Building the app with AI turned out to be the easy part.
In the last episode I told how I worked with several AIs and why I had to write rules against their memory. This is the last episode of the series. It isn't a presentation of the app. It's an honest look at what held and what I had to redo, then at the part nobody warned me about: what happens after the app is on the stores.
The numbers
From the first prompt in January to the stores in July, plus a few months of updates since. One person, a general-purpose AI as the developer, FlutterFlow as the compiler. Numbers from the project export of mid-November (the website count is from October):
| What | How much |
|---|---|
| Screens | 16 pages, 24 components |
| Custom code | 201 pieces: 94 actions, 74 functions, 33 widgets |
| Dart in the project | about 86,000 lines |
| Of which custom code | about 47,000 lines (55%), written with the AI and pasted into the editor |
| Comments in the custom code | about 6,300 lines |
| Cloud functions | 39 |
| Security rules | about 1,500 lines, half of them comments |
| Test cases on the security rules | about 550, run on the emulator |
| Languages | 5, with 273 texts in the language panel alone |
| Website | 48 pages in 5 languages |
Two numbers say more than the others.
55%. More than half of a "no-code" app is code. FlutterFlow built the screens and compiled everything, but the logic that makes the app different lives in custom code. FlutterFlow was my compiler, not my author.
6,300 lines of comments. Every AI session starts from zero. The comments are the memory between one session and the next: why a choice was made, which version fixed what, which decision not to reopen. 95 comments in the custom code carry a "FIX" note.
What held
- The colors from my first prompt. Same hex codes, months later.
- History on the phone, events that expire. Decided in February, never questioned.
- A strict line between the two tools. The app is built only in FlutterFlow's interface. The backend is deployed only from the command line, never from FlutterFlow. The first line of the backend files says it in capitals: do not overwrite from FlutterFlow.
- An order for every release. Security rules first, then indexes, then functions, then the app. It sounds bureaucratic. It saved me more than once.
- The idea itself: events on the map, not people. It took a few weeks to get there in the data and months to say it in words, but it never moved after.
What I redid
- The onboarding, three times.
- The Wish, three times.
- The place search, moved from one provider to another.
- Ads. There was a small banner at launch. I removed it at the end of August: it earned almost nothing, it limited what I could say about privacy and it would have closed the door to public bodies and non-profits.
- How phone numbers are stored, three times.
- The word "Vault", removed from the interface because nobody knew what it meant.
- The whole interface at the end of August. It was too hard to use.
The app didn't stop in October. The latest version lets associations and public bodies put their events on the map too, next to the ones of your friends. It grew the same way as everything else in this series: one real request at a time.
What I would do differently
If I started tomorrow:
- A proper schema first. I'd draw the structure of the app (data, screens, flows) with a diagram tool before writing the first prompt. A more solid skeleton would have saved me most of the rebuilds.
- More analysis of the interfaces and more use cases. Not "does it work", but "does a person understand it without help". I learned this with a full rebuild at the end of August, which cost me more than any bug.
- Intuitiveness and friction as the main requirement, not a final polish. Every extra tap is a reason to go back to the group chat.
- A real group before the first line of code. Ten people who already meet every week and agree to try it are worth more than any feature. I built for months for users I didn't have yet. This is the big one.
The hard part
I spent most of a year building. I came out of it with a clear conclusion: the hardest problem isn't building the app. It's getting it used.
An app can be good, fun and useful. Without a group that uses it and a way to spread it widely, it's still useless. An app for meeting up with friends is the extreme case. With one user it's an empty map. It only works when the people you'd go out with have it too.
And this matters to me more than the code. My conviction is that people see each other in person less and less. Most of the apps on our phones push in that direction. I built this one to push the other way: you see where something is happening and you go. An empty map can't do that.
What happened after the launch
These are the honest numbers of the last weeks. They are the real ending of this series.
- Reddit. For six weeks I answered questions in developer and student communities, without links, mostly with technical answers. A post about a SQLite fix in FlutterFlow reached about 1,800 views. In the last four weeks the website got 5 visits from Reddit. What Reddit gave me was something else: conversations with other developers. A library author thanked me by name in a release built on my suggestions, then opened an issue for another one.
- Short videos. Since November 9 I've posted 14 short videos for the app. They sit between 237 and 370 views each, with an average watch time between 1 and 2.8 seconds.
- Email. I write to student associations across Europe, one by one. It's slow. It's also the channel I believe in most: a student association is already a group that meets.
- Indie Hackers. In the same four weeks it sent 15 visits to the website, three times Reddit.
AI made building possible for one person. It didn't make distribution easier. Only use and word of mouth give an app credibility. Those can't be generated. That's what I'm working on now: not chasing downloads, but finding groups of people who already know each other and already want to meet.
One last detail
The project file of the app has a short description field, written early on, that nobody reads. It still says the app lets you "share where you are now, where you will be or where you would like to go".
The words from January survived there. The app says the opposite. I'll leave them there for a while, as a reminder.
What I took from the whole series
- A general-purpose AI can be your developer. It can't be your memory. It can't be your judgment either.
- Put your intent in the structure (data, rules, decisions written down), not only in the prompt.
- Write down every decision you don't want to defend again.
- Building is no longer the bottleneck. Getting real people to use what you built is.
Thanks for reading the series. Building something similar? I'd like to hear how you're finding your first users. Which channel brought you people who stayed?
I'm building orbiTome solo, with FlutterFlow and Firebase. More at orbitome.com.
Top comments (0)