AI Couldn’t Build My iPhone Podcast App
After the GPT-5 event, where OpenAI touted its incredible coding prowess, I had an idea: Try to make an iPhone app completely relying on ChatGPT and get it all the way to App Store approval. For context, I have zero knowledge of Swift, no coding experience whatsoever, and don’t even know how to use Xcode.
But after just a few hours vibe coding, I had a working app. A few days later, it even got through App Store approval. You can watch the whole saga here:
But that wasn’t enough. I figured if this was possible, maybe I can build a more complex app. An app I would be proud to share and use daily. I was going to build a podcast app.
I created a custom GPT, then trained it on all the Swift, Swift UI, and Apple developer documentation. I explained I knew nothing about coding and it would need to walk me through step by step. I started with the following prompt, still using ChatGPT-5, and I was off.
ChatGPT walked me through creating dozens of Swift files. Once set, I tweaked the UI and layout in natural language, and it felt like this could really happen. The app was bare bones, but it built successfully! Then, I tried to actually use it in the simulator.
At first, it would hang on searching for shows. After I seemingly fixed that, adding a show to the Library didn’t work. I kept going back and forth with GPT-5 and it kept getting worse. Every new build there was an increasing number or errors and unintended changes to UI elements I hadn’t asked for.
Up until this point, I was using the ChatGPT Mac App with “Work with Xcode” turned on, so GPT-5 could make changes to the code itself within the active window. Sometimes it would think it’s changing document X, but I had document Y opened, requiring me to revert, undo, and many times, get lost in the process.
Eventually the app failed to build and I could not fix it with ChatGPT. It felt like we were going in circles.
In my desperation, I turned to a different LLM. Many commenters on YouTube and social media suggested Anthropic’s offerings were better suited. I downloaded Claude and provided the full context of my app. I uploaded every Swift file I had created with screenshots of all the things I didn’t understand. From Core Data to build…files? I repeat, I have no idea what I’m doing.
Claude seemed to resolve some errors and I was able to successfully run the app in the simulator again, but simple functions like pulling episodes of a podcast still didn’t work. At one point I actually hit the limit of Claude Pro, my first time ever hitting an LLM’s token limit, and paid to upgrade to Claude Max! The $100 version…I couldn’t swallow spending $200 on this folly.
Then, I heard that the latest beta of Xcode allows you to connect ChatGPT and Claude directly, giving it full access to your project with the ability to make changes itself. So I downloaded the beta, connected Claude and GPT-5, and continued asking it to fix the app. At least let me see a list of episodes in a podcast!
Even with full access to my files, and despite it changing lots of code behind the scenes, the app still does not work. And the worst part is, I don’t even know how close (or far) I am to getting there! No amount of sending screenshots, using ChatGPT or Claude, or explaining what’s wrong, seems to help.
So, I’m calling it. I’m acknowledging the sunk cost fallacy and can confidently say: building a robust app is really hard, and current LLMs can’t do it on their own.
I’m sure in the hands of a skilled developer, these tools can save time, take care of menial bugs, and maybe even provide inspiration. But in the hands of someone with zero coding knowledge, they may be able to build a single-function coffee finder app, but they certainly can’t build a good podcast app.
I tip my hat once again to all the developers out there! And specifically to the makers of Overcast, Pocket Casts, Castro, Nueuecast, Queue, Fountain, Goodpods, and Apple Podcasts too: thank you for making incredible podcast listening apps. I literally don’t know how you do it.
