I fixed a production bug from my phone with ChatGPT

Introduction

This story did not start as a serious project.

I had simply vibe-coded, with Google's Antigravity tool, a very small app to help my girlfriend. This was not built with my usual dev setup, my internal framework, my skills, or my guardrails. It was really just a tiny rushed test to help her out quickly. She's at ESIRA, and during the ranking session, students choose their job postings based on their position in the class ranking. So the idea was just to make it easier to explore the available options, filter them a bit, compare them, and arrive better prepared when the actual selection started.

Nothing more.

The kind of tiny tool you throw together in an hour before going to sleep, just to help someone out.

An app for one person… then suddenly for more

At first, it was really just for her.

Not a startup. Not a product. Not a strategic side project. Just a simple helper so she could make sense of the available positions before the live selection session.

The next morning, she shared it with her cohort.

And at that point, the story changes a little. What started as a small personal tool suddenly becomes useful to other people too.

The app is here: lgaudez.github.io/amphi-session-lesira

Then it broke

During the day, I pushed a small update to improve the mobile experience.

Without testing it.

Bad idea.

An hour later, my girlfriend messaged me: the app was no longer working. Blank page when launching the filters to explore the job postings. The people in her cohort who were already on the new version were blocked.

The part that surprised me

At that moment, I was not at my desk.

I was at physio. Just my phone. No usual setup with my projects, agents, or workspaces.

So I opened the ChatGPT mobile app. The normal public one. The same app anyone can use.

That is also what makes the story interesting: this was not my usual setup at all. Normally I work with multiple agents, with my personal agent running on Hermes, sessions orchestrated in Superset, and CLI tools like Claude CLI, Codex CLI, or Gemini CLI connected to my skills framework — I talk more about that setup here: Two Years of Development with AI Tools. Here, none of that. Just a consumer app opened quickly on my phone.

I gave it access to the project through the GitHub connector, explained the problem, and let it do its thing.

That is what really stayed with me.

Not the technical trick itself. Not the exact details of the fix. Just the fact that with almost no preparation, from a phone, I could unblock a real issue on something that people were actively using.

What impressed me

I expected friction.

I expected to wait. I expected to need my laptop. I expected to postpone it.

Instead, the problem got resolved quickly, from my phone, even though I had prepared nothing for that situation.

That is the moment when it genuinely felt impressive.

Because in the end, this is not only a story about AI helping with code.

It is a story about access. About almost zero friction. About being able to open a consumer app, grant access, and get something back on track very quickly.

The real point

I think that is the interesting part.

We often talk about performance, models, benchmarks, and ideal setups.

But in real life, value sometimes comes from something much simpler: the right tool, at the right time, with almost no friction.

And in this case, on this tiny app that had originally been built just to help my girlfriend, the simplest tool turned out to be the most useful one.

Conclusion

I am not saying this is how everything will be handled from now on.

I am not saying a phone replaces a real working environment either.

But I did find the experience revealing.

A small app built quickly to help one person. Then shared with a whole cohort. Then broken. Then brought back from a phone with ChatGPT.

Honestly, not long ago, I would not have bet on that working.

And yet, it did.

If you've ever used a consumer tool to unblock a real-world problem, I'd love to hear your example.