AI Prompts for SMEs: A Practical Guide to Writing Them
Why your results with AI depend on your prompts first
In many Belgian SMEs the story repeats itself. Someone tries ChatGPT or Claude, types one vague sentence, gets a bland answer, and concludes that "AI isn't that useful". The tool is rarely the problem. The instruction is. An AI prompt (the instruction you give the assistant) plays the same role as a brief given to a new colleague: say "write an email to our customers" and you get generic text; say who it is for, why, in what tone and with which facts, and you get something you can use.
The context makes this pressing. According to Eurostat, 13.5% of EU enterprises with 10 or more employees used AI technologies in 2024, up from 8.0% a year earlier, and Belgium stood at 24.7%. Those figures date from January 2025 and have probably moved since, but the direction is clear: your teams, your competitors and your suppliers are all starting to use it. The advantage will not come from the tool, which is available to everyone, but from how you use it.
In this article I propose a simple method to write good prompts in an SME, before/after examples, a way to share the best prompts across your team, and the safeguards to respect. The aim is not to turn you into a "prompt engineering" expert, but to save you time starting this week, without taking unnecessary risks.
What a good prompt is (and is not)
A good prompt is not a magic spell or a secret formula found on social media. It is an instruction that is clear, contextual and verifiable. Anthropic's prompt engineering guide lists the basic techniques: be explicit about what you want, give examples, structure the request, assign the assistant a role, invite it to reason step by step, and split complex tasks into several requests. Nothing esoteric: these are principles of good communication.
It also helps to understand what a prompt cannot do. It does not guarantee accuracy: an assistant can produce a very convincing text that is wrong. It does not replace your data: if the information is not in the request or in a document you provide, the assistant will improvise. And it does not replace a process: a brilliant prompt used once by one employee does not improve the company, whereas an average prompt, documented and used by the whole team, does.
Finally, beware of needless sophistication. In the workshops I run, the fastest gains almost always come from a well-thought-out prompt of five or six lines, not a two-page text. Start simple, test, adjust. If you are still unsure about the tool itself, our ChatGPT, Claude and Gemini comparison for SMEs will help you choose before you fine-tune your instructions.
The 5-element method: role, context, task, format, constraints
To avoid starting from a blank page, I recommend a five-element frame to my clients. It is not official; it is a memory aid that reflects the principles above.
1. The role. Tell the assistant who it is for this task: "You are an administrative assistant in a distribution SME in Wallonia." This frames vocabulary and level of detail.
2. The context. Give the information a colleague would need: the customer, the history, the goal, the audience. It is the element most often forgotten, and the one that changes quality the most.
3. The task. Phrase one precise action: "Draft a reply", "Summarise in five points", "Compare these two quotes". One action verb per request beats three mixed together.
4. The format. Say what you want back: an email of 120 words maximum, a three-column table, a numbered list. Without guidance, the assistant chooses for you.
5. The constraints. State what to avoid or respect: formal tone, no delivery promises, do not invent figures, flag missing information. That last one is especially useful: explicitly ask the assistant to say when it does not know rather than fill the gaps.
This frame takes a few minutes to learn. Print it, pin it near workstations, and you will already have raised the level of the whole team.
Before/after examples for six common SME tasks
Here is how the method transforms ordinary requests. The examples below are illustrative: they describe frequent situations, not results measured at a client.
Replying to an unhappy customer. Before: "Write a reply to this customer." After: "You are head of customer service at a Belgian SME. Below are the complaint received and our returns policy. Write a reply in English, polite and empathetic, 120 words maximum, that acknowledges the problem, offers two solutions consistent with the policy and promises no deadline I have not given you."
Summarising a meeting. Before: "Summarise these notes." After: "Here are my raw meeting notes. Produce minutes in three blocks: decisions taken, actions (who, what, by when), open points. If a deadline or owner is missing, write 'to be confirmed' instead of guessing."
Preparing a commercial proposal. Before: "Make me a quote for a website." After: "From this client brief, propose a proposal structure in five sections (context, scope, deliverables, indicative timeline, terms). Do not invent any price: leave [TO BE COMPLETED] fields for amounts."
Analysing a sales table. Before: "Analyse these figures." After: "Here is the sales export for the last six months. Identify the three products whose sales fall the most, suggest two possible explanations for each, and state clearly what the data does not allow us to conclude."
Writing a job advert. Before: "Write a job ad." After: "Write an advert for a warehouse operator position in Charleroi, based on the job description below. Direct tone, no jargon, 200 words, with a 'what we offer' section limited to the items I provide."
