Building software with AI without losing control

Since 2022, AI has been part of how I build software: an integral tool, not a gimmick. The honest summary: I ship faster, and the risk has moved. It is no longer “can I write this function” but “do I still understand the system I am writing”. Control is the thing to manage.

Where control gets lost

Control rarely disappears in one dramatic moment. It erodes in three quiet ways:

• Copy-paste architecture. Generated code arrives as fragments. Ten sensible fragments can add up to one incoherent structure.
• Unowned abstractions. If nobody can explain why an interface looks the way it does, nobody can safely change it.
• Verification debt. Code that was written in seconds still has to be understood in hours, unless you invested in checks that do it for you.

The workflow that keeps ownership

My working rules are deliberately unexciting:

1. Decide the structure first. Data model, module boundaries and interfaces are human decisions. AI fills in the implementation.
2. Small increments, always verified. Generated code lands in small pieces against types and tests, never as a large dump.
3. Read everything that ships. The standard is simple: I don’t ship code I couldn’t explain line by line.
4. Let AI verify AI. Reviews, test generation and refactoring suggestions are where AI assists my judgement instead of replacing it.

What changes, what doesn’t

The economics of writing code change dramatically. The economics of running software do not: systems still have to be understood, debugged at 3 a.m., and evolved over years. AI-assisted development rewards people who understand systems, and exposes everyone who only assembles code.

The goal is not to write less software. It is to build better software, with judgement spent where it matters.