What Is a Prompt? How to Write the Perfect AI Prompt in 2027 (With Examples and Templates)

47 min read

What is a prompt? The answer fits in one line: it's what you write to an artificial intelligence to ask it for something. Understanding that isn't the hard part. The hard part is writing one that gives you something you can actually use without rewriting it from scratch.

You've probably been there. You ask ChatGPT for an email to a client, a LinkedIn post or ideas for a campaign, and it gives you something correct, generic and nothing like your voice. You rewrite it by hand and decide AI just doesn't work for your business.

It's almost never the AI's fault. You talked to it the way you'd type into Google, with five words, when it needed the same thing a new hire on your team would: context, a clear goal and an example of what you expect.

This is the guide I wish I'd had when I started, updated for 2027: what has changed with reasoning models, why context comes before the prompt and, at the end, a library of ready-to-copy templates for marketing, sales, SEO, ads, email, LinkedIn, customer service and data analysis.

Definition

What is a prompt in AI (and what it isn't)

A prompt is the instruction you give an artificial intelligence model, such as ChatGPT, Claude or Gemini, so it generates a response: a piece of text, an image, an analysis or an action. It can be a five-word question or a full brief, with context, examples, attached documents and the format you expect.

What it isn't: a Google search. You give a search engine keywords and it returns pages. You give an AI a brief and it returns finished work. That's why writing to it the way you'd type into Google is the mistake almost everyone starts with.

What is prompting?

Prompting is the practice of writing prompts: drafting, testing and adjusting the instructions you give an AI until you get the result you're after. When it's done methodically, for assistants, chatbots or tasks that repeat, it's called prompt engineering, and you'll find it further down.

What does “prompt” mean, and where does the word come from?

“Prompt” has several meanings in English. As an adjective, it means quick or on time (“a prompt reply”), which is what most dictionary lookups are about. As a verb, to prompt is to cause or encourage someone to do something, and as a noun, a prompt is the cue that gets them going: in theater, the line fed to an actor who has forgotten theirs; in computing, the command line waiting for your input. With generative AI, it has come to mean the instruction you type to the model, which is the sense this guide is about.

How is a prompt different from a Google search?

A search asks for pages; a prompt asks for work. That difference changes the way you write:

	Google search	AI prompt
What you type	A few keywords	A brief in plain language
What you get	A list of pages	An answer or a finished piece of work
What it knows about you	Almost nothing about your specific case	What you tell it, plus whatever it has saved in memory or in a project
If it doesn't help	You change the search	You keep the conversation going and correct it

What are the main types of prompts?

Depending on what you're asking for and where you write it, there are five types of prompts, and each one is written differently:

Type	What it's for	Example
Text	Drafting, summarizing, explaining, brainstorming	“Summarize this contract in five points for someone with no legal background”
Analysis	Drawing conclusions from data, documents or images	“Using this sales spreadsheet, tell me which three products are declining and why that might be”
Image	Generating or editing images	“Product photo of a face cream on wet stone, with soft side lighting”
System instructions	Setting how the AI behaves in every response	The instructions for a project, a Gem or a chatbot
Agent	Handing off a multi-step task that the AI carries out with tools	“Go through this month's reviews, sort them and draft a reply to each one”

Why it matters

Why a good prompt completely changes the answer

Because the AI doesn't know your business, your customer or what you consider a good result: it only knows what you tell it. With a vague prompt, it fills the gaps with whatever is most likely, and the most likely answer is the generic one. With context and a clear goal, it works on your actual case.

Anthropic, the company behind Claude, explains it with an image I use a lot: treat the AI like a brilliant but brand-new employee who doesn't know your rules or the way you work. The more precise you are, the better the result you get.

The practical takeaway: the difference between a mediocre answer and a useful one is almost never the tool; it's what you give it. And most of us write less than half of what's needed.

Anatomy

The perfect prompt structure: the seven parts of a good prompt

A good prompt has seven parts: role, context, task, format, examples, constraints and a quality bar. Google boils the essentials down to four (persona, task, context and format), and OpenAI talks about goal, context, output and limits. I add examples and a quality bar because they're the two parts that improve marketing work the most.

Here's how they apply to a real brief, the copy for the dental implants page of a clinic in Valencia, Spain:

Part	Example	Can you leave it out?
Role	“Act as a copywriter who specializes in healthcare”	Yes, if the context already makes it clear
Context	“We're a clinic in Valencia's Ruzafa neighborhood, and we want to attract patients who are comparing quotes for implants”	Almost never
Task	“Write the copy for our implants page”	Never
Format	“One headline, three 80-word sections and five FAQs”	Only for quick questions
Examples	“This copy of ours worked well: […]”	Yes, but the style will be generic
Constraints	“Don't promise results or include prices”	When there's no risk
Quality bar	“A patient who's scared of the dentist should feel reassured”	Yes, but you'll miss it when you review

What should a good prompt include?

