Few-shot / in-context learning

Including worked examples of input and desired output in the prompt. The model imitates the pattern without any weight updates; this imitation-from-context is called in-context learning. Two to five well-chosen examples often beat paragraphs of instructions.