Visual Network
In December 2024 I built a small web app with Bolt.new, one of the tools that turn a description into a working application. It generated UK lottery lines, let you add your favourite numbers, used past draws to weight the picks, and limited how many lines you could play in a week. From the first prompt to a live link took about fifteen minutes. I wrote about it at the time, with a lot of enthusiasm.
It was a good demonstration. It was also the last app I built that way.
What the one-click builders are good at
They are very good at the first ten minutes. You describe the app, it appears, you click around, and it feels like magic. For a throwaway demo, or to see whether an idea is worth thinking about, that is useful.
The trouble starts on day two. Something breaks, or you want a change the builder does not understand. The code is there, but you did not write it, the builder does not explain it, and every fix it attempts changes three other things. Some platforms make it awkward to export the code, or to host it anywhere else. You end up renting your own app.
How I build now
I write the app with Claude and ChatGPT directly, and use Google AI Studio for AI features and quick tests. The AI still writes most of the code. The difference is that I see every file, I understand what each part does, and I can take it anywhere.
Visibility Pilot, the audit app behind Visual Network, started that way in June 2026: one long brief describing everything it should do, built with Claude, then improved most evenings since. It is far bigger than the lottery app, and it is still maintainable, because nothing is hidden.
The loop I use every time
- Describe the app in one sentence. What it does and who it is for. "A checker that tells a local business whether its Google Business Profile is complete." If you cannot write that sentence, the app is not ready to build.
- Decide whether anyone signs in. If nobody needs an account, it can run entirely in the browser: fast, cheap, nothing to maintain. If people sign in, save work, pay, or use an AI model through your account, it needs a server, so your keys never sit in public.
- Get a first version running. Rough is fine. A working link beats a perfect plan.
- Fix one thing at a time. Read the error, give it back to the AI, change one thing, run it again. This is most of the skill. Calm and incremental beats clever.
- Ship it somewhere real. Your own domain, HTTPS, the code in your hands. Then polish.
The costs nobody mentions in the demo
An app that calls an AI model costs money every time it is used. Before you build, work out roughly what one use costs and what a busy month would cost. Put a limit in place so a bug, or a stranger, cannot run up the bill. I learnt that one the expensive way on a separate project, and every tool I build now shows the cost before it spends anything.
Security is the other one. A prototype that works is not the same as an application that is safe with real customers' data. Sign-in, who can see what, where data is stored, and how keys are kept private all need thinking about before launch, not after.
When a builder is still fine
If you want to see an idea on screen this afternoon, and you are happy to throw it away, a one-click builder is a reasonable way to do it. Just do not mistake the demo for the product.
If you would rather not build it yourself
That is what App Creation is for: the same loop, done for you, with the source code handed over at the end. And if the idea turns out not to need AI at all, I will say so.