Prompting Fundamentals
The art of telling a language model what you want — surprisingly similar to human communication.
What you'll learn
- 5 Prompt Frameworks compared: RACE, CRISPE, APE, RISEN, CRAFT
- 5 proven techniques for better results
- Common mistakes and how to avoid them
- Practical examples for everyday use
Introduction
The difference between 'Write something about marketing' and a precise, structured prompt is like the difference between a vague sketch and a detailed blueprint. AI systems deliver exactly what you ask for – nothing more, nothing less. In this guide, you'll learn how to achieve consistently good results through systematic communication.
Note
Why is this important?
- AI tools are only as good as the instructions they receive
- Prompting is the 'programming language' of the AI era
- Small changes in the prompt can make big differences in the result
After this guide you'll be able to:
- Explain the 3-layer model of AI communication
- Structure prompts using the RACE framework
- Apply advanced techniques like Chain-of-Thought
- Systematically evaluate and improve results
- Recognize and avoid the most common mistakes
Interactive: Three Perspectives on Prompting
A prompt can be understood in different ways. Switch between three analogies and discover how each perspective illuminates different aspects of prompting.
Foundations
Definition
What is a Prompt?
Analogy
Scientific
Definition
What is Prompt Engineering?
Analogy
Scientific
The 3-Layer Model
When you communicate with an AI, a dialogue takes place on three levels. You control only one of them – but it's your most powerful lever.
Your specific input for the current task
Full controlHidden instructions from the provider that shape behavior
Indirect (provider choice, custom instructions)The model's training – billions of texts it has 'read'
No influenceSystem prompts are normally invisible to users in the chat. They are set by the provider and shape the assistant's behavior. Some providers disclose them: Anthropic publishes the system prompts of its Claude apps in its documentation. Different system prompts partly explain why ChatGPT 'sounds' different from Claude. But a good part of the 'personality' already comes from training: Anthropic, for example, deliberately trains character traits into Claude.
"You are a helpful, honest assistant. Your goal is to fully understand the user's intent..."Caution: AI mirrors your tone
- Confirmation Tendency: 'Wasn't that a good idea?' often leads to agreeing answers – regardless of whether the idea is good.
- Priming Effect: The first sentence of your prompt colors the entire response.
- Politeness Overhead: 'Would you be so kind...' brings no reliable benefit – what matters is that the instruction is clear and direct.
Actively request counterarguments: 'Play devil's advocate.' or 'What are the weaknesses of this idea?'
Check
Checkpoint
Can you explain this in your own words? Tick what you've got.
0 of 2 done
Prompt Frameworks Compared
There are various proven frameworks for structured prompting. Each has its strengths – choose the right one for your task.
RACE is the most versatile framework for everyday use. It covers the key aspects: who responds, what should be done, for whom, and in what format. In some sources, the E also stands for Expectation.
Best for: General tasks when you need a clear structure
Role
Define who responds – what expertise, perspective, or personality should the AI adopt?
Action
What exactly should be done? Use precise verbs like 'analyze', 'create', 'compare'.
Context
Who is the result for? What prior knowledge exists? What's the occasion?
Execute
What format should the answer be in? Length, structure, style?
Role
The role is often the most powerful and economical lever. With one or two words, you narrow down the vast possibility space of the AI to a focused area.
Explain blockchain.Technical, dry, without audience focus
As a financial journalist writing for laypeople: Explain blockchain.Understandable, with everyday examples, appropriate depth
Caution
The Method-Acting Trap
AI can take roles too seriously. Studies show: With strong identities (e.g., 'As a conspiracy theorist...') factual accuracy can suffer because the AI wants to stay 'in character'.
Use role for perspective, style, and structure
No role for pure fact questions or logic tasks
Action
The Action defines the verb – what exactly should the AI DO? The more precise the action instruction, the better the result.
Do something about marketing.Create 5 headlines for a fitness app that motivate daily exercise.Note
Strong Action Verbs
Context
Context answers: For whom? Why? With what prior knowledge? Without context, the AI delivers generic output – with context, it delivers relevance.
Write about AI.Article for the Financial Times. Target audience: Skeptical mid-market managers. Focus: Practical applications and ROI.Execute
Execute transforms good content into usable results. How should the answer be structured? What style, length, format?
