Prompt Engineering · Structure · Advanced techniques

How to write prompts well

Whether an AI answer is mediocre or ready to use often comes down to the instructions. Here you'll learn how to put them together so the AI understands what you want: the five parts of a good prompt, advanced techniques and prompts to copy. All from practice.

🦴 The 5 pillars of a good prompt

Forget instructions like “Write me an article about cars.” If you want a result you can use straight away, put most of these five parts into your prompt.

1. Role (Persona)

Tell the AI who it should be. It changes the tone, the vocabulary and how expert the answer is. Describing what the role does and for whom helps more than superlatives like “elite”. It doesn't have to be IT or marketing — the AI will play anyone from everyday life.

  • IT/Business: You are a UX designer who builds online shops for small local businesses and judges them by whether you can buy something quickly and without mistakes.
  • Trade/Practice: You are an experienced car mechanic with 20 years of practice who can explain faults to a complete layperson.
  • Education: You are a patient elementary school math teacher who uses lots of playful comparisons.
  • Home: You are a Michelin-starred chef who now cooks only from leftovers in an ordinary family fridge.

2. Context (Situation)

Explain why you're giving the task and who the result is for. The AI doesn't know your background, problems or limits. Whatever you don't tell it, it will guess.

  • IT/Business: We're preparing a new online shop with watches for managers who have no time.
  • Family: I'm preparing a family party for 15 people, I only have 3 hours to get ready and half of the guests are vegan.
  • Problem: I'm trying to teach my 8-year-old son fractions, but he hates it and keeps crying at the table.
  • Law/Crisis: My neighbour keeps parking in front of my garage entrance and I need to talk to him without us ending up in a fight.

3. Task

Write clearly and briefly what exactly the AI should do. Use action verbs: write, compare, choose, fix.

  • IT/Business: Go through the text below and turn it into 3 ideas for an engaging email newsletter.
  • Home: Make me an exact shopping list and a cooking schedule for the whole Saturday.
  • Education: Come up with 3 fun games (each 10 minutes at most) that explain fractions to him with Lego or a sliced apple.
  • Car/Hobby: Give me the 3 most common causes of this noise from the clutch and how I can tell whether I can still drive the 50 km to the garage.

4. Rules and limits (Constraints)

This is where it's decided whether the result is ready to use: length, budget, what to leave out, what to stick to. It works better when you write what you want instead, and why. A bare “you must not” tells the AI where not to go, but not where to go.

And don't shout at it in capital letters. New models take rules seriously without “FORBIDDEN” and “NEVER”, and with shouting they overdo them.

  • IT/Business: Be specific: instead of “innovative” and “revolutionary”, write what the product does and how much time it saves the customer.
  • Education: Explain it in words an eight-year-old boy knows, and wrap the whole thing in a short story.
  • Hobby/Repairs: Start with the cheapest and simplest repairs. Only mention replacing the engine when nothing else makes sense.
  • Budget/Diet: Ingredients under €8 per kilo. No mushrooms, one guest is allergic to them.

5. Output format

This stops the AI from spitting out a solid wall of text. Tell it in what form you'll use the result straight away.

  • IT/Business: Give me the result as a table with 3 columns: Problem, Solution, Priority.
  • Home: List it as a checklist of tasks, split into Friday evening and Saturday morning.
  • Crisis/How-to: Short bullet points with a bold heading: Cause – How I can easily test it – Rough repair cost.
  • Training/Study: Put it in a table where the rows are the days of the week and the columns are: Exercise, Reps, What to watch out for.

Here's what all five pillars look like together, once for business and once for everyday life:

Example 1: A cold email for companies (B2B)
[ROLE] You are a B2B copywriter who writes cold emails for companies that sell software.
[CONTEXT] We sell attendance tracking software. We target owners of mid-sized companies who lose money because of attendance mistakes.
[TASK] Write one cold email to reach these owners.
[RULES] 5 sentences at most. Start straight with their problem, no polite opening line. Use the PAS method (problem – agitation – solution). End with one clear call to action, such as a short call.
[OUTPUT] 3 different versions one below the other, separated by a line.
Example 2: Home DIY
[ROLE] You are an experienced carpenter giving advice to a complete beginner.
[CONTEXT] I want to build a shoe bench out of pallets myself this weekend. I only have a cordless drill and a hand saw and I've never built anything like it.
[TASK] Come up with a simple, sturdy design and write me instructions that even a complete layperson can follow.
[RULES] Skip general advice like “measure twice”. For every step, warn me about one mistake I'll almost certainly make as a beginner (for example that the wood splits without a pilot hole).
[OUTPUT] Split it into a “Shopping list for the DIY store” and a numbered “Cutting and assembly plan”.

🚀 Advanced techniques (for connoisseurs)

When five pillars aren't enough and you're tackling a complex task, coding, brainstorming or a big audit, these methods help.

A) Few-shot (learning from examples)

Instead of explaining a style at length, show it. AI is excellent at copying structure and tone when it gets a sample. Give 2–5 examples and make them different from each other. If they're all the same, the AI copies what you don't want too, like always the same first sentence.

