Prompt engineering: how it works, the building blocks and 14 prompts to copy

What prompt engineering is

Prompt engineering is the process of learning to phrase our instructions so that an AI tool gives us the most useful answers. The principle is simple: the better the prompts, the better the results. You don’t need to know how to code. You need to be clear about what you actually want.

Generative AI tools produce surprisingly good lucky hits even without any preparation, and they can take you a long way. But until you refine your prompts deliberately, luck is all it is. And you can’t plan a working day around luck: as soon as a result has to be reliable, you need control over what comes out.

First, because it holds many people back: you can’t break anything by experimenting. A prompt is a message, not an intervention in a system, and a poor result costs nothing but another try. Only one rule really matters: don’t enter confidential data into tools your company hasn’t approved. Everything else is a playground, and playing is the quickest way to get a feel for what works.

Why prompt engineering is like playing with Lego

The difference is obvious. A well-built prompt gives the tool precise instructions about what you expect, and the result is far more likely to be what you had in mind. You don’t have to be a programmer or a product designer for this: you just have to be able to follow the instructions, and often it is enough to take a good prompt from a collection and swap in the parts that fit your goal.

In prompt engineering, technical knowledge and programming languages matter less than meta-knowledge: knowing that instructions exist, how to find and adapt them and, above all, what exactly you want to achieve.

From a good prompt to your own tool

This is where the real leap happens. A good prompt solves one task. It gets interesting when you stop writing it from scratch every time and instead record what the model needs to know about your work: your terms, your tone of voice, your audiences, your rules. The Lego robot on the box becomes a butler you can’t buy anywhere.

The common tools offer several ways to do this, all with the same aim:

This pays off from the second repetition. And it is the difference between a tool anyone can use and one that knows your work.

Natural language as a programming language

A prompt can be almost anything: a question, a statement, a request, a whole dialogue. It is the input from which the AI generates an output. What matters is the language: unlike traditional programming languages with a fixed syntax, you communicate in your own. You don’t adapt to the machine; the machine adapts to you.

Details beat chance: the quickest way in

The most effective first step is almost too obvious: put the relevant details into the prompt. Models take instructions very literally. You get what you ask for, and usually only that.

You can see the same effect with images. Two prompts, one difference: the amount of detail.

Neither image is bad. But only the second is a decision. The same goes for text: without details, a language model gives you its default answer, and default here means generic.

The eight building blocks of a good prompt

Almost all good prompts are clear and focused, no longer than necessary, and they steer the model through examples, format and style. The eight building blocks below are sorted by impact. They are explicitly not a checklist: a good prompt uses only as many as its goal requires. Too many instructions at once bury the important ones.

What has changed from the classic list: “Examples over descriptions” is new and deliberately near the top. A single sample of the result you want is more reliable than any description. “Format” has replaced the old point “Length”, because length is only part of it, and “Keywords” has been folded into “Language and style”, where it belongs.

A prompt collection to copy

14 prompts for recurring tasks, each with placeholders in square brackets. Replace the highlighted parts and leave the rest as it is. The buttons below take you straight to the area you need.

A note on language variety: models default to whichever English they see most, which is usually American. If your text should follow British spelling and conventions, say so in the prompt. That is why several examples include “British English”; change it to American English if you write for the US.

Research and analysis

You are a market analyst. Look at our competitor’s website [URL].
Compare their value proposition, how they address their audience and their pricing logic with our site [URL].
Output: a table of five points where we should catch up, each with a one-sentence rationale.
At the end, tell me which claims you couldn’t verify.

Describe the main target audience for [product] in [market].
Base it on these details: [short description, price, sales channel].
Output: three short profiles, each with their job, goal, biggest obstacle and the question that must be answered before they buy.
No demographic clichés, no invented studies.

Check the following text for three things: unclear statements, unsupported claims and repetition.
Answer as a list, with a quote and a suggested fix for each issue. Don’t rewrite the text.
Text: [paste here]

Writing and content

You are the content manager at a manufacturer of [product category].
Suggest four blog topics that answer our customers’ questions and lead to our products.
For each topic: working title, search intent in one sentence, three outline points.
No topics we have nothing of our own to contribute to.

Write the opening of a landing page for [product].
Audience: [description]. Benefit: [one sentence]. Proof: [figure, reference or certificate].
No more than 90 words, British English, address the reader directly, one idea per sentence.
No superlatives, no words such as “innovative”, “seamless” or “tailored”.

Cut the following text to [number] words without dropping any facts.
Keep the technical terms, replace filler words, break up long sentences.
Then tell me in one sentence which piece of information you would sacrifice next if it had to be shorter still.
Text: [paste here]

Email and campaigns

Here is the current copy of our email campaign: [paste text].
Audience: [description]. Goal: [clicks, replies, meetings].
Revise the copy so the benefit comes in the first sentence and only one call to action remains.
Give me three subject lines under 50 characters and justify each in a few words.

You are responsible for our social media channels. Draft three versions of a LinkedIn post about [topic].
Context: [occasion, audience, link].
Version 1 matter-of-fact, version 2 with a practical example, version 3 opening with a question.
Each no more than 700 characters, no strings of hashtags, no emojis.

Language, terminology and markets

Write [type of text] on [topic].
Use only the following terminology: [term = spelling, one per line].
Banned terms: [list].
If you’re missing a term, mark the spot with [TERM?] instead of inventing one.

Here are three texts that capture our tone of voice well: [Text 1], [Text 2], [Text 3].
First, describe in five bullet points what these texts have in common.
Then write [new text] in the same tone.
Stick to your bullet points and don’t slip into a promotional tone.

