How to write better AI prompts: 7 techniques with examples

Why most prompts disappoint

An AI assistant cannot see your situation. It does not know who the answer is for, what you have already tried, how long it should be or what you will do with it. When your prompt leaves those gaps, the assistant fills them with the most average guess, and you get an average answer.

A useful test, which Anthropic’s prompting guide also uses, is to imagine handing your message to a capable colleague who has never heard of your project. Could they do the job from what you wrote? If they would need to ask you three questions first, the assistant needs those answers too.

You do not need special wording or secret commands. Good prompts are mostly clear writing. The seven techniques below are ordered from the ones that help most often to the ones that polish the result. You can use one, or all seven in the same message.

Every example here is written to be copied. Use the Copy button, then change the details to match your own situation. If you would rather fill in boxes than write a prompt, the AI prompt builder turns a few answers into a structured prompt for you.

Seven techniques that make prompts work

1. State the goal and who it is for

Begin with the task in one sentence, then say who will read or use the result. The audience changes almost everything: the vocabulary, the length, the examples and the level of detail. “Explain photosynthesis” could produce a paragraph for a scientist or a story for a six-year-old. Tell the assistant which one you need.

The better version names the reader, the starting knowledge, a device that helps (a comparison) and a limit on length. Each of those removes a guess.

2. Give the background it cannot see

Context is the facts that make your request different from everyone else’s: your situation, your constraints, what has already happened and what you want to achieve. The more the answer depends on your circumstances, the more context you owe it. Explaining why you want something also helps, because the assistant can then make sensible choices about details you did not mention.

Notice the last sentence. Telling the assistant what you will handle yourself stops it from inventing names or placeholders you then have to delete. Keep private details out, though. Do not paste passwords, ID numbers, health records or other people’s information. Describe the situation in general terms instead.

3. Show an example of what you want

Describing a style is hard. Showing one is easy. If you have a sample of the tone, structure or format you want, paste it in and say so. One good example often does more than a paragraph of description. If you notice the assistant copying your example too closely, give two examples that differ from each other, so it learns the pattern rather than the wording.

I sell handmade candles. Here is one product description in the style I like: "Cedar & Smoke. A warm, woody candle for slow evenings. Burns for about 40 hours. Poured by hand in small batches." Write descriptions in the same style, in the same length, for these three candles: Lemon & Thyme, Fresh Linen, Vanilla Bean. Do not repeat phrases between them.Copy

This works for more than writing. You can show a sample row for a table you want filled in, a sample email signature, or a sample answer to a quiz question.

4. Set the format and the limits

If you do not say what shape the answer should take, you get whatever shape the assistant prefers, usually several paragraphs. Say whether you want a list, a table, numbered steps, a short paragraph or an outline. Add limits that can be checked, such as “five bullet points” or “under 120 words”, rather than vague words such as “short”.

The final instruction matters. Telling the assistant what to do when information is missing reduces the chance that it fills a gap with something made up.

5. Say what to do, not only what to avoid

Instructions made only of “don’t” are easy to misread. “Don’t be boring” tells the assistant nothing it can act on. “Write in a warm, plain style with short sentences and one concrete example” does. Both Anthropic’s and OpenAI’s prompting guides recommend describing the result you want rather than listing what you dislike. A few things you do not want are fine, as long as the main instruction is positive.

6. Let the assistant ask you questions

When you are not sure what information the assistant needs, ask it. This is one of the easiest improvements for beginners, because you no longer have to guess what to include. The assistant asks, you answer, and the final response is based on your real situation.

I want to plan a weekly study routine for an exam in six weeks. Before you write anything, ask me up to five questions you need answered to make a plan that fits my life. Wait for my answers, then write the plan.Copy

Keep the limit (“up to five”). Without it, an assistant can ask a long list. You can also use the same idea at the end of a task: “Is there anything in my request that was unclear or that you assumed?”

7. Treat the first answer as a draft

The first reply is a starting point. Professional writers rarely accept a first draft, and you should not either. Reply to the assistant as you would to a person: say what works, what does not and what to change. Specific feedback beats starting again from scratch, because the assistant keeps the good parts.

• Keep and change: “Keep the second paragraph exactly as it is. Rewrite the first to be more direct.”
• Adjust the tone: “Make this sound less formal, like a message to a colleague I know well.”
• Adjust the size: “Cut this to half the length without losing the dates.”
• Ask for options: “Give me three different openings, each in a different style.”
• Ask why: “Why did you choose that order? Would a different one be clearer?”

If a conversation goes in circles, or the replies keep drifting away from your instructions, start a new chat and paste in the improved prompt. A long, messy history can make an assistant lose track of what you first asked for.

A template you can reuse

You do not need every part every time. A quick question needs only a goal. A task that matters, such as a report, a difficult email or a study plan, benefits from all of them. Copy this, replace the square brackets and delete any line you do not need.

Goal: [what I want you to produce] Audience: [who will read or use it] Context: [my situation, what I have tried, my constraints] Example: [a sample of the style or format I like, if I have one] Format: [list, table, steps, paragraphs; length or number of items] Extra: [anything it should do when information is missing, such as ask me or write 'not stated'] Before you answer, ask me any questions you need.Copy

Here are three smaller prompts for everyday tasks, ready to adapt:

• Explain something simply

  Explain [topic] to someone who knows nothing about it. Use one everyday comparison, avoid technical words or explain them when you use them, and end with two questions I can use to test my understanding.Copy

• Improve my writing

  Here is my draft: [paste]. Edit it for clarity and grammar, but keep my voice and meaning. Show the improved version first, then list the three most important changes you made and why.Copy

• Compare options

  Compare [option A] and [option B] for someone who [my situation]. Use a table with rows for cost, effort, risks and best use. Then say in two sentences which you would lean towards for me and what could change that.Copy

Common prompt mistakes and quick fixes

What good prompts cannot fix

A clear prompt improves the answer, but it does not make the assistant infallible. Assistants can still state wrong facts in a confident tone, invent sources and misread your request. The better your prompt, the easier the answer is to judge, but you still need to read it critically. Our guide to using AI assistants responsibly covers checking facts and protecting private data, and the AI assistant FAQ answers common questions about mistakes, privacy and limits for ChatGPT, Claude and Gemini.

Rules also differ by place. If you use AI for school, work or publishing, check what your school, employer or publisher allows, and say when AI helped if you are expected to.

Prompts also depend on the tool. Assistants differ in what they can do, such as reading files or searching the web, and the options change over time. If a technique does not seem to work, check the provider’s help pages listed below for what your version supports.

Sources and further reading

This guide is written in our own words from the providers’ published advice. For the official wording and the latest details, use the sources below. We last checked them on 28 October 2026.

• Anthropic: Prompt engineering: be clear and direct
• OpenAI: Best practices for prompt engineering with the OpenAI API
• Google: Writing effective prompts for Gemini in Google Workspace