I almost always ask the agent rather than tell it. Even when I have a clear opinion about what should happen, I phrase it as a question.
It was a hunch. Since I don’t work on hunches, I asked the agent to analyze my messages to coding agents over the last year. Out of roughly 5000, 27% contained a question and 9% were directives — about three to one. The count is by grammatical form, not by steering behavior, but the gap is there.
I’m not the one doing the work — the agent is — so I rarely know enough to tell. Asking gets me reasoning back, or pointers to where to look in the code, so I can understand.
When I tell, and I’m wrong, the agent works on it. Instruction-tuned models lean toward following user input, so the wrong instruction goes in and gets pursued. It will push back, but only after several turns of trying.
I run several agents at once. When I send an instruction, I usually switch to another window. The wrong instruction can run for 15-20 minutes before I’m back, or before the agent gives up. Without that parallel setup, the cost would be much smaller.
Asking has a side effect. Because the agent is already prompted with my preferred patterns and architecture, the mismatch often surfaces on its own. Instead of answering the question, it goes and fixes the problem — even without me calling it out as a problem.
I’m not sure asking is actually better in principle. If I knew exactly what to do, telling would probably be faster. But I usually don’t, and the agent is the one with the code in front of it. Until I know enough to tell well, the question is the cheaper move.