Give me the answer.Structure as: Executive Summary (3 sentences), then 3 main points with one practical example each. Max. 500 words, factual tone.Try it
Complete example
You are an experienced strategy consultant, specialized in digital transformation in traditional industries.
Identify and explain the three most important AI application opportunities for medium-sized machinery manufacturers.
Presentation for a conservative board. Company with 500 employees, currently no AI usage, digitalization pressure from competitors.
Per application: Concise headline, three benefit bullets, one concrete practical example. Factual-optimistic tone, no jargon.
Another complete RACE prompt (text analysis) to copy and try is here:
CRISPE is a popular framework that offers maximum control over tone and style. Ideal for creative experiments and A/B tests. Note: In the widely shared original (usually attributed to Matt Nigh), C and R together stand for Capacity and Role, and S (Statement) is the actual task. I show a variation here, with a separate Request step and Statement as the format instruction.
Best for: Creative tasks, marketing campaigns, style experiments
Capacity (Role)
What expertise or role should the AI assume?
Request
What exactly is being requested? The specific task.
Insight
What background information or data is relevant?
Statement
How should the answer be structured?
Personality
What tone, what style should be used?
Experiment
Variations to test or alternative approaches.
Capacity (Role)
Capacity defines not just the role, but also the level of expertise. The more specific, the better.
Write an advertising text.Generic, without recognizable signature
You are an award-winning copywriter with 20 years of experience at leading agencies.Professional, with creative flair and strategic thinking
Request
The Request is the heart – here you describe exactly what you need. Be as concrete as possible.
Do something with advertising.Write three different taglines for a new vegan burger chain that appeal to different target audiences.Insight
Insights are the secret weapon for relevant results. The more context you provide, the more targeted the response.
Note
Insight Tip
Share market research data, customer feedback, or competitive analyses. The AI can use this information to create tailored content.
Statement
Statement defines the structure of the output. Do you want bullets, flowing text, or a table? This is where you specify it.
Personality
Personality is the character of the response. It determines how the message comes across.
Note
Example Tones
Experiment
The Experiment element makes CRISPE particularly valuable for A/B tests. You can request different variants in one prompt.
Note
A/B Testing with CRISPE
Explicitly request variants: 'Create Version A (conservative) and Version B (bold)'. This way you get testable alternatives in one go.
Try it
Complete example
You are an award-winning copywriter with 20 years of experience at leading agencies.
Write three different taglines for a new vegan burger chain.
Our target audience: Millennials, environmentally conscious, but no feeling of deprivation. Competitors focus on 'sacrifice', we want to focus on 'enjoyment'.
Present each tagline with 1-2 sentences about the intended effect and target audience appeal.
Cheeky, modern, with a wink – but never preachy or moralizing.
Additionally create a 'safe' and a 'provocative' version for A/B testing.
APE is beginner-friendly and suitable for most everyday tasks. Focus on audience, purpose, and execution – simple but effective. Other sources expand APE as Action, Purpose, Expectation.
Best for: Beginners, quick prompts, content creation
Audience
Who will read or use the result?
Purpose
What should be achieved? What's the goal of the content?
Execution
How should the task be concretely implemented?
Audience
The Audience is the key to relevant content. The more precisely you describe your target group, the more suitable the response.
Write a text about online banking.Generic, doesn't really hit anyone
For tech-savvy seniors over 65 who already use a smartphone but have never tried online banking.Targeted, with appropriate approach and relevant concerns
Note
Questions About the Audience
- What prior knowledge does the target group have?
- What concerns or objections might they have?
- What language and tone suit them?
Purpose
Purpose answers the question: What should happen after someone reads the content? Define a clear goal.
Write about online banking.To convince skeptical seniors that online banking is safe and convenient – with the goal of them downloading the app.Execution
Execution defines the 'how': format, length, structure, and tonality. Here you turn an idea into a concrete result.
Note
Format Tip
Be specific: 'As a friendly letter with max. 200 words' is better than 'As a text'. State length, structure, and tone explicitly.
Try it
Complete example
For tech-savvy seniors over 65 who already use a smartphone but have never tried online banking.
To convince them that online banking is safe and convenient – with the goal of them downloading the banking app.
As a friendly letter (max. 200 words) with 3 simple steps to get started. Tone: trustworthy, not condescending.
RISEN extends classic frameworks with creativity and innovation. It emphasizes step-by-step approaches and clear expectations. Note: RISEN usually stands for Role, Instructions, Steps, End Goal, Narrowing – here I show a creative variant with Input, Expectation, and Novelty.
