Prompt
engineering.
The art of addressing machines precisely and getting out of them what you really mean. Not luck, not magic. Method.
What actually is
a prompt?
A prompt is the text you give an AI, and it determines what comes back. Sounds simple. It is not. The model does not "read" your prompt like a human. It calculates which sequence of tokens is most likely to follow your words, based on everything it was trained on.
That means: Every word you choose shifts the probability distribution of the output. "Write a poem" gives different outputs from "Write a short haiku about loneliness in a big city, in the style of Rainer Maria Rilke." The second prompt has more coordinates. The model knows more precisely where to land.
A prompt is a map, not a command. You show the model where to go. You cannot force it. The more precise the map, the closer it gets to the destination.
The network
in action
Every word of your prompt activates different patterns in the neural network. Watch how signals flow through the layers. Click the network or activate one of the prompt types below to see how the activation changes.
The 5 parts
of a strong prompt
Most bad prompts fail not because of the idea but because of the structure. A good prompt has five elements: you do not always need all of them, but you should know which ones you are leaving out.
A role gives the model a starting point and a tone. Without a role it answers as a generic assistant. With a role it gets context, expertise and a voice.
→ "You are an experienced graphic designer focusing on queer zines from Berlin."Background information the model needs to answer relevantly. What do you already know? What is the purpose? Who is the output for?
→ "I am making a zine about AI and queer identity for a Berlin underground festival."This is the most important part. Be specific. Verb + object + goal is the minimum. Vague tasks = vague outputs.
→ "Write five title ideas for the zine: short, provocative, in English."List, paragraphs, JSON, table, bullet points, Markdown: say it explicitly. Otherwise the model chooses for itself, and it often misses.
→ "Answer as a numbered list. Maximum 8 words per title."Direct, academic, funny, provocative, empathetic, formal: tone changes everything. If you do not specify it, you get the most average tone possible.
→ "Tone: direct, un-academic, a bit cheeky. No corporate speak."Weak vs. strong.
The difference is in the detail.
Here you can see concretely what happens when you apply the five elements, and what happens when you do not.
| Weak prompt | Strong prompt | What is missing / what is gained | |
|---|---|---|---|
| Weak | "Explain AI." | "Explain in 3 paragraphs how large language models work, for someone without a tech background. No jargon. Use everyday comparisons." | Audience, format, tone, constraints |
| Weak | "Write something creative." | "Write a two-stanza poem about the feeling of leaving Berghain at 6 in the morning. Tone: exhausted but happy. In English, no rhymes." | Context, length, tone, language, style |
| Weak | "Help me with my CV." | "You are an experienced careers adviser. I am applying for a UX designer role at an NGO in Berlin. Here is my current CV: [text]. Improve the wording: clear, not inflated, max. 1 page." | Role, context, input, constraint |
| Strong | "What is intersectionality?" | "Explain intersectionality in 4 bullet points. Then give 2 examples of how it shows up in AI systems. Write for a queer audience that knows the term but wants to go deeper." | Format, application, specific audience |
Build your
own prompt.
Fill in the fields and the prompt assembles itself. Use it directly in ChatGPT, Claude or whichever tool you like. You do not have to fill in every field, but give it a try anyway.
Your turn.
5 exercises that actually train you.
Theory is good. Practice is better. These exercises are not for passing. They are for thinking. Open one, read the task, write your prompt. Then see what we suggest as a sample solution.
Task: The following prompt is far too vague. Rewrite it: use at least three of the five prompt elements (role, context, task, format, tone).
Original (weak)Task: You want to use AI to write a queer event text for a club night in Berlin. Write a prompt with a fitting role, the context of the event and a clear task. Goal: a short text (~100 words) for Instagram.
Task: An AI gives you the following output: long, unstructured running text. Write a prompt that covers the same topic but outputs a clear table with 3 columns: Term / Meaning / Example. Topic: "AI bias and queer identities".
Bad output example (no format specified)Task: Write the same prompt twice: once for an academic audience (factual, referenced), once for RREBELL readers (direct, un-academic, a bit cheeky). Topic: "Why should queer people understand how recommendation algorithms work?"
Version A / academic: Version B / RREBELL tone:Task: Chain prompting means using the output of a first prompt as the input for the second. Create a chain of 3 prompts that build on each other that together produce a complete piece of content. Target output: a short RREBELL article draft (title + 3 paragraphs) on an AI topic of your choice.
Prompt 1 could generate research. Prompt 2 turns it into an outline. Prompt 3 turns the outline into the finished text. In the second and third prompt, explicitly add: "Based on the following output: [OUTPUT]"
Prompt patterns
to save.
The complete prompt
For demanding tasks where output quality matters.
You are [role]. [Context]. [Task]. Answer as [format]. Tone: [tone].
Improve instead of starting over
If the first output is not right, do not start from scratch. Build on it.
Improve the previous output: make it [shorter / more direct / more concrete]. Keep [X], change [Y].
Questioning the AI critically
Make the AI question its own answer. This helps especially with complex topics.
What weaknesses or gaps does your last answer have? What might you have overlooked?
For creative work
When you need a consistent voice across several prompts.
Keep the following voice for all further answers: [description]. Confirm with "Understood."
Demanding logic
For complicated problems, "think step by step" makes a real difference.
Think step by step. Show your reasoning before you get to the answer.
Narrow it down, do not just open it up
Also say what you do NOT want: that is just as important.
Avoid [X]. Do not use [Y]. Keep the answer under [Z words]. No introduction.
It is precision.
The more you practise, the faster you learn which coordinates produce which outputs. There is no finished, only better.
← Back to the learning hub