Code isn't high quality because I looked at it A developer who spent a full week "vibe coding" concluded that code quality comes from tests and structure rather than human review, arguing that "code isn't high quality because I looked at it" but because it "contains good ideas and a structure that enforces them, like fitness tests." The developer replaced a one-PR-per-change workflow with discussing a feature or architecture change with the AI until it is clearly understood, then having the AI create the issue and implement everything in one go, relying on unit, integration, e2e, fitness, smoke and mutation tests as a coding harness. The experiment also produced failures: the AI ran mutation tests on the developer's laptop and generated more than 200GB of test artifacts, and a larger refactoring toward a clean architecture left adapters and ports strangely distributed. I spent the whole week vibe coding. No regrets and I would do it again. That was probably a reaction to my last blog post, where I realized I should be more ambitious in exploring the capabilities of AI. Initially, I kept a tight development workflow where one PR = one change. I think I did this mostly for myself, to feel in control. Given that I didn't even pay much attention to the change itself, that was rather hypocritical of me. Now I first discuss what I want with the AI e.g., a feature or a change to the architecture until I have a clear understanding of it, and then it creates the issue and implements everything in one go. Maybe that sounds weird, but it feels more honest to me. What I still had to figure out: when I'm not looking at the code anymore, how do I make sure it meets my quality standards? The solution is the same tools we already had before, linters and tests, in all their variants. I see this a bit like a coding harness. A harness tames a model, and tests tame your code. By having unit, integration, e2e, fitness, smoke, mutation tests and so on, you're forcing your code into a fixed structure. Maybe that's something I learned from my aggressive vibe coding experiment: code isn't high quality because I looked at it. It's high quality because it contains good ideas and a structure that enforces them, like fitness tests. Still, there were a lot of failures too. The AI ran the mutation tests on my laptop and almost filled my whole storage with more than 200GB of test artifacts. That was dumb. And when it did a bigger refactoring towards a clean architecture, adapters and ports ended up strangely distributed. That's when I was glad I couldn't resist glancing at the code. My head also feels constantly full. That's absolutely not a sustainable pace, or one I want to keep up for longer. I believe I have to go through this right now to figure out a workflow that feels good to me.