I've been helping friends review their AI-built projects recently, and I noticed something interesting:

The code runs. The features are there. But somehow... it's not usable.

Not a technical issue. Something else.

After a few months of coding with the Spec workflow, I kept asking myself one question: why does this "write requirements first, design next, then break down tasks" process actually work?

Then I came across Gojko Adzic's Specification by Example: How Successful Teams Deliver the Right Software, and it clicked. Behind Kiro's Spec workflow is exactly the SBE (Specification by Example) methodology. This article explains the relationship between the two — and why your AI coding keeps ending up in rework.

After more than a year of AI-assisted coding, my biggest lesson is this: vibe coding feels great in the moment, but it falls apart the moment a project gets complex. A one-line requirement, AI starts coding on its own assumptions, and what comes out is always a bit off. You ask it to fix something, it fixes that and breaks something else. A few rounds of this and your codebase has tripled in size while the number of working features barely grew.

Eventually I figured it out: the problem isn't that AI writes bad code. It's that I don't say clearly what I want.

Kiro's Spec workflow is built to solve exactly that. Let's walk through it from scratch.