Getting Better Answers From AI: The Four Part Question
- A good question carries four things: who the answer is for, the situation, the job you want done, and how long the answer should be.
- Write your question on four short lines instead of one. Bland answers almost always come from one line questions.
- Ask follow ups in the same chat instead of opening a new one, because the model still holds everything you already said.
- When an answer is wrong, name the wrong part and hand over the right fact. Saying "no, try again" just reshuffles the same words.
- Ask for the shape you want. A table, five bullets, under 120 words. You will usually get exactly that.
You type a question. The answer comes back smooth, tidy, and useless. It reads like a textbook wrote it for nobody in particular.
That happens for one reason. Your question could have been asked by anyone, so the answer got written for everyone. Fix the question and the answer changes on the first try.
The four line ask
Here is the whole method. Instead of one line, write four.
- Who it is for. Name the reader. Your boss, a 13 year old, a customer who is already annoyed.
- The situation. Facts the model cannot see. Dates, numbers, names, what you already tried.
- The job. One verb. Write, compare, explain, rank, shorten, check.
- The shape and length. A table, five bullets, 200 words, no jargon.
Say those four out loud and the question writes itself. Line one and line two are the ones people skip, and they are the ones that carry all the weight. Anyone can type a job. Only you know your situation.
The term for this is prompt engineering, which sounds far more technical than it is. You are just telling someone what you need, the way you would tell a new colleague.
Why one line questions come back bland
A model answers the question it was given. Ask "how do I get fit" and it has no idea if you are 19 or 65, injured or healthy, at home or in a gym. So it hedges. It covers every case, which means it covers yours badly.
Rule of thumb: if your question fits comfortably on one line, it is missing the situation. Every generic answer you have ever received was really a request for more context, written politely.
Being extra polite does not help either. "Please could you kindly" costs you nothing, but it adds no facts, and facts are the part that changes the answer.
Three questions, rewritten
A work email you keep rewriting
Weak:
Write an email asking for a deadline extension.
Strong:
You are writing as me, a junior analyst, to my manager Priya.
I owe her a client report on Friday. Two of the three data files arrived late, so I need until Tuesday. I have already drafted the first half.
Write the email. Own the delay, give the new date, and offer a short update on Friday instead.
Keep it under 120 words and plain. No apologies stacked on apologies.
What changed: the weak version could be sent by anyone to anyone. The strong version names the reader, the cause, the new date, and the tone. It also tells the model what to leave out, which is why the reply stops grovelling.
If drafting messages is the bulk of your day, the AI Writing Tools run these same four steps as fields you fill in rather than lines you type.
Homework you have to explain out loud
Weak:
Explain photosynthesis.
Strong:
I am 13 and I have to explain photosynthesis out loud in class tomorrow.
My teacher wants the light reaction and the dark reaction kept separate. I mix them up every time. I know a plant takes in sunlight, water and carbon dioxide.
Explain it in the order I should say it. Use one everyday comparison, then tell me the one sentence my teacher is listening for.
Around 200 words. No chemical equations.
What changed: the strong version says the age, the setting, the exact thing that keeps going wrong, and what the reader already knows. Telling the model what you already understand is the fastest way to skip the paragraph you did not need.
A decision between two options
Weak:
Should I take the new job or stay?
Strong:
I am a 34 year old nurse with a mortgage and no savings buffer.
Offer A pays 4,000 more a year, 40 minutes further away, and a manager I have not met. Offer B is my current ward, familiar, but my pay has been flat for three years.
Lay out the case for each. Name the one thing that would decide it, and the question I should ask before I answer.
A short table, then three sentences. Do not tell me what to pick.
What changed: money, distance, risk and the unknown manager are now on the page. The last line matters most. Decision questions get vague answers because people ask for a verdict, and the model softens it. Ask for the case on both sides and you get something you can actually argue with.
Word counts of the six example questions on this page. Count them yourself. Bars are scaled to the longest one, 84 words. Length is not the goal. The extra words are all context, job and shape, and nothing else.
Prompt patterns worth copying
Keep these near your keyboard. Left column is what you want. Right column is the wording that gets it.
| What you want | What to type |
|---|---|
| No warm up paragraph | "Answer in the first sentence. No preamble." |
| A table instead of prose | "Give this as a table with these columns: option, cost, main risk." |
| Options rather than a verdict | "Give me three options with one line of trade-off each." |
| The thinking, not just the result | "Show your steps first, then the answer on its own line." |
| A targeted fix | "Point 3 is wrong. The real figure is 12%. Redo point 3 only." |
| A shorter version | "Same answer, half the length. Keep the numbers, drop the examples." |
| A different reader | "Rewrite this for a busy manager who skims." |
| Fewer invented details | "If you are not sure, say so. Do not fill gaps with guesses." |
| A set tone | "Direct. No exclamation marks. No marketing words." |
| Work on what you already have | "Using the draft above, only change the second paragraph." |
More patterns, with worked examples, sit in the guide How to Write Better AI Prompts.
Follow up instead of starting over
Most people get one mediocre answer, then open a fresh chat and retype everything. That throws away the context you just built. The model still has your situation, your tone and your last draft, so the second question can be tiny.
Good follow ups are short and surgical. "Too formal, drop it two notches." "Cut the third paragraph." "Now do the same thing for the Tuesday version." Each one is cheap because the setup is already there.
One caution. Very long threads drift, and old corrections start slipping. If the answers wander off after twenty turns, start a new chat and paste in the best version you have so far as the new starting point.
How to say "that is wrong" so the fix sticks
Telling a model it is wrong works, but only if you say what is wrong. "That's incorrect, try again" gives it nothing to steer by, so it rewrites the same idea in new clothes. Sometimes it agrees with you and then repeats the error.
Use three parts: name the wrong bit, give the correct fact, say what to redo.
Paragraph 2 is wrong. Our office closes at 4pm, not 6pm. Rewrite paragraph 2 only and leave the rest as it is.
Watch for the opposite failure too. Push back on something that was actually right and many models will fold and agree with you. If a number matters, check it somewhere real. A confident tone is not evidence, and models will not always say when they are unsure.
Two dials most people never touch
Ask it to show its reasoning
For anything with steps, ask for the working before the answer. Maths, budgets, schedules, eligibility rules, comparing two plans. You are not doing it to be thorough. You are doing it so you can spot the wrong step, then correct that one step instead of the whole reply.
Skip this for simple lookups and short rewrites. Reasoning on a one line factual question just buries the answer in padding.
Name the shape of the answer
Shape is the cheapest instruction there is. "As a table." "Five bullets, no sub-bullets." "A numbered checklist I can print." "Two columns: what to do, why it matters." Say it and you get it, and you stop skimming walls of prose for the one line you needed.
Length belongs here too. "Under 100 words" and "one paragraph" work far better than "keep it brief", because brief means nothing in particular.
Check these four before you hit send
Four ticks takes about twenty seconds. It saves the three rounds of "no, more like this" that usually follow a one line question.
What to do with your next question
Take the question you were about to type and add the three missing lines. Send both versions if you want proof, the short one and the four line one, and compare what comes back. The difference is not subtle.
Try it in the free AI Chat, where you can put the same four line question to ChatGPT, Claude and Gemini in one window. Models differ in style and length, so comparing them side by side with Compare AI Models on your own question tells you more than any review will. Long threads and heavy rewriting use more tokens, and the plans page lists what each tier includes.
Write four lines. Read the answer. Correct the one part that is off. That loop is the whole skill.