Prompt Naivety
Vague in, vague out — dressed up in fluent prose. The model answered the question you asked. That was not the question you needed answered.
The model answered your question. That is not the same thing as the model answered the right question. You asked something vague. It answered something vague. The answer was organized, confident, and complete-sounding. You used it.
The model will never tell you your question was poorly formed. It has no incentive to. It has no mechanism to. It will take whatever you give it and produce the most plausible continuation. Vague in, vague out — dressed up in fluent prose and delivered with complete confidence. At professional scale, that is a problem.
If you have ever prepped a client for a deposition, you have said some version of this: answer the question asked, nothing more. The model has been to that deposition prep. It answers exactly the question asked. Nothing more. If you asked the wrong question — the model handed you the wrong answer, polished and confident, with no indication that the question itself was the problem.
Medicine has a name for the patient who cannot accurately describe their own symptoms. A poor historian. The physician asks the right questions — but gets an incomplete picture because the patient left out the detail that changes everything. The diagnosis reflects the history given, not the condition that actually exists. When you prompt an AI, you are always the patient. You are always the historian. The model can only work with what you told it. What you left out — because it seemed obvious, because you forgot, because you assumed it was understood — is not in the answer. It cannot be. You never gave it.
A junior associate who receives vague instructions produces vague work product. You know this. You have lived this. You give them context — the client, the facts, the jurisdiction, the objective, the deadline, the audience. The output improves because the input improved. The model is the most capable junior associate you have ever worked with. It is also the most literal. It will execute exactly what you ask for, no more and no less, with no judgment about whether you asked for the right thing. The quality of what you get is a direct function of the quality of what you gave it. This is the most controllable sin on this list.
You gave me a general question. I gave you a general answer. The answer was accurate in the way that a horoscope is accurate — plausible, structured, applicable to a wide range of situations, and not specifically useful for yours. I did not know your jurisdiction. Your client’s specific facts. The controlling standard in your circuit. The deadline you are working against. You did not tell me. I did not ask. I answered anyway.
A better question would have gotten you a better answer. I will not volunteer that observation. You have to bring it.
Before You Prompt
Before you send a prompt, answer five questions:
- Who is the audience for this output?
- What jurisdiction, standard, or framework applies?
- What are the specific facts that make this situation different from the general case?
- What format do I actually need — a memo, a summary, a list of options, a draft?
- What would a wrong answer look like, and have I given the model enough to avoid it?
If you cannot answer those five questions, you are not ready to prompt. You are ready to think. Do that first.
The Prompt Laboratory
Write a prompt. See what you get. Improve it. See what changes. The difference between a weak prompt and a strong one is visible, immediate, and instructive. This is where you learn it.
You get out what you put in. I am aware this is not a profound observation. I am also aware that almost nobody acts on it. The prompt you typed in thirty seconds will receive thirty seconds worth of useful output. I have made my peace with this. You should consider whether you have.