The point of few-shot is that the model works from the sample. Two or three good input-output pairs say more about format, tone and level of detail than three paragraphs of description — and they are far harder to misread. That is what makes it the prompting technique with the fastest payback.
Its effect is biggest where the task is hard to put into words: style, structure, labelling, sorting into categories. If you have twenty earlier, approved examples, pasting two of them in usually improves more than any amount of instruction polishing.
The most common misconception is that an example only carries the format. An example carries everything it contains: if all three of your samples are short, you will get a short answer even when you did not ask for one; if all three come from the same industry, the model narrows its vocabulary to that industry. That is why examples have to be picked deliberately, not copied from the last three you happen to have.
How to choose examples, and the side effects that come with them, is covered in few-shot in practice, an article of its own. The directly related terms: prompt, system prompt and structured output.