At a minimum, the task, the context and the format. With those three, you're already ahead of most prompts people write. Add the rest when the result calls for it: examples if the style is off, constraints if the AI makes things up, and a quality bar if you're going to ask it to review its own work. OpenAI's ChatGPT guide makes a sensible point: a short prompt often works, and you only need to set limits when they prevent a real problem.

Does “act as an expert in…” still work?

Yes, but not as a magic phrase. Anthropic recommends giving the AI a role because it focuses the tone and behavior, even with a single sentence. What doesn't work is an empty role: “you're the world's leading expert” adds nothing if you don't then explain who that expert is working for and what they need. It's not an official rule, but in my experience, a specific role and good context are worth more than ten adjectives.

Master template to copy

Act as [a specific role, e.g., marketing manager for a dental clinic].

Context:
- My company: [what you do, where and for whom].
- Situation: [what has happened and why you need this now].
- Audience: [who will read it and what they care about].

Task: [verb + specific result].

Format: [table / list / email / text of X words], in a [friendly, professional…] tone.

Example of what I like:
[paste something of yours that worked]

Constraints: don't make up data, prices or testimonials. If you're missing information, ask me before you start.

Quality bar: [how I'll know it's right; e.g., a customer understands it in ten seconds and knows what to do next].

Step by step

How to write a prompt, step by step (for ChatGPT or any AI)

To write a good prompt, in ChatGPT or any other AI, you start with the result you need, add the context, write the task, let the AI ask about whatever it's missing and refine its first draft. Here's how I do it:

• Result
• Context
• Task
• The AI's questions
• Iteration
• Template

1. Start with the result. Decide what you'll do with the answer: send it, publish it or present it. That defines the format.
2. Don't hold back on context. Paste in the customer's email, the product sheet or your notes. Whatever you don't give it, it makes up.
3. Write the task with a clear verb. If there are several tasks, number them.
4. Let it ask you questions. This is the step that saves me the most time.
5. Iterate on the first draft. Say what to cut, what's missing and what tone you'd change, instead of starting over.
6. Save what works. Once a prompt gives you a good result three times, it becomes a template or a project's instructions.

How do you get the AI to ask you questions before it answers?

Just ask it to, and tell it how many questions and what kind, so it doesn't put you through an interrogation. It's the best way to uncover context you didn't know was missing. This is the block I add at the end of any important request:

Ready-to-copy prompt

Before you start, ask me 3 to 5 questions about anything you need to know to do this well: goal, audience, tone, data or limits. Only ask what would change the result.
Once I answer, do the work and, at the end, tell me what assumptions you had to make.

The last sentence is the important one: if the AI tells you what it assumed, you can see in ten seconds where it went wrong.

What does it mean to iterate on a prompt, and when should you start a new chat?

Iterating means improving the result over several messages instead of chasing the perfect prompt on the first try. Google describes prompting as an iterative process in its Gemini guide, and OpenAI stresses that your first message doesn't have to be perfect.

But there's a limit. When a conversation drags on, it fills up with rejected versions, and the AI starts mixing old instructions with new ones. My rule: if it still isn't on track after three or four corrections, I ask it for a summary of what we've decided, open a new chat and paste that summary in.

Before and after

Before and after: turning a bad prompt into a good one

The difference between a bad prompt and a good one isn't length but information: the good one gives the AI what it needs so it doesn't have to guess. A typical case: a renovation company in Madrid wants to revive the quotes nobody has replied to.

Before: the prompt almost everyone writes

Write me an email for customers who haven't replied to a quote.

The AI will return a correct and completely forgettable email: “Dear customer, we wanted to follow up on our quote…” It doesn't know what you sell, to whom or what you want to achieve, so it fills in the blanks with the most generic option.

After: the same request, done right

Act as the sales manager of a full-service home renovation company in Madrid. We mostly work with apartment owners who are renovating to live there, not to rent it out.

Context: we send quotes of between €30,000 and €90,000 and, after two weeks, many customers stop replying. We know they're usually comparing us with other companies or put off by the price.

Task: write a follow-up email to send 15 days after the quote.

Format: a subject line under 50 characters and a body of 90 to 120 words, in a warm, professional, first-name tone. End with a single question that's easy to answer.

Constraints: don't pressure them or invent urgency, don't offer discounts and don't mention competitors.

Quality bar: the customer should feel we're helping them decide, not chasing them.

Give me three versions with different angles: price concerns, timeline concerns and technical concerns.

