How to Write AI Prompts: A Practical Business Guide

Almost every business now uses AI tools such as ChatGPT, Claude or Gemini: to draft posts, answer customers, summarise meetings and analyse data. Yet the results vary enormously. For some teams the tool saves hours every week; for others it returns generic text that has to be rewritten from scratch. In most cases the difference is not the tool. It is the request you gave it.

That request is called a prompt. In this guide we cover:

• what a prompt is and why it determines the quality of the answer;
• the building blocks of a good prompt;
• techniques that make results more precise;
• ready-to-use templates for common business tasks;
• what to keep in mind when you work in more than one language;
• how to save good prompts and reuse them, and which mistakes to avoid.

What a prompt is and why it shapes the answer

A prompt is the instruction you send to an AI model to get a result: text, an image, a table, code or an analysis. It can be a single sentence, or a full page with examples and attached files.

The model does not know your business, your customers or your style. It only sees what you wrote, and it fills every gap with the most likely, most generic assumption. That is why "Write a post about our promotion" returns an average post that could fit any company in the world. The same request, with details about the audience, the offer and the tone, returns a draft that only needs polishing.

A good prompt does not have to be long. For a simple task, such as translating a line or suggesting a headline, one sentence is enough. The more complex the task, the more context, output requirements and constraints you should add. A simple rule of thumb: if you gave the same task to a new hire on their first day, what would you need to explain? Write exactly that for the model too.

The six building blocks of a good prompt

There is no single perfect template. What goes into a prompt depends on the task, and you do not need every element every time. But for most business requests, running through six building blocks will noticeably improve the result.

Role

You can give the model a role: content editor, campaign manager, customer service lead, data analyst. The role steers the vocabulary and the point of view of the answer. You can also set a working style, for example: "Don't flatter my ideas. Point out their weaknesses first." This is especially useful when you want honest criticism rather than approval.

Goal

Explain why you need the result. "Copy for the homepage" and "homepage copy that should persuade clinic owners to book a demo" are two different tasks. When the model knows the goal, it can decide for itself what to emphasise and what to leave out.

Context

This is the element that affects quality the most, and the one people skip most often. Describe the business, the product, the target audience, your competitors and what you have already tried. If you ask for a marketing plan, say which market or region you operate in, what the budget is, how many hours a week you can spend and which channels already work.

Style and tone

If the result needs a particular tone, state it up front: formal, friendly, light, professional without jargon. You can also point to a sample: "Write in the style of the attached post." For images, describe the visual style: photography, illustration or 3D, plus colours and composition.

Format and length

Specify what the answer should look like: a table, a list, a checklist, an email, three versions to choose from. Set the length too: a word count, a number of paragraphs, a number of table rows. Without this, the model decides for you, and you will usually end up asking it to shorten or expand.

Constraints

If there are things the model must not do, say so explicitly: do not promise discounts, do not use certain words, do not invent data, do not name competitors. Clear constraints save you an entire round of corrections.

Techniques that improve the result

Beyond the six building blocks, a few working habits improve answers in almost any tool.

Give an example and explain what makes it good

One example is worth a paragraph of explanation. If you have a post that performed well, an email customers loved or a landing page you admire, attach it. Most importantly, say what exactly is good about it. The structure? The tone? The brevity? Without that explanation, the model may copy precisely the part you did not want.

Break a big task into steps

Do not ask for a full article, a complete marketing plan or an entire website in a single request. The result comes out generic and is hard to fix. Work in stages instead: first the outline, then each section separately, and finally an editing pass. That way you can check every stage and correct the direction before moving on.

Ask the model to ask you questions

When you are not sure what to include, add this to the end of your prompt: "Before you answer, ask me up to five questions about anything you are missing." The model will spot the gaps itself, and you get a list of the details worth providing.

Ask for critique, then for an improved version

Once you have a draft, ask the model to critique it: "What are the weaknesses of this text from the point of view of a customer comparing three suppliers?" Then ask for an improved version. This loop takes a minute and makes a noticeable difference.

Continue the conversation instead of starting over

If the answer is almost right, do not rewrite the prompt from scratch. Say what to change: "Make it 30% shorter", "Drop the third paragraph", "Make the tone less formal." The model remembers the context of the current conversation and edits on that basis.

Prompt templates for everyday business tasks

These are templates you can copy and adapt. Replace everything in square brackets with your own details.

