Prompt Engineering: A Practical Guide to Writing Effective Prompts

What prompt engineering is and why it matters

A large language model predicts the most likely next word based on your input. The prompt is therefore the main control: it determines which slice of learned knowledge the model activates and in what form it answers. Prompt engineering is the systematic practice of shaping those inputs so the output becomes useful, consistent and verifiable.

The payoff is concrete: depending on wording, the same question can yield a vague general answer or a directly usable work product. Good prompting is therefore not niche expertise but a core skill for anyone working with tools such as ChatGPT, Claude or Gemini.

The anatomy of an effective prompt

Most strong prompts can be broken into six building blocks. Not all are needed every time, but the more demanding the task, the more of them pay off.

• Role: assign the model a perspective, e.g. 'You are an experienced Swiss tax advisor.' This sharpens vocabulary and level of detail.
• Context: supply background, audience and purpose. 'For new clients of an SME who have never worked with AI.'
• Task: state the goal as a clear verb. 'Summarise', 'compare', 'draft' rather than a vague question.
• Examples: show one or two samples of the desired output (few-shot). Examples often work better than explanations.
• Constraints: set limits on length, tone, language or what to avoid. 'Max 150 words, neutral tone, no jargon.'
• Output format: request a concrete structure, e.g. table, bullet list, JSON or a fixed section layout. This makes answers easy to reuse.

Proven patterns for stronger prompts

Beyond the building blocks, a few patterns have become established and demonstrably help depending on the task.

• Zero-shot vs. few-shot: no example (zero-shot) often suffices for simple tasks; two or three examples (few-shot) markedly improve consistency for formats and classification.
• Step-by-step reasoning (chain-of-thought): ask the model to structure its reasoning before giving the answer. This helps with logic, arithmetic and multi-step tasks.
• System and role prompt: many tools let you set a persistent system prompt that fixes tone and rules for the whole session, saving repetition.
• Decompose the task: instead of one giant prompt, run several smaller steps in sequence. Intermediate results can be checked and corrected.
• Provide a reference text: for factual answers, paste the source into the prompt and instruct the model to answer only from it. This principle also underlies retrieval-augmented generation (RAG).

Common mistakes and how to avoid them

• Too vague: 'Write something about marketing' yields anything. State audience, purpose, scope and format.
• Multiple goals in one sentence: mixed tasks produce muddled answers. Separate subtasks clearly or number them.
• No output format: without a spec you get prose one time, a list the next. Ask for structure explicitly.
• Taking hallucinations as fact: models can invent convincing falsehoods. Ask for sources and verify key facts independently.
• Overrating negations: 'Don't write too long' is weaker than 'max 120 words.' Phrase limits positively and measurably.
• Pasting sensitive data: personal data or trade secrets belong in a prompt only if the legal basis and provider terms allow it.

A practical workflow: iterate toward a good prompt

Treat prompting as a loop, not a one-shot attempt. A reliable routine looks like this:

• 1. Write a base prompt with role, task and format. 2. Review the output critically: what is missing or wrong? 3. Change exactly one thing per round so you can see its effect. 4. Save working versions as templates and reuse them.
• Use the model as a partner: have it critique your prompt or ask clarifying questions before answering ('Ask me if anything is unclear').

Data protection and the Swiss context

Prompts are inputs to an external service. In Switzerland the revised Data Protection Act (revDSG/nFADP) applies, supervised by the FDPIC (EDÖB). Anyone entering personal data into an AI tool needs a legal basis, must know the processing purpose, and should check whether inputs are used for training or transferred abroad.

Practical rule of thumb: anonymise or pseudonymise data before entry, prefer business accounts with a contractual no-training guarantee where possible, and document internal prompting guidelines. That keeps prompt engineering both productive and legally sound.