A good prompt isn't longer: it's the one that doesn't force the AI to guess.

Reasoning models

How to write prompts for reasoning models

With models that reason before they answer, what works is giving them a clear goal and all the context, then letting them decide how to get there; you no longer need to walk them through it step by step. It's the biggest change in the way prompts are written, and most guides still don't explain it. If you're updating your prompts for 2027, start here.

The most advanced models in ChatGPT, Claude and Gemini think before they answer and put in more or less effort depending on how hard the question is. OpenAI sums it up with a very useful comparison: a reasoning model is like a senior coworker you can give a goal and trust to work out the details; a non-reasoning model is more like a junior one who needs explicit instructions.

Does “think step by step” still work?

With reasoning models, you don't need it anymore. OpenAI explicitly advises against asking them to reason step by step, because they already do it internally and it can sometimes make the result worse. Google says something similar about Gemini: you don't need to ask it to write out its reasoning, although for very hard problems you can ask it to think things through. With fast, non-reasoning models, the classic technique (chain-of-thought prompting) still makes sense.

Should you write in all caps or add “VERY IMPORTANT”?

No, and with today's models, it can backfire. Anthropic warns that its models follow instructions more precisely than earlier ones, and that the insistent wording that used to be necessary can now make them overcorrect. What it does recommend is explaining why. Its example: instead of “NEVER use ellipses,” explain that the response will be read aloud by a text-to-speech engine that doesn't know how to pronounce them. With the explanation, the AI applies the rule to cases you hadn't thought of, too.

Techniques

Prompting techniques that work (and when to use each one)

The techniques that actually move the needle are few: giving examples, separating content with delimiters, putting the context first and the question last, breaking big tasks down and asking the AI to review its own work.

Technique	What it is	When I use it
No examples (zero-shot)	Asking for the task directly	Simple tasks, or to see what the AI comes up with
With examples (few-shot)	Including one or more samples of what you expect	When tone, format or style matter
Delimiters	Separating instructions, data and examples with headings or tags	Whenever I paste in documents or data
Context first, question last	Pasting the documents at the top and the instruction at the bottom	With contracts, reports or transcripts
Quote before answering	Asking it to copy the relevant sentences from the document and answer using them	When I can't afford for it to make things up
Breaking the task down	Splitting a big job into stages and reviewing each one (prompt chaining)	Strategies, long documents or when the answer comes out weak: structure first, then the copy
Self-review	A draft, a review against set criteria and a final version	Copy that's going to be published or sent
Meta prompt	Asking the AI to improve or write your prompt	Templates you'll reuse

The techniques come from the official guides by Anthropic, OpenAI and Google. The “When I use it” column is my own judgment.

How many examples should a prompt include?

It depends on the tool, and this is where the official guides disagree. Anthropic recommends three to five varied examples for Claude (what's known as few-shot prompting). Google advises always including examples for Gemini. OpenAI suggests starting without examples for its reasoning models and adding them only if the format isn't coming out right. My rule of thumb: for brand content, always at least one real example of your own; for analysis, start without. And keep them varied, because the AI copies them more than you'd think.

Template for working with long documents

<document>
[paste the contract, report or transcript here]
</document>

<context>
[who you are and what you need this for]
</context>

First, quote word for word the sentences from the document that are relevant to the question. Then answer using only those sentences. If the document doesn't say, write “not in the document.”

Question: [your specific question]

The tags between < and > aren't code: they tell the AI where each thing starts and ends. Anthropic recommends them for Claude, but they work in all three. And the question goes at the end on purpose, as Anthropic and Google advise for long texts.

What is meta prompting, or a prompt improver?

A meta prompt is a prompt for writing prompts (the technique is known as meta prompting): you ask the AI to improve yours or to write it based on what you need. It's the fastest way to turn an improvised request into a template for the team.

Prompt to improve any prompt

I want to improve this prompt before I use it:

<prompt>
[paste your prompt]
</prompt>

1. Tell me what information it's missing to deliver an excellent result, and ask me questions to fill in the gaps.
2. Point out anything ambiguous or contradictory.
3. Once I answer, rewrite it with role, context, task, format, constraints and a quality bar.

Don't add instructions that wouldn't change the result.

How do you stop ChatGPT from making things up and get brutally honest answers?

No prompt guarantees it, but three instructions greatly reduce the risk: allowing it to say “I don't know,” requiring it to separate facts from assumptions and asking it to push back on you. They work just as well in Claude and Gemini, and they also keep the AI from agreeing with everything you say. This is the block I paste in when the answer matters:

Prompt so the AI doesn't make things up and pushes back

Before you answer, follow these rules:
- If you don't know something or aren't sure, say so. I'd rather get an “I don't know” than a made-up answer.
- Separate verifiable facts from your opinions or assumptions, and label them.
- Don't invent figures, sources, quotes or links. If you give me a data point, tell me where it comes from so I can check it.
- Don't agree with me by default: point out the weak spots in my thinking, the way a demanding critic would.

