[yomismoaqui]

You are right, but the responsibility of the final artifact must fall on the human.

Think about what we did before if we didn't have another human around to ask and think together about a problem.

We searched for solutions or more info on Stack Overflow, Reddit , random blogs or HN even. The we tried to evaluate the pros and cons of each possible solution and then decide what to do.

Now we should use the LLM to get that info from the internet (be it from its lossy memorized or better fresh from its search tool). Then try to ask the LLM for pros and cons and follow the links it provided if you don't trust its "judgement".

[kelseydh]

That assumes the problem is a common one others have encountered, which the examples I gave above certainly were. When you're wrangling with poorly documented legacy code operating under the context of its own internal domain logic (e.g. arcane country specific banking regulations), often the only source of good "judgement" (that's the commonwealth spelling btw) are those in the past who wrote the code the way they did.

This is an area where Claude Code is both valuable and dangerous. It can propose sweeping (correct) changes based on inconsistencies it finds within the codebase. The developer, in situations where nobody more senior is around to answer those design questions, is left making a judgement call based on vibes and what logic they can piece together about Claude's changes.