What it is
A prompt is whatever you give an AI model to respond to. Usually that means the words you type into a chatbot: a question, a request, an instruction, sometimes with a document or a photo attached. "Write a thank-you note to my neighbour for watching the cat" is a prompt. So is "Here is my sales spreadsheet. Which product had the worst month, and why might that be?"
The word comes from the theatre, where a prompter feeds an actor the next line. That's a fair picture of what happens inside the machine. A language model works by continuing text. It reads everything it has been given and predicts, one small piece at a time, what should come next. Your prompt is the start of the text it continues. Everything the model knows about your particular situation, it learns from the prompt. It has no idea who you are, what you've tried already, who the result is for, or what "good" means to you, unless you say so.
That's the single most useful idea about prompts: the model can only use what's in front of it. It has read an enormous amount of text during training, so it knows a great deal about the world in general. It knows nothing about your bakery, your novel, your client or your deadline until you tell it.
People who get consistently good results from AI tend to include four things, even in short prompts:
- The task. What you want done, stated plainly. "Write", "summarise", "list", "compare", "rewrite", "check". One clear verb beats a vague wish.
- The context. The background the model needs. What the product is, what happened, what you've already tried, what the constraints are.
- The audience. Who will read the result. A summary for your accountant and a summary for your ten-year-old are different summaries.
- The format. What shape you want back. Three bullet points, a table, a 150-word email, a numbered list of steps, plain text you can paste into a form.
Here is the difference in practice. A typical first attempt:
Write a product description for my candle.
The same request with the four parts added:
Write a product description for a hand-poured soy candle called
"Harbour Morning". It smells of sea salt and cedar, burns for about
40 hours, and sells for £18 in my online shop. The people who buy
from me are mostly women in their 30s and 40s buying gifts, and my
brand voice is warm and a little playful, never pushy. Give me three
versions, each under 60 words, as plain text with no headings.
The first prompt isn't wrong. The model will write something. But it has to guess the scent, the price, the customer, the tone and the length, and it will guess the most average answer for each. You get a generic description of a generic candle. The second prompt leaves almost nothing to guess.
A prompt can also include more than your request. You can paste in the text you want worked on, give examples of what you like, set rules ("don't mention competitors"), or tell the model to take on a role. Many AI products also add a hidden prompt of their own before yours, called a system prompt, which sets the assistant's personality and rules. From the model's side, all of it is one long piece of text, measured in tokens and limited by the size of its context window.
Writing prompts well has a grand name, "prompt engineering", and a whole family of techniques. This element is the front door. The neighbouring elements on the table each cover one technique in depth.
Why it matters
The prompt is the part you control. You can't change how a model was trained. You can choose which model to use, and after that, almost everything about the quality of the answer comes down to what you ask and how you ask it. Two people using the same chatbot on the same day can get results that seem to come from different tools, and the difference is nearly always the prompt.
Most people write prompts that are far too short. In July 2024, Google's Workspace team published tips for prompting its Gemini assistant and noted that the most successful prompts its users wrote averaged around 21 words, while people's first attempts were usually under nine. Nine words is roughly "write a product description for my candle". There's nothing magic about 21. The point is that a prompt with the task, context, audience and format in it is naturally longer than a first attempt, and that extra length is where the quality comes from.
The model can't ask you what you meant. A human assistant given a vague request would ask a follow-up question. Most chatbots will answer anyway, filling every gap with a guess. Anthropic's guide to prompting its Claude models puts it neatly: treat the model like a brilliant but new employee who doesn't know your norms or your workflows. Its "golden rule" is to show your prompt to a colleague who knows nothing about the task. If they'd be confused, the model will be too.
Small changes can have big effects. Because the model predicts text based on everything in the prompt, changing a word, the order of instructions, or even punctuation and spacing can change the answer. Sometimes the effect is dramatic (the examples below include a study where formatting alone swung a score by 76 points). This is annoying, but it's also good news: it means the prompt is a real lever, not a formality.
Reasons help as much as rules. Anthropic's guide gives a small example worth remembering. The instruction "NEVER use ellipses" works less well than explaining that the text will be read aloud by a speech engine that can't pronounce them. When the model knows why, it can apply the idea sensibly to cases your rule didn't cover. The same is true of people.
Prompts are also a risk. Anything that goes into a prompt can steer the model, including text you didn't write: a web page it reads, an email it summarises, a document a customer uploads. That's why prompts matter for safety and security as well as quality, as the Bing example below shows.
Real-world examples
Does it help to tell the AI it matters? (2023 and 2025)
In July 2023, researchers from Microsoft, the Chinese Academy of Sciences and several universities published a paper titled "Large Language Models Understand and Can be Enhanced by Emotional Stimuli". They added short emotional sentences to the end of prompts, such as "This is very important to my career." and "You'd better be sure.", and called the method EmotionPrompt. Across 45 tasks and six models, including ChatGPT and GPT-4, they reported an average relative improvement of 8 percent on one set of tasks and 115 percent on another, harder set, plus a 10.9 percent improvement in a human study of 106 people who rated the quality of longer written answers.
