prototype app user interfaces with web app backends. And as the I/O keynote continued past that bit, I found myself getting distracted. So I checked it out.

I wrote about my initial results in I Just Vibe Coded a Markdown Editor With Google AI Studio ⭐. Long story short, the Android app creation capability wasn’t available immediately, but I was more successful than expected in creating what turned out to be a series of web app-based Markdown editors. And so the next day, I tried again. As described in And Now I’ve Vibe Coded a Native Android Markdown App ⭐ – Thurrott.com, I successfully created a few native Markdown editor apps for Android too.

I did not stop there.

Android Studio + Gemini

Emboldened by this success, and by my earlier work using Stardock Clairvoyance and Anthropic Claude Code to finish my multi-document/multi-tab WinUIpad app, I turned to Android Studio, Google’s IDE for Android app developers. I knew that it offered an integrated Gemini pair-programming assistance feature, similar to GitHub Copilot in the Visual Studio IDEs. And given that AI Studio is based on Gemini and that Google has specifically trained Gemini with Android developer documentation and best practices, I figured this might work well too. “This” being me, someone who does not know Kotlin (a programming language) or Jetpack Compose (Google’s preferred framework for Android app UIs), somehow vibe-coding a real Android app using Android Studio.

Into the abyss I fell.

I’ve never been a fan of Android Studio. This thing has always felt old-fashioned and crude compared to Visual Studio. But it’s also improved a lot in recent years, and it at least presents a modern UI face to the world now. After a bit of noodling around the UI and Googling, I figured out how to connect Gemini to Android Studio through my Google AI Pro subscription. Then, I created a new Android app project with an empty activity, opened the Agent sidebar in the IDE, and asked it to create a native Android app version of a Markdown editor.

It worked on this for a long time. A very long time. It created a lot of files and made numerous changes, but eventually, it just started spitting out garbage. It was like Gemini, which had been close to perfect in AI Studio, had lost its mind. I tried to pick it up repeatedly, but each time I only got garbage responses, jumbles of nonsensical text and numbers. When I tried it the next morning, it appeared that I had run into some kind of quota limit. So I gave up on this for a bit.

After experimenting with AI Studio further on other things, it occurred to me that I could perhaps just plug Anthropic Claude code into the unfinished Android Studio project. After Googling the issue and confirming that, I added the Claude Code plug-in to Android Studio, signed in to my Anthropic account, and installed the Claude Code command line tools as instructed.

Claude Code appears in a bottom-bottom pane rather than the same Agent pane that Gemini uses on the right. Whatever. I asked Claude to figure out the problems with this app and it never really got there either. I was starting to feel like my pre-Clairvoyance issues were back, where some AI could cause a bunch of problems, I’d repeatedly ask it to solve them, it would then spend a lot of time trying to fix its own problems, all while discussing this with me more than I wanted, and then I would run into some quota issue and be unable to continue. Unless I gave it my credit card information.

No thanks. Not helping matters, I’m running Android Studio on Windows 11 on Arm, which doesn’t support using the virtual machines you can create with the IDE’s built-in emulator capabilities. So I had to connect an Android phone to the PC via USB to test apps, apps that weren’t even compiling let alone running. So I gave up for a bit.

But this gave me another idea.

Claude Code + Windows App SDK/WinUI

As I was working on WinUIpad this past spring, I kept thinking about what it would mean to turn that app, or, more ideally, a simpler single-document version of that app, into a Markdown editor specifically customized to my preferences. This is why I made web app and Android app versions of a Markdown editor with Google AI Studio, of course. But I was hoping to take my codebase or even start from scratch and make a Windows app version too.

Given my now somewhat colored recent successes with vibe coding, I thought it may be worth starting over a just prompting an AI to do this from scratch. I started with Claude Code this time, figuring Gemini was more Android/web-focused. And I used a variation on a now familiar prompt:

I would like a native Windows desktop app using the Windows App SDK and WinUI 3 that looks and works like the Markdown editor app Typora. It should include the ability to open and save files, use a WYSIWYG editing view by default, and support all the normal menus and keyboard shortcuts as per a word processor. The design should be minimalist, with no sidebar, but there can be an optional toolbar for command buttons. It’s just an editor, not a file manager.

This took a long time.

Claude said it would create a Windows App SDK app with an embedded WebView2 control hosting the Toast UI Editor, which is not an approach I’m sure I would have ever considered. So there’s a normal XAML app shell, but all the good stuff happens in JavaScript via that hosted editor. After a long wait and a lot of back and forth, it told me that the project built successfully, and it gave me a file system path I could use to run the app.

It would not run. And while it’s not worth going through all the permutations of my troubleshooting, the short version is that I loaded this project into Visual Studio 2026 Community, found and had Claude fix some .NET and Windows App SDK versioning issues, and figured out but could not fix the fact that this thing was made for x64 and could not be switched to a native Arm version that would compile and run. But I did finally get the x64 app up and running just fine. And it’s actually pretty good.

Most of the issues I have here are issues I know how to fix thanks to my previous experiences with the Windows App SDK. But the use of a web-based editor is weird to me and something I would need to learn more about before I will feel comfortable editing things and adding/changing features. Overall, this was more like my earlier AI tests, in that it was slow and error-prone. But it did get there.

So my mind wandered again.

Back to Google AI Studio for an improved Markdown editor

Given all the above experiences, and a few other small would-be apps I played around with, I returned to AI Studio and decided to use the most detailed prompt yet, to create a web app-based Markdown editor with all the features I wanted specifically spelled out. But most importantly, I also wanted to be able to install that app in Chrome or whatever web browser, just like any other modern web app/PWA.

This time, AI Studio created the first pass of this new web app in 360 seconds, or about 6 minutes. It is a PWA with offline and installing support, and I made just a few minor changes tied to styles, menus, keyboard shortcuts (like Ctrl + K for inserting a hyperlink), and an optional toolbar you can toggle on/off. 100 seconds later, that was done.

And it was mostly fine. But there was a formatting issue in the checkboxes displayed by the sample document it loads. So this morning, when I woke up, I asked it to fix that, too. 85 seconds later, the formatting issue was fixed and it also made a few small refinements tied to that. There are minor things I don’t like, such as the superfluous “Install app” button near the top and the weird Mac-like three buttons on the far left that don’t do anything. But this one is quite good.

And sure enough, you can install and use it like an app.

I could probably fiddle with this forever. But I think it’s time to move on to other things, including other apps created with or alongside AI.

Moving on…