AI coding agents, without any doubt, accelerate software development. Tools like Claude Code and Codex are increasingly capable of implementing entire features, writing tests, and reviewing code. But the better these agents become at writing code, the more important another question becomes: How do we make sure they are building the right thing? Simply prompting an AI agent to “build an application that does X” might produce impressive results. However, it often leaves too many decisions to the agent. Product assumptions, architectural choices, and implementation details become mixed together, sometimes without anyone explicitly acknowledging them. I have been experimenting with a lightweight form of  Spec-Driven Development (SDD)  to address this problem. My approach is intentionally pragmatic. I don’t want to spend weeks writing specifications before seeing working software. At the same time, I want enough structure to guide AI agents effectively, maintain consistency...