I used to put ideas on the back burner, blaming a lack of time. That excuse doesn’t hold anymore. So I recently decided to turn two small ideas for my personal use into reality:
- Airbnb Sort by Price Extension
- Spotify macOS Notification Center Widget
Given their small scope, I assumed AI could handle them. I used o3 Mini High, which was the best coding model until 3.7.
Airbnb Sorter
At first, sorting data on the initial load had hiccups, but after a few iterations, the model produced fully functional code. The only thing I had to manually provide was the HTML element IDs. Overall, great experience. It was my first time building an extension, despite being very comfortable with web technologies. The only minor inconvenience was that it requested excessive permissions, which I manually trimmed down before publishing.
Spotify macOS Notification Center Widget
I expected the biggest challenge to be pulling data from Spotify. Turns out, the real issue was macOS UserDefaults. UserDefaults on MacOS work in an odd way and AI couldn't fix it. Eventually it entered an infinite loop of suggesting the same over and over. Despite following Apple’s documentation manually, I couldn’t get them to sync between my extension and the app. I ended up temporarily disabling part of the sandbox—fine for personal use but not ideal.
My Takeaways
No surprises here: AI is great for tasks that work out of the box or are well-documented. I love how quickly I can build in unfamiliar territory. But the moment you hit an unusual issue, AI alone won’t cut it. I’m excited to try Claude 3.7 and its CLI tool to see if it handles the same challenges better.
Curious to hear others' experiences—have you built something more complex just by prompting? Have tools like Replit and Lovable worked for you? My impression is that AI is an incredible companion but still needs a developer to keep the bigger picture in mind.
This happens often as the context window length increases with the conversation. Restarting the conversation will often help.
So far my impression is that AI will be great for starting a project and iterating on small components, but not great for maintaining or enhancing a project. And there is still no replacement for good taste — the AI often just gets planning wrong and doesn’t push back if you ask it to do something dumb. So not a great partner for non-coders, yet.
> But using an editor is not the same as “no-code” I actually still prefer editors with assistants than no-code tools as I can always fallback to manually editing the code.
Deleted Comment
So I've used pretty much all the AI builders out there and seen a ton examples of other users using them.
V0 and Lovable are just ok for the UI if you have the Figma file ready but good luck connecting them with email or implement other integrations and Replit can build a prototype but it takes around 6K prompts to have a low quality mvp!
I think we are not there yet!
Deleted Comment