Best for: Innovative solutions, complex projects, when originality matters
Role
What expertise does the AI bring?
Input
What information, data, or materials are provided?
Steps
What specific steps should the AI perform?
Expectation
What exactly is expected as output?
Novelty
How innovative or unconventional should the thinking be?
Role
The role in RISEN particularly emphasizes innovation competence. Choose experts known for unconventional thinking.
Help me with new product ideas.You are an Innovation Consultant, specialized in disruptive business models in traditional industries.Input
The Input block provides the AI with all relevant data for informed analyses. Structure your information clearly.
Note
Input Structuring
Divide inputs into categories: Current state, market data, resources, constraints. The more structured the input, the more precise the analysis.
Steps
Steps are the roadmap for complex tasks. Number the steps and be specific about what should happen in each step.
Analyze and develop ideas.1. Analyze current trends in the furniture industry, 2. Identify unused niches, 3. Develop 3 concepts with reasoning.Expectation
Expectation defines the format and depth of the response. Describe exactly what the result should look like.
Novelty
The Novelty element distinguishes this RISEN variant from other frameworks. Here you give the AI 'permission' to be creative.
Note
Fostering Innovation
Use triggers like 'Blue Ocean', 'First Principles', 'What would be the Tesla moment for...?' to stimulate creative thinking. The AI becomes bolder when you explicitly ask for unconventional ideas.
Try it
Complete example
You are an Innovation Consultant, specialized in disruptive business models in traditional industries.
Our company produces classic wooden furniture. Revenue has stagnated for 3 years. 500 employees, own carpentry shop, no online sales.
1. Analyze current trends in the furniture industry, 2. Identify unused niches, 3. Develop 3 concrete concepts with reasoning.
Per concept: 100-word description, market potential rating (1-5), estimated implementation time.
Think radically new – what would be a 'Tesla moment' for wooden furniture? Ignore industry boundaries.
CRAFT focuses on task clarity with explicit tone and format definition. It structures roles and output formats very precisely.
Best for: Structured documents, formal communication, when format matters
Context
What's the situation? What background is relevant?
Role
From what perspective should the writing be?
Action
What specifically should be created or done?
Format
How should the result be structured?
Tone
What language style and atmosphere?
Context
Context in CRAFT sets the stage. Describe the situation, the occasion, and all relevant background information.
Write a press release.We're launching 'EcoSmart' next week – an AI-controlled thermostat for old buildings. First product of a CleanTech startup.Role
The role in CRAFT defines the professional perspective. Choose a role relevant to formal communication.
Action
The Action in CRAFT is particularly document-oriented. Be specific about the document type.
Write something for the press.Write a press release for trade media in the energy and sustainability sector.Format
Format is the heart of CRAFT. Here you define structure, length, and layout precisely.
Note
Format Templates
- Press release: Headline, lead (2 sentences), 3 paragraphs, boilerplate
- Executive summary: 5 bullets, max. 200 words
- Blog article: Introduction, 3-5 sections with H2, conclusion with CTA
Tone
Tone defines the character of the communication. For formal documents, the right tonality is crucial.
Note
Tones for Business Communication
Try it
Complete example
We're launching 'EcoSmart' next week – an AI-controlled thermostat specifically for old buildings. First product of a CleanTech startup with 15 employees.
As Head of Communications of an up-and-coming CleanTech startup.
Write a press release for trade media in the energy and sustainability sector.
Classic PR format: Headline, lead (2 sentences), 3 paragraphs, boilerplate with company profile. Max. 400 words.
Professional, innovative, but not over the top – facts before hype. Confident without arrogance.
Which Framework Fits You?
Each framework has its strengths. This overview helps you decide:
Note
My Tip
Start with RACE – it's easy to learn and covers most use cases. Try other frameworks when you have special requirements.
Check
Checkpoint
Can you explain this in your own words? Tick what you've got.
0 of 2 done
Advanced Techniques
RACE structures the prompt – but sometimes the AI needs help with thinking. The following techniques force the model to work more structuredly and deliver better results.
Chain-of-Thought (Step-by-Step)
Chain-of-Thought is the best-known technique for complex tasks. You get the AI to make its thinking visible – this prevents careless errors and often improves the quality of answers. That mainly applies to models without a built-in thinking mode, though. Many current models, for example in ChatGPT and Claude, now think step by step on their own before they answer. There, the instruction adds little and mostly costs time; at most I write “Think it through carefully.”