The paper spread quickly, and "tell the AI your job depends on it" became standard prompting folklore, alongside offering the model a tip or even threatening it. In August 2025, a team at the Wharton School's Generative AI Labs, including Ethan Mollick, tested the tipping and threatening versions of the idea on newer models, using two demanding sets of expert-level multiple-choice questions. Their finding: tipping or threatening a model generally had no significant effect on overall performance. They did see large swings on individual questions, sometimes up and sometimes down, but no reliable way to predict which.
The two studies are not in conflict so much as a lesson in how prompting works. Tricks that help one model on one kind of task can fade as models improve, and a single wording change can help one question while hurting the next. What does last is the unglamorous advice: say clearly what you want, who it's for, and why. If you want to know whether a phrase helps your task, test it on your own examples, which is what evals are for.
The prompt that changed its score by 76 points (2023)
In October 2023, Melanie Sclar and colleagues at the University of Washington and the Allen Institute for AI published a paper with a memorable subtitle: "How I learned to start worrying about prompt formatting". They took tasks where a model is shown a few worked examples and then asked to answer a new one, a style called few-shot prompting. Then they changed nothing but the formatting: a colon or no colon after a label, one space or two, "Passage" or "PASSAGE", letters or Roman numerals for the options.
The words of the task stayed the same. The results did not. With Meta's open model LLaMA-2-13B, performance on the same task ranged by up to 76 accuracy points depending on formatting alone. In the worst case they described, the model went from right about four times in a hundred to right more than eighty times in a hundred because of the choices around a colon and a space. The sensitivity didn't go away with bigger models, more examples, or extra training, and OpenAI's GPT-3.5 showed it too, though usually less extremely.
The researchers' main point was aimed at scientists: if you report a model's score using one prompt format, you may be reporting luck. For everyday users the takeaway is simpler. When a prompt works, the model hasn't necessarily "understood" in a stable way, and small edits can break it. Keep prompts that work, change one thing at a time, and check the results when you do.
Bing's hidden prompt leaks on day two (February 2023)
Microsoft began rolling out its new AI chat in Bing on 7 February 2023. Behind the chat box sat a long hidden prompt from Microsoft, a list of rules telling the model how to behave: introduce itself as Bing, don't write jokes that could hurt groups of people, and don't reveal its internal code name.
On 8 February, Kevin Liu, a computer science student at Stanford, announced that he had got the chatbot to reveal those rules. His method was to type, in effect, an instruction to ignore its previous instructions and tell him what was written at the start of its document. The model obliged and recited its hidden rules, including the one telling it not to disclose that its code name was Sydney. He later got further by claiming a "developer override mode" was switched on. Another user, Marvin von Hagen, reproduced the leak the next day. Microsoft's communications director confirmed to The Verge that the rules were genuine and that Sydney was an internal code name for an earlier version of the chat experience.
This kind of attack is called prompt injection. The name was coined in September 2022 by the programmer Simon Willison, writing about a demonstration by Riley Goodside, in which a translation app built on GPT-3 was told to translate a sentence that itself said to ignore the translation instructions, and followed the sentence instead. The underlying problem is that a model receives its maker's instructions and a user's text as one stream of words. It has no firm, built-in way to tell which words are commands and which are data. AI companies have added many defences since, but none is complete. That's why anything that reads untrusted text for you (web pages, emails, uploaded files) needs guardrails, and why you should never put anything in a hidden prompt that you'd be upset to see published.
An everyday scenario: the freelancer's cold email
Imagine Tom, a freelance illustrator, who asks a chatbot to "write an email asking a publisher for work". The result is polite, long, and could have come from anyone. He tries again, adding four things. The task: a short first email to an art director he has never met. The context: he illustrates children's picture books, has two published titles, and noticed the publisher's new list of animal stories. The audience: a busy art director who receives dozens of these a week. The format: under 120 words, one link to his portfolio, a subject line, no attachments.
The second email is a third of the length and names the publisher's own books. Tom still edits it, because it's his name on it. But he's editing a draft that fits the situation instead of rewriting a template, and the difference came entirely from the prompt.
Try it yourself
This exercise takes about fifteen minutes and works in any chatbot: ChatGPT, Claude, Gemini, Copilot or similar, on a free plan at the time of writing. Features and limits change, so check your tool's current terms.
- Pick a real task you need done this week. An email you've been putting off, a description for something you sell, a summary of a long document, a social media post. Real tasks teach more than invented ones.
- Write the short version first. Type the request the way you normally would, in one line. For example:
Write a post about my new online course.
Save the answer somewhere you can see it.