Even so, double-check anything important: these rules reduce errors; they don't eliminate them.

Context

Beyond the prompt: projects, memory, custom instructions and skills

The prompt is only part of what the AI reads. Before your message, ChatGPT, Claude and Gemini can load your custom instructions, what they remember about you, a project's files and saved procedures. Setting this up properly saves you from repeating the same context in every conversation.

Each tool calls these something different. Here's how they line up in practice:

What you want	ChatGPT	Claude	Gemini
For it to always know who you are and how you work	Custom instructions	Profile preferences	Personal context instructions
For it to remember what you discussed in other conversations	Memory	Memory	Personal context and past chats
To work on a topic with its own files and instructions	Projects	Projects	Gems with knowledge files
A procedure the AI applies when it's needed	Project instructions	Skills	Gems

Exact names and availability vary by plan and country. Memory, for example, doesn't work the same way on every account or in every region.

What is a “business brain,” and why does it come before the prompt?

It's a document with everything the AI needs to know about your company that isn't on the internet: what you sell, to whom, how people search for you, what sets you apart, your prices and how you talk. You upload it once to a project or a Gem and, from then on, every prompt works with information only you have. It's the best antidote to generic content: the AI stops suggesting services you don't offer or a tone that isn't yours.

The trick to doing it painlessly: don't write it; say it out loud. Recording the owner answering a set of questions, transcribing the audio and uploading the transcript works better than a polished document, because you capture the nuances nobody puts in writing. A recorded meeting or the voice notes of a talkative client work just as well. Separately, prepare a short document with the facts that never change: name, address, phone number, hours and prices.

Script for recording your business brain

Answer out loud, with real examples, as if you were explaining it to someone who starts at the company tomorrow:

1. What exactly we sell, and what we don't sell even when people ask.
2. Who our best customer is, and who we're not interested in.
3. How that customer looks for us: what they type into Google or ask AI, in their own words.
4. What customers praise most, and what they complain about.
5. Who our competitors are, and where they fall short.
6. What decisions we make that set us apart, even if they seem like details.
7. Prices, timelines and terms that always come up.
8. How we talk: words we use and words we'd never use.

Inside the project, open a new conversation for each task (an article, a landing page, a campaign): they all start from the same brain, and none of them drags along the history of the others.

What is a system prompt?

A system prompt, or system instructions, is the text the AI reads before any conversation, and it sets how the AI should always behave: its role, its tone and what it can and can't do. In chatbots and business apps, it's written by whoever sets them up. In your day-to-day work, the equivalent is your custom instructions, a project's instructions or a Gem's. Treat it like a new employee's onboarding handbook: who the company is, who it sells to, how it talks and what it must never do.

What are Claude Skills, and what are they for?

Claude Skills are reusable bundles of instructions, templates and, when needed, scripts that Claude loads on its own when a task calls for them, without you having to paste them into every conversation. Anthropic includes ready-made skills for Word, Excel, PowerPoint and PDF documents, and lets you create your own. They're the tidiest way to turn “how we do things here” into something the AI applies on its own: a report format, a brand's style or the process for putting together a proposal.

Prompt and context

Prompt engineering vs. context engineering

Prompt engineering is the craft of writing and organizing the instructions you give an AI. Context engineering goes a step further: it decides all the information the AI has in front of it while it works, including documents, memory, tools and conversation history.

Anthropic presents it on its engineering blog as the natural evolution of prompt engineering, with one key idea: context is a limited resource. As it grows, the model's ability to accurately recall what's in it declines. They call this context rot, which is exactly what it sounds like.

	Prompt engineering	Context engineering
The question it answers	How do I ask for it?	What does the AI need to know to do it well?
What you work on	Wording, structure, examples and format	Documents, memory, projects, tools and history
Typical mistake	Vague prompts	Giving it everything “just in case”
Who it matters to	Anyone who uses AI	People who use it every day and, above all, anyone building assistants or agents

In everyday terms: more context isn't always better context. Anthropic recommends finding the smallest set of relevant information that makes the result you want likely. Pasting in ten documents “just in case” usually makes the answer worse, and so does a hundred-message conversation.

The prompt is the question. The context is everything the AI knows when it reads it.

