Prompting that works has four parts: give the model a role, state your real constraints, show one example of good output, and give it permission to ask questions before answering. Use all four when you need a first draft you can send with no edits. A one-line prompt still wins for quick, low-stakes questions where any answer will do.
Takeaways
- A prompt with a role, your real constraints, one example, and permission to ask questions produces a usable first draft far more reliably than a bare question does.
- Add the example first if you can only fix one part of a weak prompt, because models copy a shown pattern more reliably than they follow a written description of one.
- Save the full four-part prompt for anything you plan to send, publish, or act on without rereading, and use a bare prompt for quick, low-stakes lookups.
- Restate your role periodically in long chats with tools that lack a dedicated system prompt field, since instructions near the start of a long conversation can lose weight over time.
- Wait for a model’s clarifying questions before adding more instructions of your own, or the permission to ask accomplishes nothing.
| Prompt parts | 4 |
|---|---|
| Quick-ask pick | Bare prompt |
| First-draft pick | Full four-part prompt |
| Added time | About 30 seconds |
Why does a role change the answer?
Telling a model who to be narrows the range of reasonable answers before it writes a single word. Learn Prompting’s guide to role prompting defines it as assigning a persona, like food critic or mathematician, to control the style or accuracy of a response, and notes it can improve both tone and factual reliability (Learn Prompting, role prompting). Anthropic’s own prompting guidance for Claude treats a defined role as a way to set expertise and tone for an entire conversation rather than one reply (Anthropic, prompting best practices). OpenAI’s prompt engineering guide makes a similar point: establishing an identity for the model narrows what a good answer looks like before generation starts (OpenAI, prompt engineering guide). A role is not decoration. Ask a model to answer as a hiring manager reviewing two hundred resumes a week, and it drops jargon a generic assistant would keep.
What counts as a real constraint?
A real constraint is a fact about your situation the model cannot guess: your actual word limit, the tool the output has to work inside, who will read it, or the deadline you are working against. Vague constraints like make it professional do less work than concrete ones: under 150 words, no jargon, active voice, written for a manager who skims on a phone. Google’s prompting strategies guide for the Gemini API recommends stating format, length and other constraints explicitly rather than trusting the model to infer them, and shows that naming a response format directly changes structure, not just wording (Google, Gemini prompting strategies). The same logic applies when you are prepping for a specific interview instead of writing a document, a comparison covered in Which AI Should You Use To Prepare For Your Next Interview. A constraint you leave out is a guess the model makes for you, and it usually guesses generic.
Why does one example beat rules?
One good example teaches format, tone and length in a single shot; a paragraph of instructions only describes them, and models are measurably better at copying a shown pattern than parsing prose about one. Researchers call this few-shot prompting, and it traces back to the paper that introduced GPT-3, which found that including a handful of input-output examples in the prompt improved task performance with no retraining at all (Brown et al., 2020). Anthropic’s prompting guidance for Claude lists examples among the most effective tools for producing exact output, and recommends at least one full example over a longer list of rules (Anthropic, prompting best practices). The Prompt Engineering Guide’s breakdown of prompt elements lists an output indicator, essentially a format example, as one of only four core components worth including (Prompting Guide, elements of a prompt). If you can only add one thing to a weak prompt, add an example.
Should AI ask questions first?
Giving a model permission to ask clarifying questions before it answers catches a wrong assumption before it gets built into a full draft, which is cheaper to fix than editing a finished piece of writing. This only works if you wait for the questions instead of piling on more instructions in the same message. Research on zero-shot reasoning found that giving a model room to work through a problem before settling on a final answer produced large accuracy gains on multi-step tasks, lifting accuracy on one benchmark from 17.7 percent to 78.7 percent and on another from 10.4 percent to 40.7 percent using the same underlying model (Kojima et al., 2022). Permission to ask questions runs on the same logic: more room before the final answer, fewer wrong guesses. It does not help every task. For a quick fact check or a single-turn chat you plan to refine anyway, asking permission to ask questions adds a step you did not need.
The four prompt parts, ranked by what breaks without them
Not all four parts cost the same when you skip them. Ranked by how much a missing part damages the output, worst first:
- Missing example: format guesswork. Without a shown example, the model guesses your preferred length, structure and tone, and that guess drives the entire first draft. OpenAI’s own cookbook on working with large language models attributes this kind of steering to few-shot examples specifically, not to instructions alone (OpenAI cookbook), which is why this is usually the most expensive part to skip.
- Missing constraints: technically fine, practically useless. Without your real limits, the answer is competent but does not fit your actual word count, tool or audience, so you rewrite it anyway.
- Missing role: generic tone. Without a role, you get an answer that could apply to anyone in any job. A 2023 study on role-play prompting found that giving ChatGPT a role instead of a generic prompt raised accuracy on the AQuA math benchmark from 53.5 percent to 63.8 percent, and on a letter-sequencing task from 23.8 percent to 84.2 percent (Kong et al., 2023).
- Missing permission to ask: silent wrong assumption. Without room to check, the model picks the single most common reading of an ambiguous request and commits to it, and you often will not notice until you have already used the output.
Which prompt style should you use?