Definition
Chain-of-Thought (Step-by-Step)
Analogy
Scientific
When to use?
- Mathematical calculations and logical problems
- Multi-step decisions with dependencies
- Analysis of complex situations with many factors
Vague vs. Precise
What is 17 × 24?Calculate step by step: What is 17 × 24?Note
Trigger Phrases
- "Think step by step."
- "Justify your answer step by step."
- "Work through the problem systematically."
Note
Complete Example
You are a financial advisor.
Analyze step by step:
1. What factors are relevant?
2. How are they connected?
3. What is the conclusion?
Question: Should a 35-year-old with $50,000 in savings invest in ETFs or individual stocks?Few-Shot Prompting
Sometimes it's easier to show examples than to explain. Few-Shot uses 1-5 concrete examples so the AI recognizes the desired pattern and reproduces it.
Definition
Few-Shot Prompting
Analogy
Scientific
When to use?
- Format is hard to describe (tone, style, structure)
- Classification tasks with clear categories
- Consistent output across many similar tasks
Example: Sentiment Classification
Classify the tone:
Text: "I absolutely loved it!" →Tone: Positive
Text: "It was okay." →Tone: Neutral
Text: "That was terrible." →Tone: Negative
Text: "Interesting idea, but..." →Tone: ?
Note
How many examples?
1-3 examples are enough for simple patterns. For complex formats or when the model doesn't match the pattern, increase to 4-5. In everyday use, more than 5 examples rarely provide benefits but cost tokens; only for difficult specialized tasks can dozens or hundreds of examples sometimes still help.
Caution
Anchoring Bias
Models can cling to superficial patterns in the examples. If all positive examples are short, the model might learn 'short = positive'. Use diverse examples that cover different cases.
Self-Critique
The AI can evaluate and improve its own work. By asking it to critically review its answer, you find weaknesses and increase quality. It works best with concrete criteria to check against (facts, tests, requirements). With no new information at all, a model often misses its own reasoning errors and sometimes even turns a correct answer into a wrong one.
Definition
Self-Critique
Analogy
Scientific
When to use?
- Quality control for important texts
- Argument verification before decisions
- Iterative improvement of drafts
Note
Example Prompt
Review your answer:
- What are the 3 weakest points?
- What counterarguments are there?
- How would you improve it?Note
Alternative Formulations
- "Play devil's advocate: What speaks against this solution?"
- "Rate your answer on a scale of 1-10 and explain."
- "What would a critic find fault with in this answer?"
Self-Consistency (Multiple Paths)
For critical decisions, it's worth solving the task in multiple ways. If all paths lead to the same result, it's probably correct.
Definition
Self-Consistency (Multiple Paths)
Analogy
Scientific
When to use?
- Critical decisions with high risk
- Complex calculations for error checking
- When you're unsure about the first answer
Note
Example Prompt
Solve this task in three different ways.
Compare the results:
- Where do they agree?
- Where do they differ?
- What explains the differences?Note
Practical Tip
Especially effective for mathematical problems: Solve it algebraically, numerically, and through estimation. If all three agree, the result is very likely correct. It gets even more independent if you ask the same question several times in new chats and take the most frequent answer – that is how the original method works.
Least-to-Most Prompting
For very complex problems, you can ask the AI to first create a plan: 'Break down this task into individual steps. Don't solve anything yet.' Then: 'Now solve each step one by one.' This prevents the model from getting tangled in its own train of thought.
Step-Back Prompting
For knowledge questions with hallucination risk, a step back helps: 'What general principles are relevant for this problem?' Then: 'Use these principles for the following question...' The model 'remembers' the basics before it answers.
Note
The Creativity Dial: Temperature
Temperature controls how 'bold' the model is with word choices. Technically, it affects the probability distribution during token selection: low values make likely words even more likely, while high values give rarer words more chances.
- Low (0.1-0.3): Consistent, focused – ideal for analysis and facts
- Medium (0.5-0.7): Balanced – suitable for most tasks
- High (0.8-1.0): Creative, surprising – ideal for brainstorming
Not all tools offer temperature settings. ChatGPT and Claude use optimized default values. The exact scale (e.g., 0-1 or 0-2) varies by model. With many of the newest models from Anthropic, OpenAI and Google (as of November 2026), temperature can no longer be adjusted via the API at all – there you steer style and creativity through the prompt.
