Prompt Engineering: Fundamentals & Techniques

Prompt engineering in one sentence

Prompt engineering is the systematic design of AI inputs with the goal of getting reliably and repeatably usable results – instead of relying on lucky hits.

The difference from simply asking questions lies in the word repeatable. A good one-off prompt is a matter of luck. Prompt engineering asks: which structure reliably reaches the goal for this type of task, even on the fiftieth run and even when a colleague uses it? The basic building blocks of an input are assumed knowledge from the introductory article What Is a Prompt? – this article is about the techniques that build on that.

Why technique makes the difference

Language models are probability machines. They generate the most plausible continuation of your text. Prompt engineering exploits exactly this property: every piece of information in the prompt shifts the probabilities in a particular direction.

An example: if you write "Analyze these numbers", the most plausible continuation is a generic description. If you write "You are a controller. Name the three most striking deviations with percentage values and a hypothesis for the cause of each", the most plausible continuation is exactly that. You're not programming – you're narrowing the space. The technical basis behind this is described in How does AI work?.

The six core techniques

1. Set a role and context

The simplest and most effective technique. Assigning a role activates a particular vocabulary, argumentation pattern, and quality level in the model. Specificity is what matters: "You are a marketing expert" achieves almost nothing, "You are a performance marketer specializing in B2B SaaS with a monthly budget of €30,000" achieves a lot.

Context adds the starting situation: target audience, previous results, constraints, and what is explicitly off the table.

2. Few-shot: show instead of describe

Rather than describing style and format, you provide examples to work from. Two to five are usually enough. This is the most reliable way to hit the right tone and structure – because one example carries dozens of implicit rules that no one could fully spell out.

For comparison: zero-shot means working without examples – faster, but less precise wherever style matters. one-shot is the variant with a single example.

3. Force reasoning steps

For tasks involving logic, calculation, or multi-step inference, the result improves markedly when the model thinks first and answers second. This technique is called chain-of-thought. The classic trigger is an instruction like "Work through this step by step and justify each intermediate step before giving the answer."

With current reasoning models, which already think internally before answering, the effect is smaller – there it's more worthwhile to specify the goal of the reasoning up front: "Before answering, check in particular whether cost and timeframe contradict each other."

A practical variant is forcing an alternative: "Name two possible explanations and then choose one with justification." This prevents premature conclusions.

4. Structure instead of running text

Long prompts without structure get processed unreliably. So separate sections clearly – with headings, dividers, or markers. This matters especially when you're including your own content as well: without clear separation, the model mixes your instruction with the material.

5. Dictate the output format strictly

The more precisely you describe the target format, the less rework. Specify the type of text, length in characters or sentences, structure, and tone. If the result will be processed further by a machine, demand a fixed format like a table with defined columns or JSON with named fields – and add: "Output only the result, with no introduction and no closing comment."

6. Build in self-checking

One additional sentence at the end noticeably raises quality: "Check your draft against the requirements above and correct any deviations before answering." It gets even stricter with an explicit checklist – for example, length, forbidden words, required elements.

Also effective against invented content: "If a piece of information isn't in the supplied material, write 'not documented' instead of estimating."

System prompt and user prompt

Almost every AI tool has two levels of instructions:

The lever lies in moving recurring instructions upward to the top: everything you retype in every other prompt belongs in a permanent system prompt – for instance in your own GPT. How that works is shown in Creating Custom GPTs.

Which technique, when?

Three patterns from agency practice

Editorial quality assurance. A prompt with embedded editorial rules checks drafts against tone, sentence length, and banned phrases – and delivers a correction list instead of a new text. The trick: explicitly demand "no rewrite, only flagged spots with justification".

Briefing extraction. Call transcript or email thread goes in, structured briefing comes out, with a mandatory field for "open questions for the client". That field alone is what separates a nice summary from a document you can actually work with.

Variants under constraints. Ten ad headlines, each under 30 characters, each with a different value proposition, none with a superlative – delivered as a table with character count per line. Requiring the character count forces the model to self-check.

What prompting doesn't solve

Prompt engineering is powerful, but it's not a cure-all. No prompt can replace missing knowledge: if the model doesn't have the information, the only fix is to supply it – as a file, a knowledge base, or via web access. Even hallucinations can be reduced with good prompts, but not switched off entirely. And for tasks that require a particular model strength, the choice of model matters more than the wording; which model suits what is covered by AI Models Compared in detail.

Rule of thumb: prompting improves form, focus, and reliability. It doesn't create knowledge.

Conclusion

Prompt engineering isn't a collection of magic spells, but a handful of systematic techniques: set a role and context, work with examples, demand reasoning steps, structure cleanly, dictate the output format, and append a self-check. Master these six and you cover most of professional practice. The next step is turning good one-off prompts into reusable templates – for that, see A Practical Guide to Better Prompts.