Match the prompt to the stakes. A bare prompt, just the question, with no role, no constraints and no example, is fine for a quick lookup you will not reuse. A role-only prompt barely improves on that, because tone without constraints still leaves the model guessing at format. Role plus constraints without an example gets close but format stays ambiguous. The full four-part prompt is worth the extra thirty seconds whenever you plan to use the output with little or no editing, including work you plan to send at your job, where it also helps to know how to explain AI-assisted work to a boss who doesn’t trust it.
| Criteria | Bare prompt | Role only | Role + constraints | Full four-part prompt |
|---|---|---|---|---|
| Best for | Quick, throwaway lookups | Tone-sensitive one-offs | Structured tasks you’ll lightly edit | Anything you’ll send or publish with little editing |
| Format guesswork | High | High | Moderate | Low |
| Time to write the prompt | About 5 seconds | About 10 seconds | About 20 seconds | 30 to 60 seconds |
| Follow-up turns usually needed | 2 to 4 | 2 to 3 | 1 to 2 | 0 to 1 |
| Risk of wrong assumption | High | High | Moderate | Low, if you wait for the questions |
Where the short prompt wins
The full four-part prompt is not the right call every time, and treating it as the default for every message slows you down for no gain. In a fast back-and-forth chat where you plan to refine the answer over three or four turns anyway, permission to ask questions just adds a turn you did not need, since you are already iterating. For a single fact lookup, a role and an example add typing without changing the answer. The three-line addition described in Fix Lazy AI Answers With Three Lines Added To Every Prompt is closer to the right amount of structure for that kind of quick, low-stakes request, and defaulting to that lighter version for anything you will not reuse saves real time over a week of prompts.
Does this work in every AI tool?
Mostly, with one meaningful difference: how well a role survives a long conversation. Microsoft’s prompt engineering guidance for Azure OpenAI models points to recency bias, meaning instructions near the end of a long prompt or chat carry more weight than instructions given at the start, and recommends repeating key instructions rather than assuming they still apply (Microsoft, prompt engineering concepts). Claude and Gemini both offer a dedicated system-level field built to keep a role active for an entire session rather than a single message, per each company’s own documentation (Anthropic, Google). The four-part structure works the same everywhere. What changes is where the role lives: a dedicated system field when the tool offers one, the first lines of your message when it does not, restated occasionally in long sessions.
Prompts you can use
Paste these straight in. Change the parts in square brackets and nothing else.
You are a prompt-writing coach helping me turn a vague request into a strong, reusable prompt. Here is what I want to do: [describe your task in one or two sentences]. Before you write anything, ask me: 1. What role should the AI play when answering this? 2. What are my real constraints: length, format, audience, deadline? 3. Do I have an example of good output to share, or should you draft one for me to react to? Once you have my answers, write the final prompt as a ready-to-paste block, then explain in two sentences why you structured it that way. Ask me anything else you need first rather than guessing.
Fill in your actual task before pasting this; it does little on a one-word topic.
You are an experienced [role, for example: recruiter, hiring manager, or client-facing consultant] reviewing a first draft on my behalf. I need: [describe the document, for example a follow-up email after a second-round interview]. Constraints: under 120 words, no generic closing lines, tone is warm but not formal, and it must reference [one specific detail from the real situation]. Here is an example of the tone and length I want: [paste a message you liked or wrote before]. Before writing the final version, ask me anything you're unsure about: the audience, the relationship, or missing details. Do not guess at facts you don't have.
Swap in your actual role, audience and a genuine example; a fabricated example produces a weaker draft than none at all.
You are a prompt engineer reviewing a prompt I already use, to tell me what is missing before I rely on it again. Here is my current prompt: [paste your prompt]. Check it against four things: does it give you a role, does it state real constraints (format, length, audience, deadline), does it include an example of good output, and does it give you permission to ask me questions before answering. For anything missing, tell me exactly what to add, then rewrite the full prompt with all four parts included. Ask me for any detail you need instead of inventing one.
Use this on a prompt you reuse often for one recurring task, not on a one-off question.
Questions people actually ask
What is role prompting in AI?
Role prompting means telling an AI assistant who to act as before it answers, such as a hiring manager or a technical editor. It narrows the range of reasonable answers and shifts tone and vocabulary toward that role, a technique both Anthropic and OpenAI document as a core prompting method.
Why does giving ChatGPT an example help?
One example shows the model your preferred format, length and tone in a single shot instead of describing them in prose. Research on few-shot learning found this improves task performance without any retraining, and it is usually the highest-impact addition to a weak prompt.
Should I let AI ask me clarifying questions before it answers?
Yes, for anything you plan to send or use with little editing, since a clarifying question catches a wrong assumption before it becomes part of a full draft. Skip it for quick, low-stakes questions where any reasonable answer is fine.
Does the four-part prompt work the same in Claude, ChatGPT and Gemini?
The structure works across all three, but where the role lives differs. Claude and Gemini offer dedicated fields that keep a role active for a full session, while a role typed into a chat can fade and need restating in long conversations.
How long should a good prompt be?
Long enough to include a role, your constraints, one example, and permission to ask questions, usually three to six sentences. Length is not the goal; each part removes a specific kind of guessing the model would otherwise have to do.
Sources
- Learn Prompting, role promptinglearnprompting.org
- Anthropic, prompting best practicesplatform.claude.com
- OpenAI, prompt engineering guidedevelopers.openai.com
- Google, Gemini prompting strategiesai.google.dev
- Brown et al., 2020arxiv.org
- Prompting Guide, elements of a promptpromptingguide.ai
- Kojima et al., 2022arxiv.org
- OpenAI cookbookdevelopers.openai.com
- Kong et al., 2023arxiv.org
- Microsoft, prompt engineering conceptslearn.microsoft.com
What happens next
Expect role and constraint settings to keep moving out of typed prompts and into dedicated interface fields, the direction Anthropic and Google have already taken with system prompts and system instructions. Watch whether more chat tools build a clarifying-question step into the interface itself instead of leaving it to the prompt. The open question is whether that makes the four-part structure less necessary to type by hand, or simply relocates it.
Take this further
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].

