A long prompt can look intelligent.
It can also be the wrong tool.
Especially when the problem is still moving.
Most prompt advice assumes you already know what you want. The task is clear. The boundaries are visible. You are refining an instruction, not discovering one.
That is not how many real problems begin.
Sometimes the work starts in fog. You know there is something important there, but the shape is incomplete. The temptation is to compensate with more prompt. More context, more detail, more structure. It feels rigorous. In practice, it often locks the conversation too early around a framing that has not earned its place yet.
I have been using a different approach.
I ask the model to generate 30 to 40 questions on the subject. Wide enough to catch angles I have not named yet. Then I ask it to choose the three that matter most.
From there, I answer one. Or I answer a fourth question the first three triggered in my head. That part matters. The model is not there to judge the path. It is there to help surface it.
Periodically, I ask for a revision of the question set. Summarize the problem as it now appears. Name what is still unclear. Drop what no longer matters. Then I reset the context and continue from that summary.
That reset is not a technical detail. It is the move.
LLMs do not stay in interview mode for long. As context grows, they start converging too early. They lose selectivity. They become more eager to complete the pattern than to interrogate it. If the context also carries contradictions, the drift gets worse.
Resetting pushes against that tendency. It gives the model a cleaner working surface and gives me a way to check whether the framing is improving or merely expanding.
The stopping point is simple.
Stop when the summary reads like what you actually meant from the beginning.
Prompts are good at producing answers.
Loops are better at finding the question.
I am still tuning the reset.