Few-shot Prompting

Few-shot prompting puts a few worked examples in the context so the model can infer the pattern: the input format, the expected output, the level of detail, the edge cases you care about. It is a form of in-context learning, since nothing is trained; the examples only live in that one request.

It is often more reliable than a paragraph of description, and it is usually more token-hungry. A careful choice of three examples can outperform a long rule list. A poor choice of three examples can pin down a wrong pattern, including a formatting quirk you did not want.

The main failure is picking unrepresentative examples. If every example you show is clean, the model will not know what to do with a messy one, and it will imitate your tidiness rather than your actual standard.