Comparison

Prompts for ChatGPT, Claude and Gemini: what changes in each one

A good prompt works in all three, but prompting in ChatGPT, Claude and Gemini has nuances that their own official guides acknowledge. This is what each company recommends for its current models, with none of my opinions mixed in:

	ChatGPT (OpenAI)	Claude (Anthropic)	Gemini (Google)
Core idea	With reasoning models, give the goal and let it work out the details	Be clear and direct, and explain the reason behind each instruction	Clear, direct and precise instructions
Examples	Start without examples and add them if the format isn't right	Three to five varied examples	Always include examples
Structure	Markdown sections and tags to separate content	XML tags to separate instructions, context and data	Consistent delimiters: XML tags or Markdown headings
Long documents	Mark them off with tags so they don't get mixed up with the instructions	Documents at the top, question at the end	All the context first, question at the end
Reasoning	Don't ask it to reason step by step: it already does	Decides for itself how much to think based on how hard the task is	Reasons internally; for very hard problems, ask it to think things through
A tip of its own	Say how you'll use the result so it comes back ready to use	Say what to do instead of what not to do	Direct by default: if you want more detail or a conversational tone, ask for it

Summary of each company's official prompting guides (links under “Sources”). Models change every few months; the principles change far less.

Does the same prompt work in ChatGPT, Claude and Gemini?

Yes, for most everyday tasks. The differences show up in long, repetitive tasks: if you're going to use a template hundreds of times, test it in the tool you actually work in and adjust it to its quirks.

Templates

Prompt library: AI prompt examples and templates to copy

These AI prompt examples are free templates for a business's day-to-day work, organized by area. They all follow the same anatomy: change whatever's in [brackets], paste in your data and, if you'll use one often, save it in a project or a Gem.

Two rules for all of them: give it real data (reviews, emails, your business's numbers) and don't publish anything you haven't read. AI speeds up the work; the responsibility is still yours.

Marketing prompts

Ideal customer based on what your customers say

<material>
[paste Google reviews, customer emails or notes from sales calls]
</material>

Act as a marketing strategist for small businesses. My business: [what you sell, where and to whom].

Using the material above:
1. Group what customers say into reasons to buy, fears and objections, with verbatim example quotes.
2. Describe the two or three types of customer that come up, what each one cares about and the exact words they use to ask for what we sell (we'll use them for ads, FAQs and keywords).
3. Suggest five marketing messages for the most profitable type of customer, in their own words.

Only use what's in the material. If something is your own assumption, label it “hypothesis.”

Sales prompts

Prepare for a sales meeting

Tomorrow I'm meeting with [company or customer profile] to offer them [service].
What I know about them: [industry, size, what they've said so far, website].

Prepare:
1. The five questions I should ask to understand their problem before we talk about price.
2. The three most likely objections and how I'd answer each one without pressuring them.
3. A three-line summary of why our service fits their situation.

Tone: a conversation between professionals, with no lines from a sales manual.

Questions from three types of customer

I want to sell [product or service] to [audience].
Put yourself in the shoes of three different customers: an expert who compares every detail, a beginner who doesn't understand the jargon and a demanding one who has already had a bad experience.
For each one, write the five questions they'd ask me before buying, in their own words.
Then group all the questions by topic and tell me which ones I should answer on the website, which ones in the proposal and which ones on the call.

Add the real questions that show up in Google's “People also ask” box, and you'll have the foundation for your FAQs and your sales pitch.

SEO prompts

AI is very good at preparing SEO work, as long as you don't ask it for data it doesn't have: it's useful for brainstorming and for cleaning up and grouping lists, but search volumes have to come from a tool like Google's Keyword Planner.

Article brief based on the competition

<competition>
[paste the H2 and H3 headings from the top five Google results for your search]
</competition>

<questions>
[paste the questions from “People also ask”]
</questions>

Act as an SEO consultant. I want to write about “[keyword]” for [type of business], and I want the article to lead to [service].
1. Tell me what every page covers (the must-haves) and what none of them covers (our opportunity).
2. Propose the article structure with H2s and H3s, noting which question each one answers.
3. Suggest where an example, a table or a template of our own would fit.

Don't make up search volumes or data.

Categories for your Google Business Profile

Act as a local SEO expert. My business: [exactly what you do, for whom and in which city; e.g., sports physical therapy in Seville].

1. Suggest the most specific primary category that exists in Google Business Profile for this business, with its exact name as it appears in Google.
2. Suggest up to five secondary categories that match services we actually offer.
3. Explain in one line why you chose each one.

Don't make up categories: if you're not sure one exists, say so.

Afterward, check in your profile that the categories exist exactly as written: AI can suggest names Google doesn't have.

Google Ads prompts

In Google Ads, AI saves hours writing ad variations. What it doesn't do is decide the account structure or the bidding strategy, which is where money is made or lost; that's what a Google Ads agency in Andorra like ours is for.

