A good prompt tells the AI what you want, why, for whom and in what shape. For chat assistants such as ChatGPT, Claude, Gemini and Copilot, that comes down to six elements.
- Goal: say exactly what you want done.
- Context: give the background the AI cannot guess, including why you need it.
- Role or audience: say who the AI should write as, and who will read it.
- Format and length: say what shape the answer should take.
- Examples: show one if you have it.
- Constraints: say what to avoid, and what to do instead.
Image prompts are different: they need a subject, lighting, a camera or style reference and setting. They have their own section near the end.
This page covers the elements of a good prompt. For writing your first prompts step by step, see How to write AI prompts. For habits that stay useful whichever model you use, see How to ask AI well.
A good prompt tells the AI what you want, why you want it, who it is for and what shape the answer should take. The more of that you say, the less the AI has to guess. Google’s guide for Gemini names four areas: persona, task, context and format.1 Microsoft’s guide for Copilot uses a different four: goal, context, expectations and source.2 This page combines what the vendors’ own guides say into six elements, with weak and better examples for everyday tasks.
Write a prompt as if you were briefing a capable new colleague who knows nothing about your situation. Anthropic, which makes Claude, uses the same picture in its own documentation: “Think of Claude as a brilliant but new employee who lacks context on your norms and workflows.”3
What makes a good prompt? The six elements
The plain point: you do not need all six every time. Microsoft’s guide says the goal is the only part you must have, and the rest makes the answer more specific.2 A prompt is simply the message you send; these are the parts that make it work.
| Element | What it answers | Example |
|---|---|---|
| 1. Goal | Answers: what do you want done? | Like: “Draft a reply turning down the meeting” |
| 2. Context | Answers: what is the situation, and why do you need it? | Like: “I am already booked that day and want to keep the client happy” |
| 3. Role or audience | Answers: who is writing, and who is reading? | Like: “Write as a friendly account manager to a long-standing client” |
| 4. Format and length | Answers: what shape should the answer take? | Like: “Three short sentences, no bullet points” |
| 5. Examples | Answers: what does good look like? | Like: “Here is a reply I liked last time: …” |
| 6. Constraints | Answers: what must it avoid, and what should it do instead? | Like: “Do not apologise more than once; offer two other dates instead” |
1. Goal: say exactly what you want done
Start with the task, in plain words. “Help with my email” leaves the AI to guess whether you want it written, shortened, checked or translated. “Rewrite this email so it is shorter and more polite” does not. Anthropic’s documentation says being specific about the output you want can improve results.3
2. Context: why you want it, and what the AI needs to know
The AI only knows what you tell it. Say what the situation is, what you already know, and why you need the answer. Google’s developer guide for Gemini advises giving the model the information it needs rather than assuming it has it.4 Anthropic adds that explaining why helps too.3
“I’m pitching to a room of 40 retail buyers who know nothing about AI” is context. “Write a pitch” is not.
3. Role or audience: who is writing, and who is reading
Telling the AI whose voice to use, and who the reader is, changes the tone and the level of detail. Anthropic says even one sentence setting a role makes a difference.3 Google calls this the “persona”. “You are a senior copywriter at a luxury fashion brand targeting women aged 25–40” is a defined role. “You are a copywriter” is not. Naming the audience works the same way: “explain it to a 12-year-old” and “explain it to a nurse” should give very different answers. A role you want in every chat can go in a system prompt or custom instructions.
4. Format and length: the shape of the answer
Say what you want back: a list, a table, a short email, three options, under 100 words. Anthropic’s guide says to be specific about format and constraints, and Google’s says the same about response format.34 “Output as a markdown table” or “write in exactly 80 words, no more” narrows the answer the way a camera and lens narrow an image.
5. Examples: show what good looks like
One good example often does more than a paragraph of description. This is called few-shot prompting. Anthropic calls examples one of the most reliable ways to steer the format and tone of an answer, and Google’s Gemini guide recommends always including them.34 Paste an email you liked, a paragraph in your house style, or the first row of the table you want.
6. Constraints: what to avoid, and what to do instead
Limits stop the AI falling back on its most common habits: “No filler phrases. No exclamation marks. Don’t start sentences with ‘I’. Avoid corporate jargon.” Google’s guide lists constraints as their own step.4 One refinement from Anthropic: where you can, say what to do rather than only what not to do.3 “Write in plain paragraphs” works better than “no bullet points”.
Goal, context, role or audience, format, examples and constraints. Only the goal is required; each extra element removes one more thing the AI would otherwise have to guess.
