Imagine you run a restaurant in Palamós and ask ChatGPT: "Reply to this one-star review". You get back a polite paragraph, complete with "we're very sorry about your experience" and "we hope to see you again soon". It's well written, and any restaurant in the world could sign it. Now try telling it what actually happened that Saturday, how you talk to your customers, what you want to offer and how long the reply should be. The answer changes completely. Knowing how to write a good prompt isn't technical at all: it's learning to hand a job to someone who doesn't know you.
What a prompt is and why it usually falls short
A prompt is what you type to an AI tool to ask it for something: the question, the task, the instructions. And it usually falls short for the same reason: we ask the way we'd ask someone who has worked with us for years.
The AI hasn't spent a single day in your business. Anthropic's prompting best practices, from the company that makes Claude, sum it up with an image that works for any tool: treat it like a brilliant but new employee who doesn't know your norms or how you work. The more precisely you explain, the better the result. They also suggest a golden rule: show your prompt to a colleague who knows nothing about the task. If they can't follow it, the AI won't either.
And this isn't just for a handful of enthusiasts. According to the INE's 2025 household ICT survey, 37.9% of people aged 16 to 74 in Spain had used generative AI tools in the previous three months, and 17.9% had used them for work.
The five parts of a good prompt
You don't need to follow a formula to the letter, but requests that work first time usually include these five things. OpenAI's guide to writing prompts suggests a very similar split: who the assistant is, what instructions it follows, what examples it has and what context you give it.
1. The task, in one sentence
What exactly you want. "Write the public reply to this Google review" is a task. "Help me with the reviews" is a conversation with no clear end.
2. Context: who you are and who it's for
What your business is, who will read the result and why. Anthropic's guide stresses explaining the reason behind each instruction: "avoid jargon, because older customers will read this" works better than a bare "avoid jargon", because the AI understands what you're after and applies it to things you didn't mention too.
3. Your information
The full review, your menu, your prices, your returns policy. The AI knows nothing about your business that you don't paste into the prompt, and it fills in what it doesn't know with great confidence. OpenAI's guide calls this context: the information the model doesn't have, such as your own data.
4. The format
How long it should be, in what tone, in which language and in what shape: "80 words maximum, informal, in Catalan, no lists". If you don't say, the AI decides for you, and it tends to overdo it.
5. The limits
What it mustn't do. "Don't offer discounts", "don't name the waiter", "if you're missing a detail, don't make it up, ask me". One line of limits saves you a lot of corrections.
Want to know what it would cost you?
We reply the same day. No commitment, no sales pitch.
Before and after: the same request, asked properly
The one-line prompt
This is how almost everyone starts:
Reply to this review: "We waited 40 minutes for the paella and it arrived cold. We won't be back."
The same request as a good prompt
And this is how it looks with the five parts:
I run a rice restaurant in Palamós and want to reply on Google to a one-star review.
What happened: on Saturday the 12th we had a first communion party of 40 and the rice dishes came out late. Theirs arrived lukewarm and they didn't tell us at the table.
Write the public reply: 80 words maximum, friendly, in English and without stock customer-service phrases.
Acknowledge the delay without excuses, explain what happened in one sentence and invite them back on a weekday.
Don't offer discounts or name anyone on the team. If you're missing any details, ask me before you write.
The second takes a minute longer to write. What comes back sounds much more like what you would have written, and the first draft rarely needs more than a small tweak.
Give it examples and save yourself the explanations
Explaining a tone is hard; showing it is easy. If you have review replies, emails or posts you like, paste them into the prompt as samples of how you write. Anthropic's guide calls examples one of the most reliable ways to steer format, tone and structure, and recommends three to five varied ones.
Make it clear which part of the prompt is examples and which part is the task. Otherwise it may copy the example almost word for word instead of using it as a sample.
Ask it to ask you questions before writing
When the task is long, one line at the end makes a big difference: "Before you start, ask me any questions you need". The AI will tell you what's missing instead of filling it in itself.
If the job has several steps, put them in order and number them. Anthropic recommends writing instructions as numbered steps when the order matters or when no step should be skipped. And if there are three different tasks, three prompts beat one that asks for everything at once.
Always check what you get back
A good prompt reduces mistakes but doesn't remove them. AI can invent an opening time, a price or a law with the same confidence it gets everything else right. Before you publish or send anything, check names, figures, dates and links. Whatever you publish carries your name. And Google doesn't penalise text for being made with AI, but it has no reason to show something nobody has checked either, as we explain in AI-generated content.
What you should never paste into a prompt
Whatever you type into an AI tool leaves your computer. Before pasting anything in, ask yourself whether you'd email it to a stranger.
- Your customers' personal data: full names, ID numbers, phone numbers, addresses or health information. If you need to work with a real case, change the names and remove anything that identifies the person.
- Passwords and logins, never.
- Contracts, payslips or confidential documents you couldn't share with a third party.
It also depends on the account. According to OpenAI's help centre, on personal ChatGPT accounts such as Free, Plus or Pro, what you type may be used to train its models unless you switch it off in the data controls in settings or use a temporary chat. On ChatGPT Business, Enterprise and the API, it isn't used by default. If your team works with customer information, that setting is no small detail.
And if AI is already part of the daily routine in your business, it's worth writing down on one page which tools are used and what information never leaves the company. In artificial intelligence in your company we explain which rules to set.
Save the prompts that work
A good prompt is a procedure, and procedures get saved. If every week you reply to reviews, prepare the newsletter or summarise orders, write the prompt once with all the context, test it until you're happy with the result and save it where the team can find it. Each time, only the day's details change.
That's how we handle tasks that repeat: the fixed instructions are written down and checked, so the result doesn't depend on who asks or what kind of day they're having. If you're not sure which tasks in your business deserve a saved prompt, which tasks to automate with AI has a four-question filter.
Five mistakes when writing a prompt
- Asking for everything in one line and expecting the AI to guess the rest.
- Not giving it your information and trusting the prices, opening times or names it makes up.
- Mixing several tasks in the same prompt.
- Not saying what format you want and getting four paragraphs where two lines would do.
- Settling for the first result. If it doesn't convince you, tell it what to change: "shorter", "less formal", "drop the last sentence". The second round is usually the good one.
What it costs for us to train your team in AI
AI training is part of our AI integration and training service, and it's built around real cases from your business: the emails, reviews, quotes and reports you produce every week. As explained on our pricing page, it's priced per project, not as a monthly fee: the price depends on what needs preparing, and we confirm it in writing before we start.
If you don't know yet which parts of your business AI would help with, the first step is an AI audit: we look at how you work and tell you where it fits and where it doesn't, with priorities and costs.
We'll tell you in a ten-minute conversation
We reply the same day. No commitment, no sales pitch.
Learn to write a good prompt with your own cases
We're three partners working from the Palau Firal in Sant Antoni de Calonge, with businesses across Girona and Barcelona. AI integrations with your software are built by Ivan, who does the programming, and the training uses the tasks you repeat every week, so the team comes away knowing how to ask for their own work rather than textbook examples.
If you'd like to see it with a real case, message us on WhatsApp with a task you repeat every week and we'll look at it in a first conversation with no obligation. You can also write to us through the contact form.
Tell us what you need and we'll give you a price
We reply the same day. No commitment, no sales pitch.