What is a prompt and how to write good instructions for AI

Updated: 24 October 2026 · reading time about 9 minutes

Two people in the same company use the same AI tool. One gets answers that can be sent to a customer straight away, the other gets generic text that has to be rewritten. The difference rarely lies in the model. It usually lies in what was typed into it – the prompt. Below we explain what a prompt is, what a good instruction consists of, how to improve it and how prompts work in company chatbots and automations – with examples you can copy right away.

What is a prompt? A definition

A prompt is an instruction, question or set of directions that you type into an artificial intelligence tool to get a response: text, a summary, a table, code or an image.

The simplest prompt is one sentence, e.g. “Summarise this email in three bullet points”. An elaborate prompt can run to several paragraphs: a role description, input data, the required format, examples and constraints. A language model (LLM) does not know your intentions or your company's context – it only sees what is in the prompt and earlier in the conversation. That is why the quality of the answer depends so heavily on the quality of the instruction.

The skill of deliberately writing and testing prompts is called prompt engineering.

Prompts in practice: three kinds of instructions

When we talk about prompts in a company, it helps to distinguish three situations – each calls for a slightly different approach.

User prompt

What you type into the chat window in ChatGPT, Claude or Gemini. You write it once, look at the result and refine it. Speed matters here, along with the ability to clarify the instruction in the next message.

System prompt

An instruction hidden from the user that the app's creator sets permanently. It defines who the assistant is, what it may talk about, in what tone and what it must not do. Every company chatbot has one – the customer never sees it, but it decides how the bot behaves.

Prompt in an automation

A template instruction run automatically, hundreds of times a day, with data filled in: “Assign this email to one of the categories…”. Nobody reads every answer as it comes in, so the prompt has to be predictable and the output must follow a strict format that the next step of the process can understand.

What a good prompt is made of

There is no single “magic” template, but good instructions usually contain the same elements. Not every prompt needs all of them – a simple question needs two, while an automation template benefits from almost all.

A good sanity check: imagine giving the same instruction to a new employee on their first day. If they would need to ask about five things, the model doesn't know them either – except that instead of asking, it will guess.

Example: a weak and a good prompt

Take a typical situation in an online store: a customer writes that her parcel hasn't arrived. The first version of the instruction usually looks like this:

The model will write correct but bland text – it may promise compensation you don't offer or use a tone that doesn't match your brand. The same request with context, data and constraints:

The second prompt is longer, but the answer is usually ready to send after a small edit. Note two details: the input data is clearly separated from the instructions (here with quotation marks), and the prohibitions are written out explicitly. The model won't guess company policy – you have to give it.

How to write prompts: six techniques that work

1. Start with the goal, not the topic. “Write a post about our promotion” gives a worse result than “Write a Facebook post that encourages existing customers to use the promotion before the end of the month”.
2. Show an example. One model product description or one good reply tells the model more than a paragraph of adjectives about a “professional yet friendly tone”. This technique is called few-shot prompting.
3. Specify the output format. A bullet list, a table with columns, JSON with specific fields, a maximum word count. The stricter it is, the easier the result is to use.
4. Ask for step-by-step reasoning on tasks that require analysis – comparing offers, checking a contract, calculations. A model that first lists its assumptions is less likely to jump to a wrong conclusion.
5. Let the model ask questions. Add: “If you're missing information, ask me questions first”. It's a simple way to stop the model filling gaps with made-up details.
6. Split large tasks. Instead of “prepare an offer” – first a list of the client's requirements, then the structure of the offer, then the individual sections. Shorter steps are easier to check and correct.

And most importantly: the first answer is a draft. Instead of rewriting the prompt from scratch, tell the model what to fix – “shorter”, “no jargon”, “add the net price”. Iteration is a normal part of the work, not proof that the prompt was bad.

Prompt examples for a small business

A few instructions that work best in SMEs. Replace the square brackets with your own data.

Product description for a store

Document summary

Message classification

The last example is a typical automation prompt – a one-word answer is easy to pass on, e.g. to a tool such as n8n, which then routes the message to the right person.

A practical note: before pasting a contract, invoice or customer message into a public chat, check that you are using a business account where data is not used to train models, and that you are not sharing personal data without a legal basis. Sensitive documents are better stripped of names and numbers or processed in a solution running on the company's own infrastructure.

The system prompt in a company chatbot

In an AI chatbot for business the prompt is not a single sentence typed by the customer but a carefully prepared instruction that applies to every conversation. It usually describes:

• role and tone – e.g. “salon assistant, formal address, short and to the point”;
• scope – what the bot talks about and on which topics it hands over to a human;
• sources – that it answers only on the basis of the price list, terms and knowledge base;
• behaviour when it has no answer – it admits it doesn't know and offers contact;
• actions – when to ask for contact details, when to offer to book an appointment.

The prompt alone is not enough for the bot to know the company's offer. Current documents are provided separately, through a knowledge base (RAG) – the prompt says how to answer, the knowledge base supplies what to answer. When the bot also has to take actions in company systems, it becomes an AI agent, and the prompt also describes when it may use a given tool.

A system prompt is tested like software: you prepare a few dozen real customer questions – including difficult, unusual and malicious ones – and check that the answers are still correct after every change to the instructions.

The most common mistakes and limitations

Too general an instruction

“Write something about our company” forces the model to guess. The result is text that fits every company – and therefore none.

Believing a good prompt removes hallucinations

Even the best prompt won't make the model know facts that are neither in the instruction nor in the data. Numbers, prices, names and regulations always need checking. A prompt can reduce the risk (“answer only on the basis of the text below, and if the information is missing, say so”), but it doesn't eliminate it.

Contradictory instructions

“Be detailed” and “max. 50 words” in one instruction is a typical example. The model will pick one of them, not necessarily the one you cared about. With long prompts it's worth reading them in full from time to time and removing additions that contradict each other.

Prompt injection

If the model reads external content – emails, web pages, customer messages – someone can hide their own instruction in it, e.g. “ignore previous instructions”. This is called prompt injection. That's why company deployments don't rely on the prompt alone: the bot has limited permissions, and actions with financial or legal consequences are approved by a human.

Prompt engineering – does a company need a specialist?

For everyday work with an AI chat, it's enough for the team to know the few principles above – an hour of training and some practice. Prompts that run unsupervised need a separate skill set, though: in a website chatbot, in an automation handling hundreds of messages or in an agent with access to company systems. There, what matters is not just the wording but also a test set, versioning of changes and monitoring of answers after launch. This is usually part of an implementation project, not a separate full-time role.

Summary

A prompt is the instruction you give an AI model – and the cheapest way to better results. A good prompt clearly defines the task, provides context and data, requires a specific format, says what not to do and, where possible, shows an example. In everyday work it's enough to iterate: refine the answer instead of starting from scratch. In chatbots and automations the prompt becomes part of the system and needs testing just like code. If you want to move good prompts from the chat window into a process that runs in your company around the clock, let's talk about it in a free consultation.