When to use which technique?
Check
Checkpoint
Can you explain this in your own words? Tick what you've got.
0 of 2 done
Practical Examples
The best way to learn prompting is through practice. Here I show you typical use cases and how to apply the techniques you've learned.
Application Areas Overview
Data Analysis
Interpret sales figures, identify trends, summarize reports
Chain-of-Thought + clear metricsCreative Writing
Product descriptions, social media posts, storytelling
Few-Shot + style guidelinesLearning & Understanding
Get complex concepts explained, exam preparation
RACE + audience contextProgramming
Code explanations, debugging, refactoring
Structured tags + concrete codeBusiness Communication
Drafting emails, structuring presentations
Role specification + tone definitionTranslation & Adaptation
Translate texts, cultural adaptation, simplification
Context + target audience definitionExample 1: Email Prioritization
You're overwhelmed with 50 unread emails. You need an assistant to help filter out the most important ones.
Note
Structured Prompt (RACE Format)
Note
Why does this prompt work well?
- Clear role gives the model a perspective for prioritization
- The context (Project Alpha, 2h until meeting) enables relevant decisions
- The Execute format prevents all 50 emails from being listed individually
Example 2: Product Descriptions
You need engaging product texts for an online shop. Instead of describing the format at length, you simply show examples.
Note
Few-Shot Prompt
Write product descriptions in the following style:
Note
Why does this prompt work well?
- The examples define length, tone, and structure without explicit rules
- The style is recognized by the model: emotional, feature-focused, with a wink
- No long explanations needed – the model learns from the pattern
Troubleshooting
The 4 Classic Pitfalls
Too vague
AI guesses what might be meant
Apply RACE framework
Too much at once
Contradictory or mediocre results
Split task into multiple prompts
Anthropomorphizing
Politeness phrases with no effect
Formulate clear, direct commands
Negation
'Don't' is easily overlooked
Formulate positively: what should be included
Caution
Why Positive Framing Is More Robust
Older language models had real difficulties with negation — 'don't' was often ignored. Modern models handle negation much more reliably. Still, positive framing remains the more robust strategy: it gives the model a clear direction rather than just boundaries.
Write about luxury watches, but DON'T mention the price.Write about luxury watches. Focus on craftsmanship, materials, and status.Quality Checklist
- Relevance: Does the answer exactly fulfill the task?
- Completeness: Were all instructions considered?
- Correctness: Are facts verifiable? (Always validate externally!)
- Format: Does the output match the format specifications?
- Tone: Does the answer hit the defined role/perspective?
Summary
Key Takeaways
- Quality Input = Quality Output – Precise prompts deliver precise results
- RACE as checklist – Role, Action, Context, Execute structure every prompt
- Techniques for complex tasks – Chain-of-Thought, Few-Shot, Self-Critique expand your repertoire
- Iteration is normal – The first prompt is rarely perfect
- Positive over negative – Describe what you want, not what you don't want
- Always verify facts – AI can be convincingly wrong
Frameworks
Techniques
| Chain-of-Thought | "Think step by step." |
| Few-Shot | [Show examples] |
| Self-Critique | "Review your answer." |
| Self-Consistency | "Solve in 3 different ways." |
Quiz
Answer the following questions to check your understanding. With at least 70% correct answers, you'll earn an achievement!
Check
Test Your Knowledge
Which RACE component defines WHO responds?
Select one answer
1. Which RACE component defines WHO responds?
2. What is Self-Consistency particularly useful for?
3. What's the problem with the prompt 'Don't mention prices'?
4. Chain-of-Thought is particularly useful for:
5. What does 'method-acting trap' mean for role assignment?
6. Which statement about the 3-layer model is correct?
7. What is Few-Shot Prompting?
8. What should be your FIRST step when the AI response doesn't meet your expectations?
9. What is Tag Annotation particularly useful for?
10. Which prompting best practice is CORRECT?
11. You receive a long, generic answer with no specific connection to your topic. What's probably missing from your prompt?
12. You need 5 product names for a new fitness app. The AI only delivers 3 and then goes off-topic. What do you do?
13. This prompt delivers inconsistent results: 'Write something cool about sustainability.' Which RACE elements are missing?