Prompt writing: how to write better AI prompts (with examples)

A prompt is the instruction you give an AI assistant, and the quality of that instruction largely decides what you get back. Ask a vague question and you get a generic answer. Write a good briefing and you get usable work in your own voice. Prompt writing is a habit you can learn, with a simple formula. In this guide we explain what a prompt is, how to write a good prompt in 4 steps, and we share copy-paste example prompts for the marketing tasks of an SME. We also look at whether you should prompt in Dutch or in English, which mistakes come up most often and what changes now that AI assistants can remember your context. Everything comes from the work we do with AI every day, for ourselves and for businesses in Flanders.

What is a prompt?

A prompt is the instruction you give an AI model such as ChatGPT, Claude or Gemini: your question, your instructions and all the context you send along, such as files or examples. The model bases its answer on that prompt, supplemented by whatever sits in your settings, project instructions or memory. Vague input therefore produces vague output.

Google puts it briefly in its help pages: a prompt tells Gemini what you want it to generate. In ChatGPT the prompt is the message you type into the input field; in an image generator it is the description of the image you want. AI researcher Andrej Karpathy summed up in a single sentence, in January 2023, how plain language is becoming the way to steer software: “The hottest new programming language is English”.

It helps to think of an AI assistant as a fast, broadly trained new colleague who does not know your business yet. Anthropic, the maker of Claude, uses the same image in its official documentation:

“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.”

— Anthropic, maker of Claude, in the documentation “Prompting best practices”

You would not give a new colleague like that a five-word briefing either. Goal, audience, style, examples and format belong in your prompt. Which assistant you pick matters little for this skill: we discuss the differences between the big models in ChatGPT vs Claude vs Perplexity and in our practical guide to choosing per task. If you want to know what prompting involves as a skill and where prompt engineering begins, read what is prompting?

How do you write a good prompt? The formula in 4 steps

A good prompt answers 4 questions: what role does the AI play, what is the goal, what context does it need and what should the result look like? That is the formula Role + Goal + Context + Output. Context carries the most weight and the role is optional. Google uses the same 4 building blocks for Gemini and adds that you do not need all 4 of them.

ClickForest teaches this formula to teams of SMEs in Flanders during hands-on AI training, because it works in every assistant. This is what each step looks like:

1. Role (optional): “You are an experienced copywriter for Flemish webshops.” A role mainly steers tone, level and perspective. According to Anthropic, even a single sentence describing the role makes a difference to the model’s behaviour and tone.
2. Goal: “Write a product description that convinces doubters.” One concrete goal per prompt. Ask for three things at once and you get three half answers.
3. Context: “The product is X, the audience is Y, our tone of voice is Z, and this is an example we like.” Also say why something matters: according to Anthropic, the motivation behind an instruction helps the model give more targeted answers. This is the step that is missing most often.
4. Output: “Give me 150 words, in Dutch, with a heading and 3 bullets.” Make format, length, language and structure explicit, and say when the result counts as good.

That last addition comes from the guide OpenAI published in October 2026 for its newest models: “Start with a clear assignment: the result you want, who it’s for, the relevant context and constraints, and what counts as done.” It is the same formula, with an explicit bar added.

Wharton professor Ethan Mollick is strikingly sober about the role. He recommends trying roles out and leaving them out when they add nothing:

”(…) roles help the AI understand your context, they aren’t magical (they do not actually turn the AI into a marketer) and sometimes giving the AI a role can actually lower accuracy.”

— Ethan Mollick, professor at the Wharton School, in “Getting started with AI: Good enough prompting” (November 2024)

So use a role for tone and perspective, and check now and then whether the answer is just as good without one. More elaborate frameworks with the same building blocks exist. If you like working with a fixed template, our article on the RACE model for ChatGPT prompts offers a worked-out variant with templates. Pick one framework and use it consistently.

Which example prompts can you use right away?

