Structure, not attention: how I guard the correctness of a service I build alone with AI

Summary

I have spent about two weeks building a paid membership service on my own, with AI agents in separate roles writing much of the code. Once you have AI write code, a problem shows up right away: who checks whether the code the AI wrote is correct, and how. Since I build alone, I cannot afford a second person whose only role is to doubt the code and investigate it, separate from whoever writes it.

So I stopped relying on the author reviewing the code carefully to keep it correct. Instead, I reshaped the design so that mistakes are less likely, and, apart from the AI that carries development forward, I set up an independent audit whose only aim is to look for how the system breaks, and had it verify the work. This way, I have caught four defects before publishing. Every one was a defect that broke only when parts were combined or a state changed, even though every per-part test passed. In this article I write how far I delegate to AI, how I get the rules to an AI that starts fresh each time, how I have an independent audit find the mistakes the author cannot find, and how I finally reshaped the design so that mistakes are impossible.

Who this is for and what you can take away

I write this for someone who, alone or in a small team, is trying to build a paid membership service while having AI write the code. By the end, you can take away a way of thinking: where AI may proceed on its own judgment, and from where a person checks first and then instructs, drawing that line not by the kind of operation but by the conditions that must be met. It is based on the record of building this site alone with AI.