Blog article or post

Template: "You write content for a [industry] business in [market or region]. Target audience: [who they are and what they care about]. Goal: [for example, generate enquiries for service X]. Suggest an outline for an article on [topic]: a headline, 6 to 8 subheadings, and two sentences on what each section covers. Do not write the article itself yet."

Once you approve the outline, ask for each section separately. That gives you room to add personal experience and examples from your business at every step, which the model cannot know on its own. We cover a structured process for producing content with AI in our guide to building an AI content pipeline.

Image for a post or banner

Template: "Create an image for an Instagram post about [topic]. Style: [photography / illustration / 3D]. Colours: [brand colours]. Composition: [what is in the centre, what is in the background]. No text in the image. Aspect ratio: 4:5."

It is usually better to add headlines and captions yourself in a design tool: image generators still get text wrong fairly often, especially in languages other than English. We wrote about image prompts separately in our guide to AI photoshoots.

Data analysis

Template: "Attached is a file with [leads / sales / campaign data] for [period]. The columns are: [short description]. Find out which channel brings the most leads, what the cost per lead is in each channel, and which anomalies stand out. Show a table and three conclusions. If the data is not sufficient to answer, say so explicitly and do not guess."

The last line matters: it reduces the chance that the model invents numbers. How to prepare the data and check the output is explained in our guide to AI for data analysis.

Replying to a customer enquiry

Template: "You are a customer service agent for [business]. Our tone: [warm, to the point, no exclamation marks]. The customer wrote: [message]. Our policy on this: [returns / delivery times / pricing]. Write a reply of up to 80 words, using 'we', that offers a clear next step. Do not promise anything that is not in the policy."

Proposal or sales follow-up email

Template: "Help me draft an email to a prospect after an intro call. About the prospect: [industry, size, main pain point]. What we agreed: [points]. What we are offering: [service, timeline]. Write a short email with three paragraphs and a single call to action. Do not include a price; I will add it myself."

Planning your work week

Template: "These are my tasks for the week: [list with time estimates]. I have [number] working hours a day, recurring meetings on [days and times], and I focus best in the morning. Build me a weekly plan as a table, put the demanding tasks in the mornings and leave one hour a day for the unexpected."

Preparing for a meeting or negotiation

Template: "Tomorrow I am meeting a [type of client / supplier] to [goal]. Play the other side: ask me the toughest questions they are likely to ask, one at a time, and give me short feedback after each of my answers." In the app's voice mode you can rehearse this out loud, like a real conversation.

Prompting in several languages: what to know

Many businesses in Europe work in two or three languages at once: English for international clients, plus Greek, Portuguese, German or another local language for the home market. Leading models now write well in all major European languages, but a few points are worth knowing.

Write the request in the language of the result. If you need text in Portuguese, write the prompt in Portuguese, or at least state the output language clearly at the start. When the request is in one language and the output in another, the model tends to "translate in its head", and the result reads like a translation.

Set the form of address. Many languages force a choice that English does not: formal or informal "you", and in some languages grammatical gender as well. Say it up front, for example "address the reader formally" or "use the informal form throughout". Otherwise the model may mix them, especially in longer texts.

Ask for natural, local phrasing. Add a line such as "Write in plain, natural [language], no literal translations from English and no corporate buzzwords." Watch for clichés like "at the end of the day" or "take it to the next level" and their direct equivalents, which instantly sound machine-made. If the audience is in a specific country, name it: European and Brazilian Portuguese, for example, differ noticeably.

Keep terminology consistent. Decide whether you write "CRM" or a local term, "lead" or "enquiry", and ask the model to stay consistent. In mixed-language text, also check punctuation, number formats, dates and currency symbols, which often follow the wrong convention.

Proofread before publishing. Even when the text looks good, read it through once, ideally out loud, or have a native speaker check it. Agreement errors, wrong register and awkward word order are the most common problems.

Save and reuse: custom instructions, projects and templates

If you type the same context every time, you are wasting time. All major tools let you save it once.

Custom instructions

ChatGPT, Claude and Gemini all let you set personal instructions that apply to every conversation: who you are, what your business does, which tone to use and which language to answer in. This is the place for details that do not change.

Projects and custom assistants

For recurring tasks, set up a dedicated workspace: Projects and GPTs in ChatGPT, Projects in Claude, Gems in Gemini. Upload your standing documents, such as a brand guide, an internal price list or customer FAQs, and write instructions. Every conversation inside that workspace starts with that context already loaded.

