Give AI These Four Things And The Answer Gets Sharper

A prompt works when it gives the model four things: a role that matches the actual job, the real constraints instead of vague hints, one example of the output you want, and explicit permission to ask questions before it answers. Skip any one of these and you get a generic, guessed-at response instead of a useful one.

What Makes A Prompt Work?

A prompt works when it removes guesswork for the model, the same way a good brief removes guesswork for a new hire. Four elements do that: a role that frames the perspective you want, the real constraints of the task such as length, audience, format and deadline, one example of what good looks like, and a line telling the model to ask before it answers. Anthropic’s own prompting guidance for Claude covers role, clarity and examples as separate techniques you combine rather than one single trick. Most people use zero or one of the four. Using all four turns a vague request into something closer to a finished brief, which is why the output changes so much between a one-line prompt and a four-part one. If you have ever gotten back a flat, generic answer, the fix is usually one of these four missing pieces, and I’ve also written about three lines that fix lazy AI answers as a faster patch while you build the full habit.

Four Prompt Parts, Ranked By Impact

Here is the order I teach, ranked by how much the output changes when each part gets added on top of a plain request.

1. An example of good output. Anthropic’s own prompting guidance for Claude lists showing examples, often called multishot prompting, as one of its core techniques for improving accuracy, because the model copies structure and tone instead of inventing its own.
2. The real constraint. A number, a format, a deadline or a length limit closes off the thousand reasonable answers the model could otherwise give you. OpenAI’s own prompt engineering guide puts it plainly: models benefit from precise instructions that explicitly provide the logic and data required, not general direction.
3. Permission to ask questions first. A model that asks before answering catches wrong assumptions before they turn into a wrong answer. A 2019 study on clarifying questions in conversational search found that asking a single well-chosen question improved retrieval performance by more than 170 percent compared with guessing at what the user meant.
4. The role. Smallest effect of the four on its own, but still measurable: a 2023 study on role-play prompting found that giving ChatGPT a role raised its accuracy on a reasoning benchmark from 53.5 percent to 63.8 percent, using the same question either way.

Stitched together, those four parts might read: “You are a technical writer who explains security concepts to non-technical employees. I need a 150-word explanation of multi-factor authentication for a company intranet post. It should avoid jargon and end with one action the reader should take this week. Here is an example of the tone I want: [paste a paragraph from a post you liked]. Before you write it, ask me anything you need to know about the audience or the intranet format.” That is a role, a constraint, an example and permission to ask, inside four sentences.

Why Does A Role Help The Model?

A role helps because it narrows the model’s choice of words and structure down to one job, instead of averaging across every job it has ever been asked to do. A 2023 study on role-play prompting tested the same questions with and without an assigned role and found consistent gains: on one reasoning benchmark, ChatGPT’s accuracy rose from 53.5 percent to 63.8 percent, and on a separate word-based task it rose from 23.8 percent to 84.2 percent, just from adding a role to the same question. Specificity is what makes it work: “you are a senior editor who cuts jargon for working professionals” does far more than “you are an expert.” A vague role changes almost nothing, because the model still has to guess which kind of expert you meant. A role also points the model toward using whatever long context or tools it has access to with the right priorities, instead of the wrong ones.

What Counts As A Real Constraint?

A real constraint is specific enough that two different people reading it would produce a similar answer: a word count, a reading level, a deadline, a list of what to exclude, or a format such as “three bullet points, no headers.” Vague instructions like “make it good” or “be concise” leave the model guessing at a target it cannot see. OpenAI’s own prompt engineering guide makes this explicit, noting that models benefit from precise instructions that explicitly provide the logic and data required, rather than general direction. Anthropic’s prompting guidance for Claude covers the same ground from a different angle, treating being clear and direct as a teachable skill rather than something a good model should just intuit. Either way, the lesson is the same: narrow the brief before you hand it to AI.

Why Show An Example Of Good Output?

