The most useful AI prompt examples share one pattern: they say who the output is for and what shape the answer should take. A weak prompt asks for "a summary". A strong one asks for "three bullet points for a manager who missed the meeting, each under twenty words". The rewrites below show that change across writing and images among other jobs, so you can copy the pattern along with the prompt.
"Write me something good" is a prompt in the same way "food" is a dinner order. The model will bring you something, and you will eat what arrives.
Each example comes as a before and an after, with a line on why the second version works. The why is the part to keep, because the exact wording will age faster than the reasoning behind it.
What makes AI prompt examples work
A good prompt leaves the model nothing important to guess. Every gap you leave gets filled with the most average answer available.
Anthropic's prompting best practices suggest treating the model as a brilliant new employee who lacks context on your norms. Their golden rule is a useful test: show the prompt to a colleague with no background on the task. If the colleague would be confused, the model will be too.
Most strong prompts carry the same parts. They do not need labels, but each one should be findable:
- The task: one verb and one object, such as "rewrite this email".
- The reader: who will use the output and what they already know.
- The context: the source material, pasted in full.
- The format: length and structure, stated as numbers where possible.
- The reason: why the constraint matters, so the model can handle cases you did not foresee.
The last item is the one people drop. The same Anthropic guide notes that explaining the motivation behind an instruction helps the model deliver more targeted responses. "Keep it short" is an instruction. "Keep it short, because it will be read on a phone between meetings" is an instruction the model can reason about.
Generative AI prompt examples for everyday writing
Writing tasks improve the most from a rewrite, because vague writing prompts produce fluent, empty text that looks finished.
Before:
Summarise these meeting notes.After:
Summarise the meeting notes below for a manager who missed the meeting.
Give three bullet points: decisions made, open questions, and who owns
each next step. Keep each bullet under 20 words. If the notes do not
record an owner, write "owner unclear".
<notes>
[paste notes]
</notes>The second version names a reader and fixes the length. It also tells the model what to do when information is missing. That final instruction matters most. Without it, the model will invent a plausible owner, and the meeting gains an action item for someone who was on holiday.
The same shape works for email replies:
Draft a reply to the customer email below. The customer is frustrated
about a late delivery. Apologise once, give the new delivery date
(14 October), and offer free shipping on the next order. Keep it under
120 words, plain sentences, no exclamation marks.Tone instructions land better as concrete rules. "Be professional" gets interpreted. "Apologise once, no exclamation marks" gets followed. For longer pieces where voice matters, the guide on how to prompt AI to write like a human covers pasting voice samples, which beats describing a tone in adjectives.
Resume prompts for AI that keep the facts yours
AI is good at tightening a resume and bad at knowing what you did. Give it the facts and ask it to edit them.
Before:
Write a resume for a project manager.After:
Rewrite the resume bullets below for a senior project manager role at a
logistics company. The job ad is pasted under the bullets.
Rules:
- Use only facts that appear in my bullets. Do not add numbers, tools
or achievements I have not listed.
- Start each bullet with a past-tense verb.
- Where a bullet has no measurable result, flag it with [ADD METRIC]
so I can fill it in myself.
<bullets>
[paste bullets]
</bullets>
<job_ad>
[paste job ad]
</job_ad>The rule about facts carries the weight here. Asked to make a resume "stronger", a model will add percentages that sound right, and you will be the one explaining them in an interview. The [ADD METRIC] flag turns that failure into a to-do list.
Pasting the job ad lets the model match the employer's language without inventing experience. It also cuts down on the verbs every model reaches for. Recruiters have now read "spearheaded" more often than any person should, usually about a spreadsheet.
AI image prompt examples, including negative prompts
Image prompts work best as a description of a photograph that already exists, down to the light and the lens.
Before:
A Christmas family photo.After:
A candid photo of a family of four laughing on a sofa on Christmas
morning, wrapping paper on the floor, warm window light from the left,
shot on a 35mm lens at eye level, shallow depth of field, natural
skin tones.
Negative prompt: text, watermark, extra fingers, plastic-looking skinThe rewrite reads like a photographer's shot list. Light direction and lens choice give the model decisions it would otherwise make at random. Ask for "a festive family photo" with no detail and you tend to get six strangers in matching jumpers, all holding the same mug.
The last line is a negative prompt, a separate field that lists what to keep out of the image. Tools built on Stable Diffusion accept one directly. Chat-based image tools usually do not, so fold it into the main prompt as a plain sentence: "no text or watermarks in the image". To reverse-engineer a prompt from a picture you already like, see how to create an AI prompt from a photo.
Prompt engineering in AI: putting examples inside the prompt
The strongest prompt technique is to show the model an example of the output you want. Describing a format takes a paragraph. Showing it takes one sample.
This is called few-shot prompting. OpenAI's prompt engineering guide describes it as including a handful of input and output examples in the prompt, and recommends covering a diverse range of inputs. Anthropic suggests three to five examples, each wrapped in <example> tags so the model can tell them apart from the instructions.
Classify each support ticket as billing, bug or feature request.
<examples>
<example>Ticket: "I was charged twice in March." → billing</example>
<example>Ticket: "The export button does nothing." → bug</example>
<example>Ticket: "Can you add dark mode?" → feature request</example>
</examples>
Ticket: "[ticket_text]" →Pick examples that differ from each other. Three near-identical samples teach the model one narrow case. The model also copies examples faithfully, including their mistakes, so a typo in example two will turn up in the output with full confidence.
Where the prompt sits changes how it is weighted. Chat products and APIs separate the system prompt, a standing instruction applied to the whole conversation, from your individual messages. OpenAI's guide treats developer instructions as higher priority than user messages. That is why the same prompt can behave differently across the system prompts and models of AI tools: each product wraps your words in instructions you cannot see. Test a prompt in the tool where it will run.
A reusable template for your own prompts
Every example above fits one skeleton. Copy it, fill the slots, and delete any line that does not apply:
Task: [one verb, one object]
Reader: [who uses the output and what they already know]
Why: [what the output is for]
Format: [length, structure, tone rules as concrete limits]
If information is missing: [what to do, e.g. write "unknown"]
<example>
[one sample of a good output]
</example>
<input>
[the material to work on]
</input>Before you reuse any prompt, run it on three real inputs, including one awkward case. A prompt that works once may have been lucky. One that works three times is worth saving, and the guide to building an AI prompt library covers how to store it so your team can find it again.
For a deeper treatment of prompt components and the techniques behind them, AI Prompt Engineering covers foundational prompting through to prompt systems that hold up over time.
The habit to build is small: before you press enter, ask whether a new colleague could do the job from your prompt alone. If they would need to ask a question, put the answer in the prompt.