Comment by cruffle_duffle

I think this generalizes to a lot of professional work.

LLMs are very good at satisfying the requirements you give them. They’re much worse at knowing which requirements should be challenged, reframed, or ignored because you don’t know enough about the field to know what actually matters. That’s why they’re such force multipliers for experts: the expert can spot when the model chose the wrong abstraction or checked every box and still produced something bad. A novice often can’t. I’d argue the model mostly raises the novice toward “average.”

“Tell it to push back” doesn’t really solve this either. Too little and you get “aye aye captain!” Too much and you get Opus 5 / claudise, where it seems compelled to find something wrong with everything.

I see this in a Home Assistant project I’ve mostly vibe coded. I can ask Codex to make the kitchen brighter on school days and it’ll happily implement something. But maybe an HA expert would say “you really shouldn’t model it this way.” Maybe they’d even question HA itself. The model could conceivably know that too, but it only sees a tiny slice of my world. There are infinite side quests it could raise, and a good professional’s real skill is knowing which one matters enough to interrupt you about and which 99% to silently ignore.