On this page
The golden rule
Write for a brilliant new colleague who knows nothing about your situation. Anthropic's guide puts it this way: show your prompt to a colleague with little context; if they'd be confused, so will the AI. (source)
AI can't read your mind or see your screen. Everything it knows about your task is in the words you give it.
The six habits every AI lab agrees on
We read the official prompting guides from Anthropic, OpenAI, Google and Microsoft. These six habits come up again and again across them.
1. Be clear and specific
Say exactly what you want. "Make it shorter" is vague; "cut it to three sentences" is clear. Numbers, names and specifics beat adjectives.
2. Give the background, and say why
Explain the situation and the purpose. "Write a product description" gets a generic one. "Write a product description for our handmade candles; our buyers are gift shoppers who care that they're plastic-free" gets a useful one. Saying why helps the AI make good choices you didn't spell out.
3. Describe the result you want
Name the format (list, table, email, code), the length, the tone and the audience. "Explain this to a 10-year-old in under 100 words" changes everything.
4. Show an example
One good example often works better than a paragraph of description. Paste a past email you liked and say "match this style". A few varied examples work best; if they're all alike, the AI may copy them too closely.
5. Give it your material
Paste or attach the document, data, notes or code you want it to use, and say which to rely on. An AI working from your actual material is far more accurate than one working from memory.
6. Treat the first answer as a draft
Reply with what to change: "more formal", "drop the second point", "now make a table". Most great results take two or three rounds.
Three more that help
- Give it a role. One line such as "You are an experienced tax accountant" focuses expertise and tone. One sentence is enough; a long backstory adds little.
- Separate your instructions from your material. Put pasted text in quotes or under a heading like "Here is the article:" so the AI doesn't confuse it with your request.
- Ask the AI to improve your prompt. Write a rough version, then ask: "Rewrite this prompt to get a better result, and ask me anything you need to know." Google and Anthropic both suggest some version of this.
A simple formula
If you like a template, Google's Persona · Task · Context · Format is easy to remember. Microsoft teaches a similar one: Goal · Context · Expectations · Source. You rarely need every part; the task is the only must-have.
Prompt builder
Before and after
Write an email about the meeting.
Write a short, friendly email to my team saying Thursday's planning meeting moves to Friday at 10:00. Apologize briefly, keep the same video link, and ask anyone who can't make it to reply by Wednesday. Under 80 words.
Why it's better: it says who it's for, what changed, the tone, the one action needed, and the length.
Explain inflation.
Explain inflation to a 15-year-old who gets a weekly allowance. Use one everyday example, then say in two sentences why governments try to keep it low. No jargon.
Why it's better: a clear audience, an example to anchor it, and a defined length.
Summarize this.
Here is a 20-page report (pasted below). Summarize it for my manager, who has two minutes. Give the three most important findings, one risk, and one recommended next step, as bullet points. Quote the report's own numbers; if something isn't in the report, say so rather than guessing.
Why it's better: it puts the document first, names the reader, fixes the shape, and tells the AI not to invent.
Copy-ready templates
Each template has blanks in [brackets]. Copy it, fill the blanks, paste it into any AI.
- Write
Write a [type of text] for [audience] about [topic]. The goal is [purpose]. Tone: [tone]. Length: [length]. Here is an example of the style I like: [example].
- Summarize
Summarize the text below for [who will read it]. Give [number] key points, then [one risk / one action]. Use only the text; say if something is missing. Text: [paste].
- Explain
Explain [concept] to [audience and their level]. Start with a one-sentence answer, then an everyday example, then [number] details. Avoid [jargon / math].
- Decide
I'm choosing between [option A] and [option B] for [goal]. What matters most to me is [criteria]. Compare them in a table, then recommend one and say what would change your mind.
- Brainstorm
Give me [number] ideas for [goal] for [audience]. Make them [constraint: cheap / quick / unusual]. Group them by [theme] and mark the three you'd try first.
- Code
Write [language] code that [does what]. Inputs: [inputs]. Output: [output]. It runs in [environment]. Explain any part a beginner might not follow, and list how to test it.
- Feedback
Act as [role, e.g. a tough editor]. Review my [text / plan / code] below for [what to judge]. List the five most important problems, most serious first, each with a fix. [paste]
- Learn
I want to learn [skill] in [time]. I already know [level]. Make a plan with one short lesson per [day/week], each with a practice task, and quiz me at the end of each lesson.
What's changed: habits to drop
Some popular prompting advice no longer holds, according to the labs' current guides.
- Stop shouting. ALL CAPS, "CRITICAL" and "you MUST" can make models overreact. Say it once, calmly. (Anthropic, Google)
- You rarely need "think step by step". Reasoning models already think before answering, and OpenAI notes that telling them to can even hurt. For hard problems, choose the AI's deeper "thinking" mode instead. (OpenAI)
- Say what to do, not what not to do. "Write in plain prose paragraphs" works better than "don't use bullet points".
