Prompt Engineering
Updated 11.08.20261 min
Prompt engineering is writing requests so a model does the right thing. Not magic phrasings but ordinary precision: a clear task, the context it needs, a defined output format, and a criterion for judging the result.
What actually works
First, specificity. "Write about bitcoin" and "explain the halving to someone hearing about crypto for the first time, in three paragraphs, without jargon" produce answers of entirely different quality.
Second, examples. One or two samples of the wanted answer convey a format better than a paragraph describing it. Models copy the structure of what they are shown.
Third, the material. If the answer must rest on a specific text, provide it. Asking a model to recall a document by its title is the fastest route to invention.
Techniques that pay off
- Name the role and the audience"Explain as an editor would to a newcomer" sets the level of detail, the tone, and what may be left out.
- Define the format up frontNumber of points, length, headings, output language. Otherwise you get whatever the model considers default.
- Allow "I don't know"Without that permission the model fills the gap with a guess, because continuing the text is what it must do.
- Ask for reasoning before the answerOn counting and logic tasks this measurably improves accuracy: the model stops jumping straight to a conclusion.
- Iterate instead of writing one long promptA short request, a look at the result, then a correction beats trying to anticipate everything in advance.