Few-shot examples are "user question—ideal answer" pairs embedded in the prompt to demonstrate the expected answer style, format, and reasoning pattern, achieving in-context learning without fine-tuning.