Compare the source text and the translation section by section.
Flag only: missing content, wrong figures and units, inconsistent terminology, passages that need different wording in the target market.
Format: a table with source text, translation, finding, suggestion.
Don’t retranslate anything.

Adapt the following text for [target market].
Check: units of measurement, date and number formats, currency, examples, tone and forms of address, legal disclosures.
List all changes with reasons before you output the new text.
Text: [paste here]

Roles and perspectives

Answer as [role, e.g. marketing controller, civil engineer, head of purchasing].
Task: [question or assignment].
Name three points that non-specialists overlook in this context and justify each in one sentence.

You are a sceptical buyer reading the following text for the first time.
Ask the five questions you would ask before making an enquiry.
Then mark the places in the text that leave these questions unanswered.
Text: [paste here]

Style and role: your strongest lever

Language models have learned from huge amounts of text which phrasings are usual in which context. When you specify a style or a role, you tell the model which of these areas to draw on. In other words, you choose the language of the answer: not just English or German, but also the language of a specialist field, of reportage or of a children’s book.

The language in your prompt works like a map. Write in marketing terms and the model answers in marketing terms. Ask for the perspective of a civil engineer and it reaches for the matching vocabulary. And you always set a style, whether you mean to or not: if you don’t name one, the model simply picks up the tone of your own input.

Length: two kinds of model, two rules

With text models, plenty of context helps. With image models, the opposite is often true, as these two attempts show. First, a very detailed prompt:

The first image is missing the sweater, the necklace and the walking stick, even though they were in the prompt: too many details, and the important ones get lost. The result was also “highly realistic” but not a photograph, because nobody had asked for one. Models guess when you leave it to them. In the second prompt, “holding a walking stick” comes near the start and “photography” has been added. Same idea, different priorities, different result.

One word, nine results

The principle is clearest when everything stays the same and only the style changes. The same sentence nine times, with one word swapped each time:

• PromptBarbara Kruger style image of a woman talking with a robot
• PromptDe Stijl style image of a woman talking with a robot
• PromptHighly realistic image of a woman talking with a robot
• PromptPointillist style image of a woman talking with a robot
• PromptASCII image of a woman talking with a robot
• PromptRococo style image of a woman talking with a robot
• PromptMinimalist style image of a woman talking with a robot
• PromptPolaroid style image of a woman talking with a robot
• PromptArt Nouveau style image of a woman talking with a robot

The creativity here doesn’t happen during generation but long before it: when you come up with the subjects, styles and words. And afterwards, when you choose. Which of the four or forty results works, and why? Which wording do you keep? The thinking stays with us; it just moves into the preparation and the follow-up.

For reference: styles, authors, visual languages

The lists below are there for exactly the moment when the right word escapes you. Open a list, copy a term, drop it into your prompt.

Let the AI write the prompt

The most useful trick is also the most often overlooked: you don’t have to write the perfect prompt yourself. Describe in everyday language what you want to achieve and have a prompt written for you. The model knows better than any guide which information it needs, and it will ask for whatever is missing.

I want to use an AI tool to [goal in everyday language].
Write me a prompt for it.
First, ask me for everything you need: audience, context, format, length, tone and anything that must not appear.
Then give me the finished prompt as a block I can copy, and explain in three bullet points why it is built that way.

Here is my prompt: [paste prompt].
Assess it on six points: role, goal, context, material, format, constraints.
For each point, say what is missing or ambiguous.
Then give me an improved version without changing my content requirements.

We regularly write [type of text] for [audience].
Turn this into a reusable prompt template with placeholders in square brackets.
The template should include role, goal, context, format and constraints. It should work for similar cases without changes.
At the end, name the three placeholders that will need adjusting most often.

A tip from practice: check the generated prompt before you run it. Models like to add assumptions you never made, such as an audience or a tone. Delete whatever isn’t right, and the prompt becomes a real template.

Six tricks you rarely find in guides

The building blocks cover the structure. These six points decide how good the result really is in everyday use.

Follow-up prompts: the work starts after the first attempt

The real strength of chat-based tools is the dialogue. A follow-up prompt is simply an answer to the answer: you say what worked and what missed the goal, and you shape the result step by step. Four types cover almost every case.

With image models, the same principle works differently: instead of discussing, you vary the prompt, swap individual words or move important details to the front. That is a dialogue too, just a one-sided one.

From prompt to skill: build it once, use it every time

This last step is the one most often skipped. Once a task has gone well, don’t save the prompt in a document; save the procedure in the tool itself. After that, one sentence is enough: “Do it the usual way for our product descriptions.”

Most tools now offer a way to do this, under various names: saved skills, projects with their own rules, custom assistants or persistent instructions. What they have in common is that they store three things you would otherwise type again every time: the role, the material and the procedure.

That worked well. Summarise the procedure we just used as a reusable workflow:
what role you take on, what information you need from me, the order you work in and how you check the result.
Write it so I can save it and call it up with one sentence next time.
Call the workflow [name].

This is the point where the tool becomes a colleague who knows your work. And it is why the effort of writing a good prompt pays off: you write it exactly once.

Two things to bear in mind. First, document what goes into a skill, or in six months nobody will know why the results look the way they do. Second, before sharing a skill, check that no confidential details have ended up in it, such as client names from the sample material.

Limits, confidentiality and quality

Prompt engineering makes results more predictable, not more true. Three things remain your job: checking the facts, setting the terminology and deciding what may be entered into a tool in the first place.

For texts that will be published or translated, the next step is proofreading AI texts and post-editing to ISO 18587.

To take away

Both PDFs contain the content of this page in a format you can print and share with your team. They are currently available in German only.