I vibe-coded a Pebble Time 2 app, and I learned some tough lessons

If you spend enough time on the internet, you’ll be led to believe coding is now easy for everyone thanks to various AI models. Claude, ChatGPT, and others can turn even a novice into a master programmer, and software engineers should seriously consider a second career as pastry chefs. Unsurprisingly, that’s a gross mischaracterization. Large language models do give people without formal training tools they could’ve never imagined before, but it’s not that simple.

So, I decided to find out for myself. I recently wrote about how much I love my Pebble Time 2, and part of the appeal is PebbleOS. It’s open-source, and people can write their own programs to fit their needs. I decided to lean heavily on nostalgia and see if I could recreate the Microsoft MSN Direct service from the mid-2000s. I wanted my Pebble Time 2 to remind me of my Fossil Abacus, so I enlisted AI to help.

Here’s what happened.

It’s not as simple as just giving one prompt

To start, I had to choose a model. I ended up going with a combination of ChatGPT and Codex. It really boiled down to me not using the Codex capacity I already had with my $20-a-month Plus subscription. A couple of other friends, including 9to5Google’s Daniel Bader, are convinced Claude is the better option, but I wanted to go with what I knew.

Unfortunately, vibe coding takes a lot more effort than giving Codex a single prompt and letting it work while I accessorize my Lambo. I had to be an active participant, and it started with design. I knew the aesthetics and functionality I wanted to translate over from the old Wrist.NET service, but getting there was a challenge. I used ChatGPT to optimize my prompts and walk through what I wanted; that way, I gave Codex clear instructions.

At first, I was impressed. Codex went out onto the web and researched what MSN Direct used to look like, down to fonts and layouts. It took a few rounds of back-and-forth, but we arrived at something that approximated the Nina font Microsoft used 20 years ago. A few prompts later, I had a backbone built out, complete with Stocks, News, Weather, and Sports categories. Codex pulled up the PebbleOS emulation tool, and the app appeared to be running well, but then the issues started to hit.

More goes into app creation than you think

You’ll often see software engineers push back that it’s easy to create an app with vibe-coding, but good luck actually deploying it, and I found that to be the case even on my small-scale app. What Codex created looked the part, but the functionality took more effort than I expected. It’s not just as simple as my app going to ESPN and porting over the scores. I needed to set up reliable API feeds for stocks and sports. I don’t plan to sell my app or release it outside my own ecosystem, but if I did, I’d have licensing concerns to contend with.

The combination of ChatGPT and Codex was undoubtedly user-friendly. It walked me through the steps to get the beta version running on my watch and helped with each problem along the way, but it still took hours of testing and fixing bugs. At first, stocks weren’t refreshing properly, and then sports scores weren’t showing up on the day the games were played. Each problem required several prompts and plenty of input from me on how to fix the issues that arose.

The entire process required more patience than I was expecting. AI models are fantastic and have advanced a great deal, but they can still get caught in loops that require someone to step in and stop them. ChatGPT can sometimes feel like a wonderfully gifted teenager that needs proper focus to achieve its best. The information pipelines were eventually sorted, and I’ve been using my PebbleNET app for a few days now, but it took hours and hours of work to get it up and running.

I have a new appreciation for developers

I didn’t think vibe coding would be easy, and it’s certainly given me a new appreciation of how apps are built. The experience has also helped me reconcile a larger question about AI I’ve had. I love that I can use Linux and code apps in a way I never could have before. My brain doesn’t work that way, and I could study for 100 years and still never write a line of Python.

Still, I’m not foolish to equate what I can do with a few prompts in ChatGPT to what professionals can do. AI can get us in the door and let us enjoy skills and experiences we wouldn’t have otherwise, but it’s still a stretch to consider it a suitable, wholesale replacement.

I also have a newfound appreciation for those who vibe-code. It’s not effortless, and just because you don’t have the technical expertise doesn’t mean you can’t create. Your vision and preferences still have to go into the app, and it’s a challenge to keep AI on track to achieve them. I couldn’t give Codex one prompt that made me a millionaire, but I could use it as a tool to recreate a bit of nostalgia on my wrist — and that’s good enough for me.