How to write better AI prompts: a simple structure, real examples and mistakes to avoid

Ask an AI assistant to "write an email to my boss" and you'll get something correct and bland. Tell it who your boss is, what happened, what you want to achieve and what tone you need, and you'll get something you can almost send as-is. The difference isn't the tool — it's the prompt, the instruction you write.

This guide covers a structure that works with any assistant (ChatGPT, Gemini, Claude, Copilot and others) and lines up with what the companies behind them recommend in their own documentation.

#The short answer

A good prompt usually has four parts: context (who you are and what's going on), task (exactly what you want), format (how you want the answer) and criteria (what makes the answer good or bad). It doesn't need to be long — it needs to be specific.

#The four-part structure

#1. Context: what the AI can't guess

The model knows nothing about your situation except what you tell it. Include what a person would need to know to help you well:

• Who you are or what role you're writing from ("I teach Year 8 science", "I run a small online shop").
• Who the output is for ("for customers who aren't technical").
• What's already happened or what you have ("here's my current draft", "I already tried X and it didn't work").

#2. Task: a clear verb and a goal

Use a concrete verb — summarise, compare, rewrite, suggest, fix, explain — and add the purpose. "Summarise this report" is different from "summarise this report so I can decide whether it's worth reading in full."

#3. Format: how you want to receive it

If you don't say, the assistant will choose for you. Specify:

• Approximate length ("under 150 words", "three paragraphs").
• Structure ("a table with columns X, Y and Z", "a numbered list of steps").
• Tone ("professional but friendly", "no jargon").

#4. Criteria: what to avoid or achieve

Much of the quality lives here: "if you're not sure about a fact, say so", "don't make up numbers", "avoid clichés", "prioritise free options".

#Before and after

Weak prompt:

Write an email asking for an extension.

Better prompt:

I work at a small consultancy and owe a client a report on Friday. Because the client sent us their data late, I need one more week. Write a short email (under 120 words) to the client's project lead asking for an extension. Professional and warm, without directly blaming the client. Suggest the following Friday as the new deadline and offer a 15-minute call if they'd prefer.

The second version has context (consultancy, client, reason), a task (ask for an extension), a format (short, under 120 words) and criteria (tone, no blame, propose a date and a call). The result needs far fewer edits.

#Techniques that improve almost any prompt

#Show examples of what you want

If you're after a particular style, include one or two examples. Models are very good at copying a pattern once they see it. This is especially useful for repetitive formats such as product descriptions, customer replies or quiz questions.

#Keep instructions separate from material

When you paste in long text, separate it clearly from your instructions — with triple quotes or a label like "TEXT:". That way the model doesn't confuse what it should do with what it should work on. Both OpenAI's and Anthropic's guides explicitly recommend delimiters like these.

#Break complex tasks into steps

For multi-step problems — calculations, decisions with many factors, debugging code — ask the assistant to lay out its reasoning before the conclusion, or split the task yourself and ask for one step at a time. Intermediate steps are much easier to check.

#Give it permission to say "I don't know"

A line like "if you don't have reliable information about something, say so instead of guessing" reduces the tendency to fill gaps with invented details. It doesn't eliminate it, so you still need to fact-check the answers.

#Iterate instead of starting over

The first answer is a draft. Reply with specific edits: "shorter", "drop the second point", "make the tone more direct", "add an example with numbers". Each change builds on the last within the same conversation.

#A reusable template

Copy this and fill it in:

Context: [who you are, who it's for, what's going on]
Task: [verb + what you want + why]
Material: """[text, data or draft, if any]"""
Format: [length, structure, tone]
Criteria: [what to avoid, what to prioritise, what to do if information is missing]

You won't need every line every time. For simple questions, context plus task is enough.

#Common mistakes

#What a good prompt can't fix

Even an excellent prompt won't turn a model into a reliable source of current facts, or an expert on topics where its knowledge is thin. If you need recent information, use a tool with web search and check the sources. And if the job calls for a different kind of tool — transcribing audio, generating images, working with spreadsheets — see which AI tool to use for each task.

#The bottom line

Good prompting mostly means giving the AI what someone who knows you would take for granted: context, goal, format and criteria. Start with the template, refine the first result, and always check important facts. To see which habits do the most damage, read about the most common chatbot mistakes.