Below are 6 copy-paste example prompts for common marketing tasks of an SME, each built along the formula. Replace the text in square brackets with your own details. Wherever you can, add an example of your own of what you consider good: in our practice that is the fastest way to get more usable output.

That last prompt deliberately puts the document at the top and the question at the bottom. Anthropic recommends this for long inputs and reports that a question at the end can improve response quality by up to 30% in its own tests, especially with complex input spanning several documents. Anthropic is just as clear about examples: “Examples are one of the most reliable ways to steer Claude’s output format, tone, and structure.”

Keep your best-performing prompts in one place for the whole team, so nobody has to start from scratch every time. That is also the bridge to structural AI use, covered in more detail in our AI marketing guide. Which tools are strongest per marketing task is compared in the best AI tools for marketers.

Should you prompt in your own language or in English?

For Dutch marketing copy you are best off prompting in Dutch, and you state the desired output language explicitly. Research shows that tasks where culture and tone matter benefit from a prompt in the native language. For knowledge and reasoning tasks, studies often find an advantage for English instructions, especially with older models. We found no research specifically on Dutch.

What exactly do those studies say? A study by Liu and colleagues at the 2025 NAACL conference concludes: “For culture-related tasks that need deep language understanding, prompting in the native language proves more effective as it better captures the nuances of culture and language.” A study in JMIR Medical Education from March 2026 tested German, French and Italian on medical multiple-choice questions. English prompts generally scored better there, while the strongest models achieved comparable results with prompts in their own language. Those models have since been succeeded, so read it as a direction.

Anthropic adds a practical point: even though the model detects the target language by itself, explicitly stating the input and output language makes the result more reliable. For Dutch output, also specify the region if that matters for your brand, for example “Belgian Dutch, with Flemish word choice”. Anthropic recommends taking cultural and regional context into account alongside the language, and in our experience, without that instruction, Netherlands-Dutch words regularly slip into the text.

Unsure for your task? Have the model make two versions, one from a Dutch instruction and one from an English instruction, and compare. It takes a minute.

Want your whole team to learn to prompt better? ClickForest gives hands-on AI training to SMEs in Flanders, from prompt basics to fixed working routines per department. Take a look at the AI training or get in touch.

What are the most common prompt mistakes?

The most common prompt mistakes are too little context, several tasks in one prompt, not asking for a format, only saying what is not allowed, expecting the AI to fill in what you did not ask, treating the first answer as the final product, not checking and sharing sensitive data. The formula plus two habits, iterating and checking, catches most of them.

This is how you recognise them, and how you fix them:

• Too little context. The AI does not know your business, your customer or your house style, unless you provide them or have set them up in a project.
• Several tasks in one prompt. Split them up. One goal per prompt gives sharper answers.
• Not asking for a format. Without length, structure and language the model chooses for itself, and that is rarely what you were looking for.
• Only saying what is not allowed. Anthropic recommends telling the model what it should do instead: replace “do not use bullet points” with “write in flowing paragraphs”.
• Expecting the AI to fill in the obvious. Recent models follow instructions more literally. Anthropic writes: “If you want Claude to do more than the task strictly requires, explicitly request it rather than relying on the model to infer this from vague prompts.” If an instruction should apply to all five of your products, say so.
• Treating the first answer as the final product. Give the core right away and refine afterwards: “shorter”, “less formal”, “give three alternatives for the opening line”. Or as Ethan Mollick puts it: “Working with AI is a dialogue, not an order.”
• Not checking. Verify facts, figures and claims in AI output before you publish, because a fluently worded answer can be wrong. Whether an AI text still reads as AI is something we tested in AI copywriting versus the human writer.
• Sharing sensitive data. Do not put customer data or confidential figures into a tool whose settings you do not know, and make team agreements about it.

The best test for any prompt again comes from Anthropic’s documentation:

“Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they’d be confused, Claude will be too.”