Responsive search ad

Act as a Google Ads specialist for local businesses.
Business: [what you sell] in [area]. Real advantage: [what sets you apart, with data].
Search the ad responds to: “[keyword]”.
Landing page: [URL or summary].

Write a responsive search ad:
- 15 headlines of up to 30 characters: 5 with the keyword, 5 with benefits and 5 with proof or a call to action.
- 4 descriptions of up to 90 characters.

Don't use superlatives we can't back up, and no exclamation points in the headlines.

A practical detail: AI is bad at counting characters. Run them through a character counter before you upload them.

Meta Ads prompts

On Facebook and Instagram, your ad competes with content from friends and creators, so what performs best is testing different angles. AI is perfect for generating them; choosing which ones to test and reading the results is the job of a Meta Ads agency in Andorra or whoever runs your campaigns.

Creative angles and copy for Meta Ads

Act as a social media advertising creative.
Product: [what you sell]. Customer: [who they are and what they care about]. Offer: [what you're offering and until when].

1. Suggest six different creative angles: problem, desire, social proof, comparison, objection and curiosity.
2. For each one, write primary text that makes sense in the first two lines, a headline under 40 characters and a visual idea for the image or video.
3. Rank them from the one you'd test first to the last one, and explain why.

Talk the way the customer talks, not like a brochure.

Email prompts

Welcome sequence for your newsletter

Write a three-email welcome sequence for people who sign up for [business]'s newsletter.
Email 1 (immediately): deliver what was promised and explain what they'll get and how often.
Email 2 (after 3 days): tell this real story about a customer or the business: [paste the story].
Email 3 (after 7 days): introduce [product or service] with a single call to action.

For each email: three alternative subject lines under 45 characters and a body of 120 to 180 words.
Tone: [how the brand talks]. An email of ours that worked: [paste the email].

LinkedIn prompts

LinkedIn post based on a real experience

I want to post on LinkedIn about something that happened to me: [tell it the way you'd tell a friend, with details].
I'm [your job title] at [company], and my readers are mostly [type of audience].

Write two versions of the post with:
- a first line that stands on its own and sparks curiosity without exaggerating,
- paragraphs of one or two sentences and one concrete takeaway,
- a closing open-ended question,
- 120 to 200 words, no filler hashtags and no emoji on every line.

This is how I write (copy the tone, not the content): [paste two of your posts].

Customer service prompts

Reply to a negative review

A customer left this [number]-star review on Google: “[copy the review]”.
What happened, from our side: [your version, with facts].

Write a public reply of under 100 words that:
- thanks them for the review and names the specific problem,
- explains, without excuses, what we're going to do or have already done,
- offers to continue the conversation privately at [phone or email].

A future customer should read it and think, “This company handles it well when something goes wrong.” Don't argue, and don't share any of the customer's personal details.

Instructions for a customer service assistant

You're the customer service assistant for [company], which [what it does] in [area].

Your goal: answer questions about [products, hours, shipping, appointments] using the information in the attached documents.

How you respond:
- In whatever language the customer writes in (Spanish, Catalan, French or English).
- With short, friendly, specific answers. If there are steps, number them.
- If the answer isn't in the documents, say so and refer them to [email or phone]. Don't assume prices, timelines or terms.
- If the customer is upset, acknowledge the problem before offering solutions.

Never promise refunds or exceptions, never ask for payment details and never share information about other customers.

That last one isn't a prompt for a single day: it's a set of system instructions. Paste them into a project or a Gem along with your FAQs, terms and prices, and test them with the twenty toughest questions you've ever been asked before you trust them.

Data analysis prompts

Analyze a sales or campaign spreadsheet

I'm attaching [file] with [what data: sales by product and month / campaign results].
Context: [what the business is and what decision I want to make with this analysis].

1. Before analyzing anything, describe in three lines which columns you see and whether there's any odd, missing or duplicate data.
2. Answer this question: [a specific question; e.g., which products have declined and since when].
3. Give three possible explanations, from most to least likely, and which data point would confirm each one.
4. Finish with a summary table and a recommendation.

Do the calculations using the file, not from memory, and separate what the data says from your hypotheses.

Prompt to humanize AI text and make it sound like your brand

“Prompt to humanize AI text” and “make ChatGPT write like a human” get searched a lot, usually with the aim of fooling AI detectors. I come at it from the opposite direction: make the text sound like your brand, because that's what your customer notices.