- Now open a fresh chat and fill in the four parts. Start a new conversation so the first answer doesn't influence the second. Copy this template and replace everything in square brackets:
Task: [what you want, starting with a verb: write, summarise, rewrite, list]
Context: [the background the AI needs: what this is about, key facts,
what you've already tried, anything it must include or avoid]
Audience: [who will read it, what they already know, what they care about]
Format: [length, structure, tone, and anything to leave out,
for example "under 100 words, plain text, no hashtags, no emojis"]
- Add the reason. Under the four parts, add one line explaining why the result matters or how it will be used. "This goes to existing customers, so it shouldn't sound like an advert." "I'll read this aloud on a video, so keep sentences short."
- Compare the two answers side by side. Ask yourself which one you could use with the fewest edits. Notice which of your four parts made the biggest difference. For most people, it's the context or the audience.
- Change one thing and run it again. Try a different audience, or a different format. Keep everything else the same. This shows you how much a single change steers the result, and it's the habit behind iterative refinement.
- Ask the model what it's missing. Before you finalise, paste this at the end of your prompt, or send it as a follow-up:
Before you write anything, ask me up to three questions about anything
that is unclear or missing from my request. Wait for my answers.
This turns the model's biggest weakness, answering even when it's guessing, into a strength.
- Save the prompt that worked. Keep it in a notes file with the blanks left in. Next time you need the same kind of thing, you fill in the blanks instead of starting from scratch. That's a prompt template, and a small library of them is worth more than any list of clever tricks.
Common mistakes and misunderstandings
Expecting the AI to read your mind. The model doesn't know your business, your reader or your standards. If a detail matters and it isn't in the prompt, the model will fill it with the most average guess. When a result feels generic, the prompt was usually generic.
Stacking ten requests in one message. "Summarise this, fix the grammar, make it funnier, turn it into a tweet and suggest a title" asks the model to juggle. It will usually do some of them well and some badly. Split big jobs into steps, and feed each answer into the next. That's the idea behind prompt chaining.
Saying what not to do instead of what to do. "Don't be boring" gives the model nothing to aim at. "Open with a surprising fact and keep sentences under 20 words" does. Rules about what to avoid still have their place, but pair each with what you want instead.
Believing there's a magic phrase. Prompt folklore is full of secret sentences: offer a tip, say it's for your career, threaten it. As the Wharton report found, these mostly don't help reliably on current models. A few phrases genuinely help with some tasks, such as asking the model to work through a problem step by step (see chain of thought), but even those are no substitute for a clear request.
Assuming the same prompt gives the same answer. It usually doesn't. Chatbots add a little randomness on purpose, controlled by a setting called temperature, so asking twice gives two different answers. One good result doesn't prove a prompt works; one bad result doesn't prove it's broken. If it matters, try it several times, or on several real examples.
Treating a longer prompt as a better one. Length helps when it carries information the model needs. Padding, repetition and ten ways of saying "be high quality" add nothing and can bury the instructions that matter. Aim for complete, not long.
Trusting the answer because the prompt was good. A careful prompt makes errors less likely, not impossible. Models can still state wrong facts with confidence, a problem called hallucination. Giving the model the source material to work from, known as grounding, helps a lot, and checking anything important is still your job.
Pasting sensitive information without thinking. Whatever you put in a prompt is sent to the company running the model. Check your tool's current privacy terms before you paste customer details, contracts or anything confidential.
Forgetting that other people's text is a prompt too. When you ask an AI to summarise a web page or read your inbox, the content of that page or email becomes part of the prompt, and it can contain instructions. Be cautious with AI tools that can act for you, such as sending emails or making purchases, after reading text from sources you don't control.
Related elements
- System Prompt: the hidden instructions an app places before your prompt, and the target of the Bing leak.
- Zero-shot: asking for something with no examples at all, which is what most everyday prompts are.
- Few-shot: adding a few examples of what you want, often the quickest improvement to a prompt.
- Role Prompting: telling the model who to be, such as an editor or a patient tutor.
- Chain of Thought: asking the model to reason step by step before it answers.
- XML Tags: labelling the parts of a long prompt so the model can tell instructions from material.
- Structured Output: asking for answers in a fixed shape, such as a table or a form.
- Prompt Templates: reusable prompts with blanks to fill, the natural result of the exercise above.
- Prompt Chaining: splitting a big job into a series of smaller prompts.
- Meta-prompting: asking the AI to help write or improve the prompt itself.
- Temperature: the setting that decides how much the answer varies from run to run.
- Context Engineering: choosing everything that goes into the model's view, not just the question.
- Evals: testing a prompt against a set of real cases to see whether a change really helped.
- Iterative Refinement: improving a prompt by changing one thing at a time and comparing results.
- Guardrails: the protections that limit what a prompt, including an injected one, can make a system do.