Translating and adapting. Before: "Translate into Dutch." After: "Translate this email into Flemish Dutch, professional but warm register, keeping product names as they are. Flag any expressions you could not render naturally."
In each case the "after" version takes thirty seconds longer to write and saves several minutes of proofreading. For more ideas of tasks to hand to AI, browse our article on tasks SMEs can automate with AI.
Iterate: the conversation beats the perfect prompt
A common mistake is hunting for the perfect prompt on the first try. In practice, the best results come from a multi-turn exchange. You send a first version, read the answer, then correct: "Shorter", "Add a figure that I will supply", "Rewrite the second paragraph, it is too salesy".
Three habits speed up the loop. First, ask for variants: "Give me three email subject lines, from the most sober to the most compelling." Choosing is faster than writing. Second, have the assistant explain its assumptions: "List the assumptions you made to produce this answer." That shows you what was made up. Third, split long tasks. Rather than ask for a full report in one go, ask for an outline first, validate it, then have it written section by section. Splitting work into steps is also among the techniques named in the Anthropic documentation mentioned above.
Keep one hygiene reflex too: open a new conversation for each distinct subject. Mixing a quote, a job advert and a reminder email in the same thread blurs the context and degrades answers.
Build a prompt library shared by your team
If one employee finds a prompt that works and keeps it to themselves, the company gains nothing. The solution is a prompt library: a deliberately modest shared document where the team files the instructions that have proven themselves.
To keep it alive, each entry should fit in a few lines: the task name, the prompt with fields to replace in square brackets, an example of an acceptable result, the date of last check and the name of an owner. A simple shared table is enough. Avoid complicated tools at the start.
Begin with five to ten prompts matching your company's most repetitive tasks: replies to incoming requests, meeting minutes, reminders, product descriptions, first drafts of internal documents. Appoint one person to review additions once a month and remove what is no longer used. This light discipline stops the library turning into a graveyard of obsolete texts.
It is also an excellent training aid. A newcomer learns in an hour what the team took months to discover. To structure that learning, see our guide to training your team for AI adoption.
Confidentiality, GDPR and the AI Act: what never to paste into a prompt
A good prompt is also a careful prompt. Before pasting a document into an assistant, ask yourself a simple question: "Would I be allowed to send this content to an outside provider?" If the answer is no, or "I don't know", don't.
In practice, avoid including identifiable personal data about customers or staff, information covered by professional secrecy, trade secrets (purchase prices, margins, confidential contracts) and logins or passwords. When a task requires handling such data, replace it with fictitious or anonymised data, or use a business-grade plan whose data-processing terms you have read. Rules depend on the tool and the subscription: check them, don't assume. For the general framework, see our article on data security when using AI in SMEs and, for privacy questions, the website of the Belgian Data Protection Authority.
On the regulatory side, the European Commission states on its AI Act page that AI literacy obligations have applied since 2 February 2025. In practical terms, training your teams in sensible use, for example with a prompt guide like this one, is part of a responsible approach. I am not a lawyer: for your specific obligations, rely on the texts in force and, where needed, on legal advice. Our article on the EU AI Act for Belgian SMEs gives a first orientation.
Finally, always proofread what the assistant produces before sending it to a customer or using it to decide. The human remains responsible for the result.
Next step: from good prompts to real processes
Mastering prompts is an excellent first step, but it has limits. When the same task comes back twenty times a week, copy-pasting instructions into an assistant becomes a chore in itself. That is the moment to ask whether the task deserves to be automated: connected to your tools (email, CRM, accounting), triggered automatically, with human review at the critical points.
My advice is to proceed in order. Start by establishing good prompt practices in the team for four to six weeks. Spot the most-used prompts: they are your automation candidates. Measure the time actually saved on those tasks, without extrapolating from impressions. Only then does it make sense to invest in a custom integration. This progression avoids the classic mistake of paying for a complex project to meet a poorly understood need; it is one of the traps described in our article on AI integration mistakes to avoid.
If you want to move faster, I can help you audit your team's current usage, write a first prompt library suited to your business, then identify what is worth automating. You will find the detail of our approach on the AI integration page.
Conclusion: five actions to start this week
Writing good prompts is not reserved for specialists. It is a communication skill any team can acquire in a few weeks, provided it goes about it methodically.
This week you can: 1) print the role-context-task-format-constraints frame; 2) rewrite two or three requests your teams already make, in "after" form; 3) create a shared table to collect the prompts that work; 4) remind everyone what must never be pasted into an assistant; 5) set a date, a month from now, to review the time saved.
Would you like an outside view or support to structure all this? Get in touch for an initial conversation, or explore all of our services.