What is a prompt?

A prompt is the text you give an AI model to tell it what you want. It includes not only the request itself, such as “Write a…”, but also any background, source material you want the answer based on, and examples of what a good answer looks like.

AI chat tools can give very different answers to the same request depending on how you phrase it. When you ask for a class announcement, for example, just adding who it’s for and what you want to get across can bring the draft much closer to something you can use. Knowing the basics of prompts helps you get better results from AI, and helps you spot the problems that no amount of rewording will solve.

What is a prompt made of?

A prompt can contain five kinds of content. You don’t need all five every time. Add what the request calls for.

Part	What it is	Example (yoga studio announcement)
Instruction	What you want done	Write an announcement for a trial class
Background	Purpose, situation, audience	We want people who haven’t exercised in a while to feel welcome
Source material	Information the answer should be based on	A table of dates, prices and what to bring
Examples	Samples of the answer you want	A past announcement people responded well to
Format	Length, structure, tone	Friendly, no headings, short paragraphs

Asking without examples is called zero-shot prompting. Including a few examples is called few-shot prompting. Examples are especially good at conveying a tone or structure that’s hard to describe in words.

An AI model reads your prompt and writes a natural continuation of it. What is a large language model (LLM)? explains how that works. The model can’t guess things you didn’t write down, so it helps to think of the prompt as the only briefing it gets.

How is a system prompt different from a user prompt?

Inside an AI-powered service, there’s usually more than one prompt. There are two main kinds.

	System prompt	User prompt
Who writes it	The people who build the service	The person using it
Example	“Answer as the front desk of a cooking school. For booking changes, give the phone number.”	“Are there still places in next Saturday’s class?”
Visible to users?	Usually not	Yes
Applies to	The whole conversation	That one request

A system prompt sets the AI’s role, tone, what it may answer and the rules it follows, before any user types anything. In OpenAI’s prompt engineering guide these are called developer messages, and they take priority over user messages. Different system prompts are one reason two services can answer the same question differently.

A user prompt is what you type into the chat box. As a conversation goes on, the earlier messages are sent to the AI along with each new one and shape the next answer.

Some chat services also let you save a description of your work or your preferred tone in the settings, so it’s added to every request. If you keep typing the same introduction, a setting like that saves time.

How do you write a request that gets understood?

It helps to imagine the AI as a capable new colleague who doesn’t yet know anything about your business. Anthropic’s prompting best practices suggest a simple test: if a colleague with little context would be confused by your prompt, the AI probably will be too.

When you ask, include:

1. The goal: what the result is for (“an announcement that gets people to try a class”)
2. The audience: who will read it (“people who haven’t exercised in a while”, “a client’s project manager”)
3. Conditions: what to stick to (length, information to include, phrases to avoid)
4. The format: a table, a bulleted list, an email
5. The reason: a few words on why a condition matters (“most people read this on a phone, so keep paragraphs short”)

Here’s how a small design studio might ask for a reply to an inquiry:

	Request
Hard to act on	“Write a reply to this inquiry.”
Easy to act on	“Write a reply email to a first-time client who asked how much a logo costs. Use only the prices in the price list below. Keep it polite but not stiff, and end by asking for times that suit them for a call.”

Saying what to do works better than only saying what not to do. Instead of “Don’t use jargon”, try “Use words someone new to websites would understand”. When the order of steps matters, number them.

When you share source material

When you need an answer based on facts such as prices, dates or terms, paste the source material in with your request. Mark clearly where it starts and ends, for example with “Source material:” and “End of source material”, so the AI doesn’t mix up your instructions and the material.

Say you run an online course and want a draft reply to a refund question. Paste in your course terms and add: “If the terms don’t cover it, say you don’t know” and “Note which part of the terms you used”. That makes invented details less likely and makes the answer easier to check. When there’s too much material to paste at once, a technique called RAG is often used. See What is retrieval-augmented generation (RAG)?

When you give examples

Examples make the tone and format more consistent. But the AI may copy their wording too closely. Show a few examples from different situations, and say “use these for the format only” to keep it from copying.

What do you do when the answer isn’t right?

You won’t always get the answer you want on the first try. Google’s prompt design strategies for Gemini describe prompt design as an iterative process of trying and revising. Some ways to revise:

• Say exactly what’s wrong. Instead of “make it better”, try “make the wording softer and cut it to half the length”.
• Add the conditions or source material that were missing.
• Split a big task. Ask for an outline first, then the full text once you’ve agreed on the outline.
• When a conversation gets long, summarize the key points and start a new one. Conditions from early in a long conversation can lose their effect.
• Save prompts that work as templates, with blanks for dates or names, and reuse them.

Even the same prompt produces slightly different wording each time. For important text, it’s worth asking for a few versions and comparing them.

What can’t a prompt fix?

Better prompts get better answers, but some problems can’t be solved by wording alone.

Confident mistakes

Writing “Be accurate” or “Don’t make mistakes” won’t stop hallucinations, where an AI states something false as if it were true. A model can produce a natural-sounding answer even when it doesn’t know. What is a large language model (LLM)? explains why and how to reduce it. For answers that involve facts, provide sources and have a person check before using them.

Instructions hidden in the text

When you paste an email or web page and ask for a summary, the text itself may contain instructions aimed at the AI. This is called prompt injection, and careful wording alone can’t reliably prevent it. What is an AI agent? explains how to guard against it.

What happens to what you paste

Writing “Don’t remember this” doesn’t change how your input is stored or used. The service’s terms and settings decide that.

• Don’t paste passwords, card numbers or API keys.
• Replace customers’ names and contact details, for example with “Customer A”, unless the task needs them.
• Before using a service for work, check how it handles your data, including whether your input is used to train its models.

Summary

• A prompt is your request to an AI, made up of instructions, background, source material, examples and format.
• There are system prompts, written by the people who build a service, and user prompts, written by the person using it.
• State the goal, audience, conditions, format and reasons, and provide facts as source material.
• Don’t expect perfection on the first try. Say exactly what’s wrong and revise.
• Wording alone can’t prevent confident mistakes, hidden instructions or data-handling issues, so rely on human checks and be careful about what you paste.