Thread

1. hombre_fatal · 1786032618 · focus · HN ↗

   We have this idea of AI software development where a human watches over its shoulder and shouts out things like "use uuidv7 for the id since timing is important" or "use a sum type here to make that bad state unrepresentable".

   But the thing is that most of this can be encoded in a markdown file for agents to read if it doesn't already come out of the box in the next round of sota models. And funnily enough as agents get smarter, you risk being overprescriptive where you hamstring the agent from making a pivot that would have led to better engineering.

   The future is pretty clear to me at this point that we won't need software engineers looking over the shoulder and instead it's just a "user with taste" asking for revisions.

   1. sailfast · 1786035069 · focus · HN ↗

      A software engineer user with taste, yes.

      Have you ever been really deep in a thread with an agent on a technical deep dive? The trade-offs conversations would not make sense to someone that is not able to decode the real decisions they’re making.

      You can dumb down the agent, but the reason domain language matters is because of the precision of the word sometimes - not to be a gatekeeper.

      Anyway - user with taste that understands all the elements of their decisions would be my modifier :)

      1. hombre_fatal · 1786036498 · focus · HN ↗

         Long ago I added an instruction to my global AGENTS.md that every time the agent asks me something, make it multiple choice that ranks the top options, give a quick trade-off summary of each one, recommend one, and justify it.

         Over the last year, I've noticed that the recommended option for technical questions has become one that is always sufficient to the point that I'm once again in a tastemaker position, not in the "here's a better idea btw" position.

         And if that's true, then my participation in the intermediate steps isn't strictly necessary to get a well reasoned position that's only directionally wrong.

         1. user43928 · 1786048471 · focus · HN ↗

            The Claude Code App automatically does this by default.

            The Codex app only does it in Plan mode, I believe. I did not touch the Plan mode for a while.

            1. StilesCrisis · 1786073783 · focus · HN ↗

               Claude does, but I've also seen it give a list and then afterwards, it changes its mind and recommends option B or C instead. Really drives home how LLMs don't actually know what they're going to say until they say it.

Unofficial Hacker News client; not affiliated with Y Combinator.