— Anthropic, maker of Claude, in the documentation “Prompting best practices”

What changes now that AI assistants remember your context?

AI assistants such as ChatGPT and Claude can now keep context through memory, personal instructions and projects. What you used to repeat in every prompt, such as your business, your audience and your tone of voice, is better set once in the instructions of a project. Your prompt gets shorter as a result. The formula stays the same.

At the end of 2024 Ethan Mollick still advised treating AI like “an infinitely patient new coworker who forgets everything you tell them each new conversation”. That is no longer entirely true. Since August 2026 Claude’s memory works both in chat and in Cowork in the cloud, and it is on by default on the Free, Pro and Max plans. For Team and Enterprise organisations it is off by default. What Cowork is exactly, you can read in Claude Cowork explained.

In ChatGPT you can set per project that memory only uses conversations from that project. Since July 2026, Plus, Pro, Business, Enterprise and Education users can save up to 5,000 characters of personal instructions, up from 1,500 before. That is enough for a short company description, your audience and your house style rules.

In practice: create a project for each recurring type of work, for example “webshop copy” or “newsletter”, and put your fixed context there. Your prompt then only describes today’s task. Because memory is on by default on some plans, it also belongs in your team agreements about which data may go into an AI tool.

How do you use prompts for recurring tasks?

Turn your best prompts into templates: a fixed structure per recurring task, with fill-in fields for whatever changes each time. That way your whole team gets the same quality. You store those templates in a shared document, or in the place your AI assistant offers for it: skills, plugins or project instructions, depending on the platform.

Those built-in places are changing fast. Google wrote on 30 September 2026 that skills will soon replace Gems for giving Gemini instructions for specific tasks, and calls them “saved shortcuts to faster prompting”. OpenAI announced on 11 September 2026 that it plans to retire custom GPTs in ChatGPT, with a migration path to plugins that combine reusable instructions with connected apps. Claude already works with skills; how they work and how you share them, you can read in Claude Skills explained. A skill or plugin is only as good as the instruction inside it, so the formula remains the foundation. That is why you should also keep your templates in a document of your own, so you can take them along with every switch.

Sam Altman wrote early on that this investment pays off:

“writing a really great prompt for a chatbot persona is an amazingly high-leverage skill and an early example of programming in a little bit of natural language”

— Sam Altman, CEO of OpenAI, on X in February 2023

In practice it works like this: pick your three most recurring writing tasks, work out one template for each along the formula, test it for a week and sharpen it based on what the output misses. Agree where improvements are kept. From a handful of templates onwards you have the basis of an AI way of working. ClickForest supports SMEs in Flanders that want to go further from there, with AI implementation and with agents; how such an agent works in marketing is covered in our guide on AI agents in marketing.

When are you ready for prompt engineering?

You are ready for prompt engineering as soon as the basic formula falls short: with complex reasoning tasks, with output that is never quite good enough, or with workflows you want to make reliable enough to automate. Then techniques such as giving examples, having the model reason step by step and chaining prompts together help.

Giving examples is called few-shot prompting; Anthropic recommends 3 to 5 examples for the best results. Prompt writing is a habit for everyone who works with AI, prompt engineering the deep dive for those who build systems with it. The techniques, and when to use which one, are in our follow-up article what is prompting and prompt engineering?, which also covers how to build a prompt library for your team. A short definition is in our glossary. If you then want to truly automate recurring tasks, marketing automation is the next step. Only start on that once the basics are in place.

What should you remember about writing prompts?

A good prompt is a good briefing: a clear goal, enough context with the why included, and an explicit format, supplemented by a role where it helps. Prompt in Dutch for Dutch copy and name the output language. Save what works in templates, projects or skills, refine the first answer, and check facts before you publish.

ClickForest helps SMEs in Flanders prompt better and use AI structurally, through hands-on training and guidance on the broader AI approach. Curious what is in it for your team? Get in touch and we will look at your most time-consuming tasks together.