- Write sentences, not search keywords. "best laptop students" gets a generic answer; a sentence with your budget, use and must-haves gets a useful one. (Google's guide)
- Don't over-pad. A long role-play setup or paragraphs of "you are the world's greatest…" add little. Real context does the work.
When the answer is wrong
| Problem | Fix |
|---|---|
| It made something up | Say "If you're not sure, say so." Ask for sources and open them. Give it the document instead of relying on its memory. |
| The answer is generic | Add your audience, goal and specifics. Paste an example of what good looks like. |
| Wrong length or format | Give a number ("5 bullet points", "under 150 words") or show the format. |
| It ignores a "don't" | Rephrase as a "do": say what you want instead. |
| Long document, weak answer | Put the document first and your question last, and ask it to quote the relevant parts before answering. |
| Your request was ambiguous | Ask it to ask you clarifying questions before it starts. |
| Math or counting errors | Ask it to use code or a calculator tool if it has one. |
| Different answers each time | That's normal. Run it again, compare, and keep the best. |
Prompting image generators
Image AIs follow the same spirit with a few extra habits, drawn from OpenAI, Google and Midjourney:
- Describe a scene in a sentence, not a pile of keywords: "a red bicycle leaning on a blue door in a sunny Lisbon street" beats "bicycle, door, street, colourful, 4k".
- Be specific about the look: subject, setting, lighting, colours, style (photo, watercolour, 3D), mood.
- Use camera words: close-up, wide shot, from above, shallow depth of field.
- Give exact numbers: "three cats", not "cats".
- Say what you want, not what you don't. Naming an unwanted object can make it appear.
- Put any words that must appear in quotes, and check the spelling in the result.
- Change one thing at a time when refining, and say what should stay the same.
- Set the shape: square, portrait or widescreen.
- Midjourney is the exception on length: it favours short, simple prompts, so trim to the essentials there.
Use AI safely
- Check anything that matters. Facts, quotes, numbers, citations, medical, legal and money advice. AI can be confidently wrong.
- Mind what you paste. Don't share passwords, ID numbers or other people's private information. Read each AI's data settings; most let you stop your chats being used for training: ChatGPT, Claude, Gemini.
- Watch for instructions hidden in content. If you ask an AI to read a web page or email, that text could contain instructions meant to trick it, called prompt injection. Be careful letting AI act on your behalf.
Watch: a prompt built live
Anthropic's prompt engineers build a real prompt from scratch, step by step, showing why each part is there.
Go deeper: the official guides
TRY IT · OPENAIPrompting best practices for ChatGPT(opens in a new tab) TRY IT · ANTHROPICPrompt engineering overview(opens in a new tab) TRY IT · GOOGLE WORKSPACEWriting effective prompts(opens in a new tab) TRY IT · MICROSOFTLearn about Copilot prompts(opens in a new tab) TRY IT · LEARN PROMPTINGA free, beginner-friendly prompting course(opens in a new tab) TRY IT · DAIR.AI PROMPT ENGINEERING GUIDEEvery prompting technique, catalogued (technical)(opens in a new tab)Questions people ask
How do I write a good AI prompt?
Say clearly what you want, give the background and the reason, describe the result you want (format, length, tone, audience), include an example or your own material if you have it, then refine the first answer with follow-up messages. The major labs' guides share these basics.
What is the best prompt formula?
A simple one from Google's guide is Persona, Task, Context, Format: who the AI should act as, what to do, the background, and the shape of the answer. You rarely need all four; the task is the only must-have.
What is prompt engineering?
The practice of wording instructions so an AI model gives the result you want. For everyday use it mostly means being clear, specific and complete. For developers it also covers testing prompts systematically and building them into apps.
Is prompt engineering still needed?
Less trickery is needed than before, since models follow plain instructions well and reasoning models plan on their own. But the basics matter more than ever: an AI can't read your mind, so clear goals, context and examples still decide the quality of the answer.
How long should a prompt be?
As long as it needs to be to remove guesswork. Google found the most effective prompts in its research averaged around 21 words, while people typically type fewer than nine. For big tasks, pasting pages of background is fine.
Do prompts work the same in ChatGPT, Claude and Gemini?
The fundamentals work everywhere, because the major labs recommend similar core habits. Each model has small quirks, so if one gives you a poor result, rephrasing or trying another AI can help.
What is few-shot prompting?
Including a few examples of the input and the output you want in your prompt, so the AI copies the pattern. One example is 'one-shot', several are 'few-shot', none is 'zero-shot'.
How we made this
Parth Amin runs thisorthis.ai and has spent over a decade building software products, most recently enterprise generative-AI systems. @_aminparth on XHow we made this: drafted with AI assistance, then fact-checked against the linked sources and edited by hand. Reviewed . Spotted a mistake? Tell us.
How AI works
Which AI should I use?
AI glossary