Showing an example works because models are built to continue a pattern, and an example is the clearest pattern you can hand one. Google’s own prompting guidance for Gemini recommends you always include few-shot examples in your prompts, stating plainly that prompts without them are likely to be less effective. Anthropic’s prompting guidance for Claude treats multishot prompting the same way, as one of its core techniques rather than an optional add-on. One example of a good email, summary or headline tells the model your tone, length, structure and format in a way that paragraphs of description cannot. If you only have time to add one of the four elements to a prompt you already use often, add the example first.

Should You Let The Model Ask Questions First?

Yes, and the easiest way is one line at the end of your prompt: “Before you answer, ask me anything you need to know.” This works because most wrong answers come from the model answering the wrong question, not from reasoning badly once it has the right one. A 2019 study on clarifying questions in open-domain conversational search found that asking a single well-chosen question improved retrieval performance by more than 170 percent compared with systems that guessed at what the user meant and proceeded. The catch is that this only helps when you read the questions and answer them. If you skip past them or tell the model to “just proceed,” you have cancelled out the benefit before it started.

Why Doesn’t Being Polite Help?

A popular claim says piling on “please” and “thank you” makes AI answers better. That part is a myth. A 2024 cross-lingual study testing prompts across three languages found that impolite prompts often produced worse answers, but overly polite language did not reliably produce better ones, and the ideal level varied by language and model. Politeness has a floor, not a ceiling: rudeness can hurt you, but extra courtesy past ordinary manners does not reliably help. If you are choosing where to spend limited editing time on a prompt, spend it on the role, the constraint, the example or the permission to ask questions. Those four change the output in ways politeness does not.

How Do You Know A Prompt Is Working?

You know a prompt is working when you stop rewriting the output after the model answers and start lightly editing it instead. That is a measurable change in your own time, not a feeling. A 2023 controlled study on GitHub Copilot found developers completed a coding task 55.8 percent faster than developers without AI assistance, finishing in just over an hour compared with nearly two hours and forty minutes for the control group. A 2025 study from METR found the opposite effect in a different setting: experienced developers using AI on unfamiliar, mature codebases sometimes took 19 percent longer than developers working without AI, largely because they trusted vague output instead of checking it. The gap between those two results sits in the prompt and the task fit, not the tool alone, which is also why I think about explaining AI-assisted work to a skeptical manager as a prompting problem as much as a communication one.

You are [a specific role tied to the task, for example: a technical writer who explains security topics to employees with no technical background]. I need [the specific task, for example: a 150-word intranet post explaining multi-factor authentication]. Here are the real constraints: [length, audience, tone, format, deadline ,  be specific, for example: 150 words, written for non-technical staff, friendly but not casual, ending with one action the reader should take this week]. Here is an example of the kind of output I want: [paste a real example, even a rough draft or a piece you liked]. Before you write anything, ask me whatever questions you need answered to do this well, and wait for my answers before drafting.

I need your help with [describe the task in one sentence], but I have not fully worked out the constraints yet. Act as a thinking partner, not just an executor. Before doing any of the actual work, ask me 3 to 5 questions that would change how you approach this task. Cover at minimum: who the audience is, how long the final output should be, what tone fits, what format it needs to take, and what 'done well' looks like to me. Wait for my answers before you draft anything.

Here is a prompt I use regularly: [paste your prompt]. Acting as a prompt engineering reviewer, check it against four elements: a clear role, real constraints (length, audience, format, deadline), an example of good output, and a line giving you permission to ask clarifying questions before answering. Tell me plainly which of the four are missing or weak. Then rewrite the prompt to add whichever elements are missing, without changing what I originally asked for. If anything about my original intent is unclear to you, ask me before rewriting it.

Act as a blunt hiring manager who has read ten thousand resumes, not a career coach. I will paste my full resume and the job description I want. Rewrite the whole resume for that role, section by section, in this order: summary, experience, skills, education. Rules: every experience line leads with impact, not duty. Use bracketed placeholders like [8 percent] for any number I did not give you, and list at the end every placeholder I need to replace with a real figure. Keep it to one page of text. Plain formatting only, no tables or columns, so screening software can parse it. After the rewrite, tell me the three weakest claims that need evidence before I send this anywhere. My resume: [paste resume]. The role: [paste job description].