Skip to main content
Guides & How-To

Prompt engineering: how to get better answers from AI

Prompt engineering is the art of phrasing your instructions to an AI so that it delivers a usable answer the first time. A prompt is the message you send to the model, and the quality of the answer almost always depends more on how you ask than on which model you use. Most bad AI experiences are not caused by a bad model, but by a prompt that is too vague. The good news: it can be learned in an afternoon.

Author: Agent360 editorial team
Published: 31 August 2026
Last updated: 3 September 2026
Reading time: 7 min read

What actually makes a good prompt?

A good prompt is an instruction so clear that a capable new colleague could carry out the task from it alone, without having to ask follow-up questions. Think of the AI as a quick, willing employee who knows a lot in general, but knows nothing about your business, your customers or the specific task until you tell it.

Bad prompts typically have one of three problems: they are too short ("write a text about our product"), they lack context (the model doesn't know the audience or the purpose), or they leave the format to chance. The result is generic, and you spend more time correcting the answer than you would have spent writing the piece yourself.

A good prompt does the opposite: it is specific, it provides context, and it states clearly what the output should contain and what it should look like.

Five principles that lift almost any prompt

The five techniques below are the most effective, and they work across tasks, whether you're writing emails, summarising documents or analysing data.

1. Be specific

Remove all ambiguity. Say exactly what you want, how long it should be, who it is for and what it should achieve. "Keep it short" is imprecise; "write 3 bullet points of no more than 15 words each" is an instruction the model can hit.

2. Give context and a role

Tell the model who it is and what it is working with. A role ("you are an experienced Danish B2B copywriter") calibrates tone and level of expertise. Context ("the text will go on our website and be read by purchasing managers at mid-sized manufacturers") makes the answer relevant instead of generic. It also helps to attach background material (an extract from your earlier texts, a product description or a set of minutes) so the model has something to build on.

3. Show an example

A single example of the output you want says more than three lines of explanation. If you want answers in a particular format or tone, show the model one finished example and ask it to imitate the pattern. This technique is often called "few-shot" and is among the most reliable ways of steering output.

4. Ask for an approach on complex tasks

For tasks with several steps (an analysis, a calculation, an assessment), it helps to ask the model to think in steps ("first go through the material, then set out the pros and cons, and conclude at the end"). The model makes fewer mistakes when it doesn't have to jump straight to the answer, and you can follow the reasoning along the way.

5. Describe the format you want

State explicitly how the answer should be structured: a table, a bullet list, a paragraph, JSON, an email with a subject line. If you don't ask for a format, the model picks one itself, and then you spend time reshaping it afterwards.

Before and after: the same task, two prompts

The difference between a weak and a strong prompt is rarely more work, just more precision. Here are three typical examples (generic illustrations):

Task Weak prompt Strong prompt
Customer email "Write a reply to an unhappy customer." "You are a customer service rep. Write a friendly, apologetic reply of no more than 120 words to a customer whose delivery is 4 days late. Offer a solution, avoid apologies that sound like a template, and end with a concrete next step."
Summary "Summarise this document." "Summarise the attached minutes in 5 bullet points for a management team that didn't attend. Highlight decisions and who is responsible, not the discussion. No more than 20 words per point."
Idea generation "Give me ideas for marketing." "You are a marketing adviser for a Danish B2B software company. Suggest 5 angles for LinkedIn posts that speak to IT managers about data security. For each: a headline and one sentence on the angle. Avoid buzzwords."

Note the pattern: role, context, concrete boundaries, format. The strong prompt takes thirty seconds longer to write and typically saves several rounds of corrections.

A simple template you can reuse

When in doubt, fill in four fields before you hit send:

  • Role: who is the model? ("You are a Danish legal assistant…")
  • Task: what should it do, specifically?
  • Context: what does it need to know to get it right? (audience, purpose, background material)
  • Format: what should the answer look like? (length, structure, tone)

These four fields cover the vast majority of everyday tasks, and they are easy to teach to a whole team.

How to iterate your way to a good result

The first prompt doesn't have to be perfect. Often the quickest route is to send a reasonable prompt, look at the answer and then adjust: "That's too long, halve it." "The tone is too formal, make it more direct." "Add a concrete example in the second paragraph." The AI remembers the conversation, so you can fine-tune in dialogue instead of writing the perfect instruction from scratch. Treat it as a conversation, not a one-off order.

When prompt engineering isn't enough

Here is the important limit that many businesses hit: prompt engineering makes a single answer better, but it doesn't solve a task that has to run again and again, pull data from your systems or take actions on its own.

A prompt is a one-off instruction in a chat window. If you find that you write the same kind of prompt every day, copy data in and out by hand, or want something to happen automatically without a person at the keyboard, you have gone beyond what a prompt can do. Typical signs:

  • The task repeats: same prompt, new data, many times a week.
  • It needs access to your data: customer system, emails, documents, database.
  • It has to carry out actions: create a case, send a reply, update a record, not just produce text.
  • It has to run unattended: at night, on an event, at a set time.

When one or more of these points apply, the answer is not a better prompt but a system: an AI agent or an automation in which the prompt is built in, connected to your data and set up to run on its own. Prompt engineering is still the foundation (an agent builds on exactly the same principles), but the instruction becomes a fixed, tested part of a workflow instead of something an employee writes from scratch every time.

In short: prompt engineering makes your employees faster. An agent takes the task off their plate altogether.

Getting started

Start small. Pick one recurring writing task in your day-to-day work, build a good prompt with role, task, context and format, and save it somewhere the whole team can find it. Share your best prompts internally; they are worth just as much as a good template. When you begin to feel the limit (the repeated tasks, the data access, the actions), that is the point at which it makes sense to look at automation.

Shall we look at your workflows together?

Agent360 helps Danish businesses get from "we use AI now and then" to AI that genuinely lifts the work, both good prompt practice and the agents and automations that take the recurring tasks off your hands entirely. We advise on where you'll get the most value, and build the solutions when that is the right way forward.

Book a no-obligation AI consultation with Agent360, and we'll look at your specific workflows and work out where you can make progress fastest.

What takes up your time?

Write down the task that takes up the most time. We reply whether it is worth building, and we say so if it is not.