Rewrite in your brand voice

<our_voice>
[paste three pieces of your copy that represent you well]
</our_voice>

<text>
[paste the text you want to improve]
</text>

Rewrite the text in the tone of the examples: same vocabulary, same sentence length, same level of warmth.
Remove clichés, empty adjectives and repetitive structures (“not only… but also…”, lists of three in every paragraph).
Keep all the facts and the meaning. If something can't be said in our voice, flag it instead of changing it.

If you'll use it often, turn it into a Gem or a tone-of-voice project, with the brand's examples and red lines, and run each text through it section by section. It's far more reliable than asking it to “make it sound more human.”

Prompt to translate and adapt content into another language

In Andorra, businesses work in Catalan, Spanish, French and English every day; in much of Spain, they sell in Spanish and in Catalan, Galician or Basque; and in the US, plenty of businesses sell in both English and Spanish. Literal translation is the fastest way to make a text sound foreign. This prompt asks for adaptation, not translation:

Translate and adapt

Translate and adapt this text into [Catalan / French / Spanish] for [audience: customers in Andorra, French tourists…].
Don't translate word for word: adapt expressions, examples and calls to action so it reads as if a native speaker wrote it.
Keep brand names, prices and data exactly the same.
At the end, list the expressions you adapted and why, so I can review them.

[paste the text]

Prompts to find out what AI says about your business

It's the flip side of everything above: asking the AI the same thing a customer would. Make a list of the twenty most important questions about your business, run them through ChatGPT, Gemini, Claude and Google's AI Mode, and note whether you're mentioned, whether you're cited as a source and what's said about you.

What AI knows about your company

Without searching the web, using only what you already know: what can you tell me about [exact name of your company] in [city]?
Tell me what it does, what sets it apart and what its customers think of it. If you don't know it or aren't sure, say so clearly instead of guessing.

If your tool lets you turn off web search, do it for this test: that way, you see what the AI has learned about your brand, not what it finds in the moment. If it knows nothing or gets it wrong, that's where GEO work comes in.

Images

Image prompts for realistic photos: how to ask for them properly

A good image prompt describes the scene in full sentences, not as a list of loose keywords: what's in it, how it's framed, what's happening, where it is and in what style. Google lists these elements in its tips for generating and editing images with Gemini:

Element	What to write
Subject	Who or what appears, in detail: material, color, age, clothing
Composition	The framing: close-up, wide shot, overhead view, portrait
Action	What's happening or what the subject is doing
Location	Where the scene takes place and what the background looks like
Style	Realistic photography, watercolor, flat illustration, 3D render…
Editing instructions	If you're modifying an image, exactly what to change, directly and specifically

Template for a realistic product photo

Generate an image: a realistic product photo of [product, with material, color and size] on [surface], with [background].
Framing: [overhead view / close-up at product level / three-quarter view].
Lighting: [soft natural light from a side window / studio lighting with strong shadows].
Mood: [what it should convey: understated luxury, freshness, craftsmanship].
Key detail: [what must be sharp; e.g., the texture or the label].
Landscape format, 16:9, no added text.

Two tips from Google's guides that make a difference. Describe what you want to see in positive terms: instead of “no cars,” write “an empty street.” Some tools have a separate field for what you don't want, known as a negative prompt, but Google recommends describing the scene positively. And when you're editing, change one thing per message, within the same conversation, so the image keeps the same people, lighting and style.

Template for editing an image

Using the image I'm attaching, change only [specific element: the background / the color of the jacket / the lighting].
I want [description of the change].
Keep everything else exactly the same: framing, people, products, colors and style.

What prompt should you use to improve image quality?

One that asks for a specific improvement and forbids changing the content. Attach the photo in Gemini or ChatGPT and describe what you want to improve: sharpness, noise, resolution or scratches on an old photo. Keep in mind that the AI rebuilds details, so check that it hasn't changed faces, text or objects.

Prompt to improve photo quality

Improve the quality of this image: sharper, less noise and higher resolution, with natural colors and balanced lighting.
Don't change the framing, faces, text, objects or their position.
If it's an old photo, repair scratches and stains without adding new elements.

What is Nano Banana, and how do you write a prompt for it?

Nano Banana is the name Google gives to Gemini's image generation and editing models, which have already gone through several versions. You talk to it the way we've covered: full sentences, a well-described scene and one edit at a time. Its strength is editing images you already have and keeping characters or objects consistent from one image to the next, so it works very well for product variations or pieces of the same campaign.

How do you create a professional headshot with Gemini or ChatGPT?

With a well-lit photo of yourself and a prompt that asks it to keep your features and change only the style of the photo. It works reasonably well if you're specific:

Professional headshot from your photo

Using the photo I'm attaching, create a professional headshot for LinkedIn and the company website.
Keep my face, features, hair and age exactly as they are: don't make me look younger or change the shape of my face.
Clothing: [navy blazer and white shirt / plain sweater in a neutral color].
Background: [plain gray studio backdrop / bright, blurred office].
Soft studio lighting, framed from the chest up, looking at the camera with a friendly expression.
It should look like a real photograph, not an illustration.

