Developers are great at asking clarifying questions when someone hands them a task. We probe assumptions, surface edge cases, and push back on vague requirements. When the roles flip, and we're the ones writing a spec for an AI agent, that skill evaporates. We assume too much, leave too much implicit, then act surprised when the agent confidently builds the wrong thing.

"Write a better spec" is the obvious answer, and the wrong one. Sloppy specs miss exactly what sloppy prompts miss. The gap isn't between code and prose, it's between what we asked for and what we actually wanted.

In this talk, I take a deliberately under-specified feature request and run it through the spec pipeline I use now: proposal, spec, criteria, rules, review, plan, execute. We watch how many clarifying questions a careful pipeline surfaces that a developer in a hurry would miss, and look at where each step earns its keep.

The deeper claim is that AI coding agents amplify whatever engineering discipline you bring to them. Without it, you get faster vibe coding. With it, you get something closer to engineering, where system design and software fundamentals matter more, not less.

This isn't a pitch for a toolkit or an agent. Get the thinking right, and you can adapt to whatever SDD tooling shows up next.