A team prompt library

A prompt that works well is an asset. Store it in a shared document, in Notion or in whatever tool your team already uses, together with a sample output and a note on when to use it. That way a new team member does not start from zero, and quality does not depend on one person.

From prompt to automation

When the same request goes out every day or every week, such as a lead summary or a first reply to a website enquiry, you can turn it into an automated workflow. Tools like n8n, Make or Zapier pull in the data, send it to the model with your saved prompt and pass the result to your CRM, your inbox or a messaging app. The next step is an agent that carries out a series of actions on its own, as we explain in our guide to AI agents for business.

Privacy and fact-checking

Two things you must never skip, however good the prompt is.

What not to paste into public tools

Do not paste customers' personal data into a chat: ID numbers, phone numbers, payment details, health information. Before you upload a file, remove or mask those columns. If your team handles sensitive data regularly, look at the business and enterprise plans of the tools: many of them let you ensure your data is not used to train the model and let you manage permissions per user. Always read the provider's current terms of use.

If you operate in the European Union or serve customers there, the GDPR applies to any personal data you process, including data you send to an AI provider. In practice that means having a lawful basis, sharing only what is necessary, and having a data processing agreement with the provider. Whether you can hand a customer database to an external service is a question to check with a legal adviser or your data protection officer. This is general information, not legal advice.

Check every fact

A language model writes convincingly even when it is wrong. It can invent a number, a date, a quote or a source that does not exist. So:

• ask up front: "If you are not sure, say that you are not sure";
• ask for sources, and open them yourself before relying on them;
• check numbers and calculations in a spreadsheet, not in the chat;
• verify anything related to law, tax or health with a qualified professional.

Common prompt-writing mistakes

In most cases a weak answer is not the model's fault but the request's. These are the mistakes we see most often.

A request that is too vague

"Write a post." "Give me marketing ideas." A request like that gets an answer that fits everyone, which means it fits no one. Add the audience, the goal and the context.

No format

If you need a table, a list or an email, say so. Otherwise you get a five-paragraph essay and have to ask for it to be restructured.

Several tasks in one request

"Come up with ideas, write the posts, pick the images and build a content calendar" in a single prompt. Every part comes out shallow. Break it into steps.

No examples

When you have an example of what you want and do not attach it, you leave the model guessing. One example with a short explanation beats a paragraph of description.

Accepting the first draft as it is

The first answer is a starting point. Ask for critique, a shorter version, an alternative. Two or three quick rounds make most of the difference.

Using the output without checking it

Text that is published on your website or sent to a customer in your company's name is your responsibility. Check facts, numbers, names and language before it goes out.

How RJL Studio helps businesses put AI to work

A good prompt is the first step. The next one is connecting AI to the processes your business already runs: website forms, your CRM, messaging apps and reports. At RJL Studio we build websites on Webflow and WordPress, connect AI assistants and chatbots to them, and set up automations and integrations with n8n, Make and Zapier: from a first reply to a new enquiry to a weekly lead summary that lands in your inbox on its own.

We start with a short call: we find out which recurring tasks eat up most of your time and suggest what is worth turning into a template, what into an automation and what to leave to a person. Get in touch and we will get back to you.

Frequently asked questions

How long should a good prompt be?

There is no single right length. For a simple task, one sentence is enough. For a complex one, such as writing a landing page or analysing data, a prompt of one or two paragraphs with context, format and constraints will return a better result. The test is not length but whether a new team member would understand exactly what to do from your request.

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

In most cases, yes. The principles of role, goal, context, format and constraints work across all major models. The results themselves can differ in style and length, so when a prompt matters to your business, it is worth trying it in two tools and choosing the one that suits you.

Should I write prompts in English to get better results?

If the output needs to be in another language, it is usually better to write the request in that language, or at least state the output language clearly. Leading models handle major European languages well, and when the request and the result share a language, the text sounds more natural. English can still help for technical tasks, such as code or tools whose interface is in English.

How do I stop the model from making things up?

You cannot prevent it completely, but you can reduce the risk: attach the correct information as a file or text, ask the model to answer only on the basis of what you provided, and tell it to say "I don't know" when information is missing. In any case, check numbers, names and sources before you use the output.

When should I turn a prompt into an automation?

When the same request comes up regularly, the data arrives from a consistent source such as a website form or your CRM, and the output is already good enough without manual edits. At that point you can connect the prompt to a tool like n8n or Make and leave a person to review only the exceptions.