Mistakes

Common prompt-writing mistakes (and how to avoid them)

The most common prompt-writing mistakes are asking without context, settling for the first answer, giving it data it shouldn't have and trusting figures without checking them. Almost all of them can be fixed with one more sentence or a good habit:

Mistake	What happens	How I fix it
Writing like it's Google	Generic answers	Saying who you are, who it's for and what you want it for
Asking for everything in one message	A mediocre result in every part	Breaking it into steps and reviewing each one
Settling for the first version	You end up with a draft, not finished work	Asking for specific changes: what to cut, what's missing, what tone
Only saying what you don't want	The AI fixates on exactly what you ruled out	Saying what you want instead
Trusting data and sources without checking	Mistakes published under your name	Asking it to separate facts from assumptions, and verifying anything important
Giving it a URL and assuming it read it	Sometimes it can't open it and answers with guesses	Pasting in the text or giving it screenshots
Never-ending conversations	The AI mixes up old and new instructions	A summary of what's been decided, and a new conversation
Repeating the context every day	You waste time and forget things	Custom instructions, projects or Gems

AI doesn't make fewer mistakes because you ask more forcefully. It makes fewer mistakes when you give it what it needs so it doesn't have to guess.

Summary

How to write the perfect prompt: 8 key rules

The perfect prompt isn't written on the first try: you build it with a good brief, just the right context and two or three rounds of conversation. These are the key rules:

Example

Full example: an AI marketing assistant for a ski shop in Andorra

Here's how I'd set up AI for the marketing of a ski rental and equipment shop in Andorra that's getting ready for the season. It works for any business: change the industry and the framework stays the same.

The first step isn't a prompt but the business brain: a project in ChatGPT or Claude, or a Gem in Gemini, with these instructions and four files: catalog, rates, customer reviews and copy that worked in previous years.

Project instructions

You're the marketing manager for [name], a ski rental and equipment shop in [parish], Andorra.

Our customers: families who come to ski for the weekend, and local skiers who buy gear. They choose us for our boot fitting and the way we treat them, not for price.

How we write: warm, clear, with a light sense of humor and no jargon. Examples of our voice are in the “copy-that-worked” file.

Always: use the prices and terms in the “rates” file; if a piece of information isn't there, ask.
Never: promise snow, discounts that aren't in the rates file or comparisons with other shops.

Languages: Spanish by default; Catalan and French when I ask, adapted rather than translated.

With the project set up, the day's request can be short, because the context is already inside:

First request inside the project

We want to fill up season rentals for families before the slopes open.
Put together a four-week plan for Instagram, Google and email: what to post, what to advertise and what to send each week, with the goal of each action.
Before you start, ask me any questions you need.

Stage	What happens	Why it works
The AI's questions	It asks about the ad budget, the opening date and whether there's an early-booking offer	It uncovers the context missing from the request
First version	A four-week plan in a table, with pieces for each channel	It arrives in the requested format, with no rework
Iteration	“Week 2 is too salesy; replace it with content on how to choose boots”	A specific correction on a solid base
Pieces	The copy for week 1, using this guide's Meta Ads, Google Ads and email templates	Each piece inherits the project's context
Review and languages	“Review everything with this criterion: would a family that has never skied understand it?” Then Catalan and French versions	The AI reviews against a criterion, not by eye

Each day's prompt is short because the context already lives in the project. That's the difference between using AI and working with it: the first time, you spend half an hour explaining the business, and from then on, every request takes minutes.

What I wouldn't do

• Copy “the 100 best prompts” without adapting them to my business.
• Pay for “secret” prompt packs. What works is in the official guides, and they're free.
• Ask the AI for figures, studies or sources and publish them without checking.
• Fill prompts with all caps, warnings and “VERY IMPORTANT.”
• Paste in customers' personal data or confidential information.
• Use AI to make a text fool a detector instead of making it sound like my brand.
• Leave the final decision to the AI. It prepares, suggests and speeds things up; I decide.

The bottom line

Writing good prompts isn't about tricks. It's about the same things that make any brief to a person good: knowing what you want, explaining the context, showing an example and reviewing the result.

• Goal
• Context
• Example
• Conversation
• A result you can use

And there's a second takeaway that, as an agency, I think matters even more. If you ask AI, so do your customers. Many buying decisions now start with a prompt like “which renovation company do you recommend in Barcelona?” You already know how people ask AI. The next step is making sure that when they do, it talks about your business.

Frequently asked questions