Original Summary

I&#x27;ll preface this by saying where I <i>haven&#x27;t</i> encountered this problem to set up some contrast:<p> Logical code search - want to know what component(s) do something or how something has changed<p> Design brainstorming - need to solve X but don&#x27;t like (or cannot use) my current ideas<p>* Investigation assistance (sometimes!) - execute tools to gather data and analyze it<p>These are generally things that I can &quot;verify&quot; fairly quickly: whether the code presented is what I thought I&#x27;d find or whether the brainstormed suggestions are novel and realistic. Investigation can be a mixed bag depending on how complex the problem is, but usually if I know the system well enough I can spot when it gets stuck on a red herring and either redirect it or pick up the rest myself. In any case there is some amount of manual effort being offloaded and giving me leverage.<p>However... when it comes to design documents (writing) and implementation (coding), I seem to be in the minority that finds AI slowing me down, and I think this is a problem with how I use it. I always review the generated output myself, to ensure that it:<p>1. Matches my intent<p>2. Maintains proper abstraction and comprehensibility, and<p>3. Surfaces edge cases I missed before getting into the details<p>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 &quot;oops, I thought this was negligible but it is load-bearing&quot;. In the coding phase, I will argue with it about what it has produced or interrogate its logic to align myself with its approach.<p>Sometimes I do find subtle issues, and on rare occasions I will find major issues. I&#x27;ll admit that I&#x27;m not sure anything Claude&#x2F;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.<p>By the end of it, I&#x27;ve expressed as much thinking in text as I would have in my head, and I&#x27;ve reviewed every character it produced. It doesn&#x27;t seem to have saved me time over typing everything in myself and I&#x27;m exhausted on top of that.<p>So I&#x27;m wondering: what am I missing when it comes to AI-driven design&#x2F;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?

  • 情报分类：商业与市场研究
  • 分类依据：内容涉及商业、投资或市场动态
  • 信息来源：Hacker News 新项目
  • 发布时间：2026/9/30 06:08:51