Weak prompt vs better prompt: everyday examples
The plain point: the better versions are not cleverer. They just answer the questions the weak versions leave open.
| Task | Weak prompt | Better prompt |
|---|---|---|
| An email | Weak: “Write an email to my landlord about the heating.” | Better: “Write a short, polite email to my landlord asking for the heating to be repaired. It stopped working on Monday and I reported it by phone on Tuesday. Ask for a repair date this week. Under 120 words, friendly but firm.” |
| A summary | Weak: “Summarise this.” | Better: “Summarise this report for my manager, who has two minutes before a meeting. Give the three main findings as bullet points, then one sentence on what we should decide. Leave out the methods section.” |
| A plan | Weak: “Make me a workout plan.” | Better: “Make a four-week beginner plan for someone who can walk 30 minutes but has never run. Three sessions a week, no gym. Show it as a table with one row per session.” |
| An explanation | Weak: “Explain inflation.” | Better: “Explain inflation to a 14-year-old in under 150 words, using an example about the price of a school lunch. End with one question to check they understood.” |
Then check what comes back. Microsoft’s guide warns that the same prompt can give different answers on different runs, and tells users to review and verify the responses.2
A simple formula you can copy
The plain point: fill in the brackets you need and delete the rest.
Short is fine if it is specific. Google’s guide for Gemini in Workspace says to state requests “in brief — but specific — language”.1 And expect a second round: Microsoft says most people follow up on the first answer with another prompt.2 The step-by-step version is on how to write AI prompts.
Copy the formula, fill in only the brackets that matter for the task, and plan to refine the first answer rather than expecting it to be final.
If you are making images
The plain point: image and video models need different elements: a clear subject, lighting, a camera or style reference, a setting, and care with quality tags and negative prompts. Each vendor publishes its own formula, collected on what the vendors actually say, and there is a cheat sheet per model and a printable images and video cheat sheet.
A clearly defined subject
The model needs to know exactly who or what it is drawing. “A woman” gives the model infinite interpretations. “1woman, late 20s, Southeast Asian, long dark hair, angular jawline, confident expression” removes nearly all of them.
The subject should usually come early, because order matters on at least some models — Black Forest Labs states this outright for FLUX.2, that it “pays more attention to what comes first.”5
Not every vendor agrees. Google’s published formula for Veo opens with the camera, not the subject: [Cinematography] + [Subject] + [Action] + [Context] + [Style & Ambiance]. If you are prompting video on Veo, lead with the shot. Formula per vendor.
Specific lighting
“Good lighting” means nothing to a model. “Soft Rembrandt key light from upper left, warm fill from the right, subtle rim light” means something specific. Named setups include Rembrandt, clamshell, chiaroscuro, golden hour, blue hour, window light and practical neon. Describing the light works better than borrowing a name: a director’s or cinematographer’s name transfers a look unpredictably, and some tools restrict imitating living artists’ styles. See director styles in prompts. Left unspecified, lighting is chosen for you, and often comes back flat and even.
A camera reference and a setting
Naming a camera and lens is a quick way to signal a photographic look: “Shot on Canon EOS R5, 85mm f/1.2” suggests shallow depth of field and the compression of a portrait lens. A camera family works too, such as “medium format Hasselblad” or “ARRI ALEXA anamorphic”. The setting carries mood: “Rain-soaked alley at midnight with neon reflections” communicates colour, light and atmosphere in one phrase, where “Outdoor setting” communicates nothing.
Quality and style anchors cut both ways
Terms like “8K, ultra-detailed, commercial photography standard” pull output toward polished, professional, advertising-grade imagery, because that is the material they are associated with in training.
Know which way you are pulling. That polished look is the opposite of documentary realism — if you want something that looks like an actual photograph rather than a campaign image, these are the first words to delete. See why AI output looks fake.First-hand: an earlier version of this element recommended these tags without qualification, which contradicted this site’s own guidance on realism. Both effects are real; the tags help for commercial polish and hurt for realism, and the page now says which is which. Corrected 31 Aug 2026.
Negative prompts: check your model first
Telling the model what to avoid depends entirely on the model. Image models split: Stable Diffusion and ComfyUI pipelines expose a real negative-prompt field, while Black Forest Labs states that FLUX.2 does not support negative prompts and asks you to describe what you want instead.6 Video models mostly want identity-lock instructions rather than a ban list. Where there is no field, turn the list into positive description: “visible skin texture and pores” does the job that “no plastic skin” was doing. More on negative prompts.
For images, put a tightly defined subject first unless the vendor’s formula says otherwise, name the light, choose quality tags for the look you want, and check whether your model reads negative prompts before writing a ban list.
Questions people ask
What are the key elements of a good prompt?
