However, when it comes to design documents (writing) and implementation (coding), I seem to be in the minority that finds AI slowing me down. I think this is a problem with how I use it.

For example, I always review the generated output myself, to ensure that it: (1) matches my intent, (2) maintains proper abstraction and comprehensibility, and (3) surfaces edge cases I missed before getting into the details

This leads to me spending lots of time going back and forth with the agent. In the design phase, I often get to a point where the approach seems sound, but then upon pressing for 15-20 minutes the agent tells me something like "oops, I thought this was negligible but it is load-bearing". In the coding phase, I will argue with it about what it has produced or interrogate its logic to align myself with its approach.

Sometimes I do find subtle issues, and on rare occasions I will find major issues. I'll admit that I'm not sure anything Claude/Codex has generated in the past 3 months would be catastrophic if shipped as-is. Still, I work on systems whose primary objective is reliability and performance, rollout times can be long, and blast radius can be difficult to contain -- so I need to be careful and a measure-twice-cut-once approach is often necessary.

By the end of it, I've expressed as much thinking in text as I would have in my head, and I've reviewed almost every character it produced. It doesn't seem to have saved me time over typing everything in myself and I'm exhausted on top of that.

So I'm wondering: what am I missing when it comes to AI-driven design/development? Is there something I can change in my assumptions or process to get a smoother outcome, or has this been difficult for others too?