An email in my style
I want you to write emails in my style. Here are 3 of my emails that turned out well:

<examples>
[PASTE EXAMPLES]
</examples>

Write a new email on the topic “Delivery delay”. Keep the same tone, sentence length and greeting as in the examples, but write new content.

B) Let the AI think

For complex tasks (logic, calculations, code, planning), choose a model or mode that thinks longer. Every AI names it differently and the names change often. The current overview is in the Which AI for what guide.

You no longer need the old “magic formula” “think step by step”. Today's models think on their own and according to guides from OpenAI and Google such a sentence is more likely not to help. Instead, write what the result should be and how you'll know it's done. Leave the method to the AI.

The goal instead of the method
Goal: [what you want, in one sentence]
It's done when:
1. [first condition, for example “the totals add up”]
2. [second condition, for example “I can do every step with what I have at home”]
3. [third condition]
Choose the method yourself. Before you answer, check the result against these points.

C) Phases and pauses (gamebook)

For big tasks (a whole book, a big website audit, a holiday plan), don't ask for everything at once. Stop the AI halfway and approve the direction. That way it won't head off the wrong way and produce ten pages of unusable text.

Stop and wait for me
Work in 3 steps. Now do only STEP 1 (the outline). Then stop and wait until I approve it. Move on to STEP 2 only when I write “OK”.

D) Reverse prompting (let the AI ask you)

Often you don't know how to write the prompt or what the AI needs for a good result. Let it ask by itself. Or let it turn your request into a full prompt with the five pillars.

Let the AI ask first
I want to write a good CV for a manager position. Before you start writing, ask me 5 questions about my experience that you really need for it. Wait for my answers.
Let the AI write the prompt for you
Here's what I want from you: [describe in your own words what you need]
Don't do anything yet. First rewrite it into a prompt with five parts: Role, Context, Task, Rules, Output. Where information is missing, ask me. Show me the finished prompt so I can edit it and then use it.

E) Self-correction

The first version is usually average. When you add a review step to the prompt, the AI finds and fixes mistakes before it shows you the text. It works best when you tell it what to check against.

Strict editor
When you've written the first version, read it as a strict editor. Find the 3 weakest spots (clichés, boring parts, mistakes, unclear sentences) and give me a corrected version 2.0 right away. Below it, write in one sentence for each spot what you changed.

F) Multi-persona (a debate of several experts)

Need to see a problem from several sides? Let the AI act out a debate between several experts. It's great for brainstorming and for finding blind spots in an idea.

Three experts on an idea
You are the moderator of a debate about my new business idea: [describe the idea]. Let three experts discuss it: 1. an enthusiastic visionary, 2. a very sceptical accountant, 3. a practical lawyer. Let them argue over 3 rounds. At the end, sum up the main risks and what I should verify first.

G) Stop the AI making things up (hallucinations)

Language models want to please you, even when they make up a book, a law or a statistic along the way (so-called hallucinations). When you work with facts, let the AI say “I don't know” and ask for sources. That helps more than demanding “absolute certainty”, which a model can't measure.

Facts with sources
When you're not sure about something, say so. For numbers, laws, books and statistics, give a source. At the end, separate what you know for sure, what you're estimating and what I should check.
Answer only from the pasted text
Answer only from the text below, not from your own knowledge. Add a word-for-word quote from the text to every claim. Mark anything that isn't in the text as “not in the text”.

<text>
[PASTE TEXT]
</text>

Even this won't stop hallucinations completely. Always double-check health, money and law.

H) Fresh information and sources

AI only knows the world up to when it was trained. For prices, news, app versions and anything from the last few months, turn on web search and ask for links.

Grok can also search posts on X. Handy when you want to know what people are writing about something right now.

Grok: what people are posting on X now
Find posts on X from the last 7 days about [topic]. Summarise the main opinions for and against and give 2–3 links to posts for each. Separate people's opinions from verified facts and write what needs checking elsewhere.
For any AI that can search the web
Search the web for current information about [topic]. Use sources no older than [year]. Give a link and a date for every piece of information. When sources contradict each other, say so.

📝 Formatting and Markdown

How do you paste source texts and keep your prompt tidy? When a prompt is longer than half a page, separate the instructions from the data with headings or tags (XML tags). The AI finds its way around better. With long material (a whole document, a contract, a transcript), put the text at the top and the task below it.

How to separate instructions from data
You are the proofreader of our company website.

<SOURCE_TEXT>
[Paste your long text here]
</SOURCE_TEXT>

### RULES ###
- We address customers politely and formally.
- We write in the active voice (“we fixed it”, not “it was fixed”).

Fix the grammar in the text above and stick to the rules. Write out the corrected text and a list of changes below it.

💎 Big prompts from this site

You know the theory now. Here's what longer prompts built from these pillars and techniques look like:

There are more big prompts in the other guides. We're rewriting them step by step by the same rules — some of them still write prohibitions in capital letters:

You don't have to write the role and rules over and over

What you write into every prompt again (who you are, how you want answers, what to avoid) you can save once in custom instructions. How to do it is in the Set up AI section. The generator puts the rules together for you.