A clear goal, the context, a role or audience, the format and length you want, an example, and any constraints. Vendors group them differently: Google names persona, task, context and format, and Microsoft names goal, context, expectations and source, but both say a clear task comes first.2
How long should a prompt be?
As long as it needs to be to remove the guesswork, and no longer. Google’s Workspace guide asks for brief but specific language, while Anthropic says the more precisely you explain what you want, the better the result; length on its own is not the aim.1
Do the same rules work in ChatGPT, Claude and Gemini?
The basics do: OpenAI, Anthropic, Google and Microsoft all advise clear instructions, context, examples and a stated format. But OpenAI notes that different kinds of model can need different prompting, so test a prompt that matters in the app you will actually use.7 Prompting for LLMs covers the differences.
Why does the same prompt give different answers?
Chat assistants do not give a fixed answer to a fixed question, so a rerun can come out differently. Microsoft says so directly in its Copilot guide, which is one more reason to check the answer you keep.2 See temperature and prompt engineering for the terms.
Related on this site: Prompting Techniques Explained · Starter Prompts for Everyday Tasks, by Job · Free Prompt Checker · The Six-Part Prompt Checklist · Prompt Frameworks Compared
The weak and better prompts, the formula, the table examples and the new-colleague comparison on this page are this site’s own ways of explaining, not claims from these sources; the “brilliant but new employee” picture is Anthropic’s own.
SOURCES- 1Google Workspace, Writing effective AI prompts for business (Prompting guide 101), read at source 5 Oct 2026“The four main areas to consider when writing an effective prompt are: Persona, Task, Context, Format.” ↑a“Be concise and avoid complexity. State your request in brief — but specific — language.” ↑b“Be concise and avoid complexity. State your request in brief — but specific — language.” ↑c
- 2Microsoft Support, Get started writing prompts in Microsoft Copilot, read at source 5 Oct 2026“Prompts can include four parts: the goal, context, expectations, and source” ↑a“All that’s required is a clear goal.” and “You’ll often need to include more than a goal to get the results you want.” ↑b“Review and verify responses you get from Copilot.” and “Using the same prompt multiple times can result in different responses.” ↑c“Most likely, you’ll follow up on the results with another prompt. Expect some back-and-forth conversation to get the results you’re looking for.” ↑d“All that’s required is a clear goal.” ↑e“Using the same prompt multiple times can result in different responses.” ↑f
- 3Anthropic, Prompting best practices (Claude developer docs), read at source 5 Oct 2026“Think of Claude as a brilliant but new employee who lacks context on your norms and workflows. The more precisely you explain what you want, the better the result.” ↑a“Claude responds well to clear, explicit instructions. Being specific about your desired output can help enhance results.” ↑b“Providing context or motivation behind your instructions, such as explaining to Claude why such behavior is important, can give Claude more information about your goals and help it deliver more targeted responses.” (wording re-read at source 12 Nov 2026) ↑c“Setting a role in the system prompt focuses Claude’s behavior and tone for your use case. Even a single sentence makes a difference” ↑d“Be specific about the desired output format and constraints.” ↑e“Examples are one of the most reliable ways to steer Claude’s output format, tone, and structure. A few well-crafted examples (known as few-shot or multishot prompting) improve accuracy and consistency.” ↑f“Tell Claude what to do instead of what not to do”, with the example: instead of “Do not use markdown in your response”, try “Your response should be composed of smoothly flowing prose paragraphs.” ↑g
- 4Google AI for Developers, Prompt design strategies (Gemini API docs), last updated 17 Sep 2026, read at source 5 Oct 2026“You can include instructions and information in a prompt that the model needs to solve a problem, instead of assuming that the model has all of the required information.” ↑a“You can give instructions that specify the format of the response.” ↑b“We recommend to always include few-shot examples in your prompts.” ↑c“Specify any constraints on reading the prompt or generating a response.” ↑d
- 5Black Forest Labs FLUX.2 prompting guide, read at source 17 Sep 2026“Word order matters - FLUX.2 pays more attention to what comes first.” An earlier version of this note quoted the guide’s negative-prompt sentence instead, and said all models weight earlier tokens more heavily. ↑
- 6Black Forest Labs, FLUX.2 prompting guide, read at source 23 Sep 2026“FLUX.2 does not support negative prompts. Focus on describing what you want, not what you don’t want.” First-hand: an earlier version of this element said avoiding was "as important as" including and presented negative prompting as a general image-model technique — which contradicted this site’s correction of 25 Aug 2026. Corrected 31 Aug 2026. ↑
- 7OpenAI, Prompt engineering (API docs), read at source 5 Oct 2026“Different model types (like reasoning versus GPT models) might need to be prompted differently to produce the best results.” ↑
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK