By Joseph Batrin Jr.
You ask the AI a question. The answer comes back generically, so you rewrite the question. You add a role and tell it to act like a marketing consultant. You tell it to be specific and give examples, same result. You try a third and a fourth time. Now you are getting frustrated.
The fourth prompt is a far better written prompt than the first one. The fourth answer is no more useful than the first one.
That is the moment worth stopping on. You rewrote the prompt four times and the output barely moved, because all four attempts were only one part of a prompt, reworded. There are five parts to a properly structured prompt. You were polishing one of them.
The Myth
The myth is that prompting is a matter of wording, that somewhere out there is a magic phrase that unlocks a better answer. It is an appealing idea, because the fix is free and the whole problem fits inside one sentence you can keep editing while you sit there.
Wording is not the skill. Writing a complete prompt is the skill.
The Reality
A complete prompt has five parts, and almost every disappointing answer you have ever gotten was missing three or four of them.
The first is Role, which tells the tool who it is supposed to be. An experienced copywriter and a general contractor with twenty years of estimating will answer the same question differently, and that one line changes the vocabulary the tool reaches for and the standard it holds the work to.
The second is Context, which is the situation the tool is walking into. Who the audience is, what has already happened, and what the real constraints of your business are. When that is missing the tool does not stop and ask you for it, it invents whatever it needs to keep moving, which is why the rest of this article is mostly about this one part.
The third is Task, which is the action you want taken, stated with a verb that means something, such as write, summarize, analyze or compare. The instinct is to soften it into help me with this, and a soft verb gets you a soft answer.
The fourth is Constraints, which are the rules the answer has to live inside: how long it should be, what it must not say, and which of your lines it cannot cross. Every rule you set removes an entire category of answer you did not want.
The fifth is Format, which is what the finished product should look like when it lands in front of you. It might be a table with three columns you name, or a numbered list, or a subject line followed by a body and one alternate subject line. If you don't tell the model what the format should look like, you're going to get a standard paragraph output, and then you have to spend the time you were trying to save editing and iterating it into the shape you needed all along.
What That Looks Like
Here is the prompt most people write.
Write me a follow-up email to a customer who never got back to me.
That is Task and nothing else, and a generic email is the only honest answer to it. Now here is the same request with all five parts in it.
You are a sales manager with fifteen years in home services, and you write the way an owner writes rather than the way a marketer writes. The customer got a quote for a water heater replacement nine days ago, told me the price looked fair, and has not answered since. We are a family shop, we do not discount, and we do not chase. Write the follow-up email. Keep it under 120 words, do not offer a discount, do not open with just checking in, and do not ask whether they got my last email. Give me a subject line, then the body, then one alternate subject line.
Same tool, same underlying request, and the second one comes back usable on the first try. Nothing about that was clever phrasing. Four missing parts got supplied.
Why Context Is the One That Breaks
Look at how that second version is built. Role, Task, Constraints and Format were each a single line, and I could have written all four from memory in fifteen seconds without knowing anything about the business. Context was the paragraph in the middle, and Context was the part that required somebody who actually knows how the shop runs. That is exactly why it is the part that gets left out, and it is the part carrying the weight.
In a business that context sits in four layers, and the layers you leave out are the ones the tool will quietly invent for you. Your identity is what you actually do, who you serve, and what you refuse to be. Your operations are how the work really moves through the business, including the bottlenecks and the workarounds nobody ever wrote down. Your standards are what good looks like to you, meaning your voice, the logic behind your pricing, and the line you will not cross even to close a sale. Your history is what you have already built, already tried and already ruled out, so nothing gets rebuilt from scratch and nothing that failed once comes back a second time.
Those four layers are the entire distance between an answer that could have been written for anybody and an answer that could only have been written for you.
Why This Costs You
An incomplete prompt does not cost you a wrong answer. It costs you another round. You ask, you read something that is close but not right, you adjust, you ask again, and every one of those trips takes a few minutes and a little more of your patience.
I have done this to myself on real work. I have had a document run to seventeen versions before it was right, and not one of those seventeen was the model failing at anything. I handed it an incomplete picture at the start and then paid for that on the back end, one version at a time.
There is an old line that fits this better than it fits most of what people use it on. The five P's: Proper Planning Prevents Poor Performance.
The Move Almost Nobody Makes
If you are unsure what belongs in a prompt, there is a shortcut, and it is the one I still catch myself forgetting to use. Do not write the prompt at all. Give the model the overall picture of what you want and need, and ask it to write the correct prompt for you.
The wording I use is close to this. I want this result. Before you do anything, write the prompt that would get me that result, and ask me any questions you need answered first.
What comes back is a structured prompt with all five parts laid out, plus a short list of questions only you can answer. That question list is the valuable half, because it names exactly which pieces of your context are missing, and you usually only have to make a couple of changes before you ask for the final output.
One honest note. None of this makes the tool infallible. A fully informed model still gets things confidently wrong and still needs a person reading the output before it reaches a customer. What a complete prompt removes is the round trip, not the need for judgment.
Do This Today
On your next real task, hand the model the overall picture and let it write the prompt. Then read the questions it asks you, because most of them will be about your business rather than about your request.
Once those questions are answered in the chat, tell the model to turn your answers into one context document you can reuse. Have the model store it inside the tool itself. Most of the major models have somewhere to keep it permanently, whether that is called custom instructions, a project, or saved information. Put it there once and you never paste it again. Problem Solved...