Zero-shot / few-shot

Zero-shot prompting is what most users do most of the time: describe the task, ask for the output, done. Modern models handle a huge range of tasks zero-shot because their training exposed them to countless instances of similar tasks. "Summarize this article" doesn't need an example — the model has read the whole internet.

Few-shot prompting adds 2-5 worked examples inside the prompt itself, showing the model exactly the input-output pattern you want. It's the right tool for tasks where the output shape is specific or unusual — a particular table format, an unusual classification scheme, a house style the model can't infer from generic instructions.

The practical rule: try zero-shot first, and only add examples if you find yourself repeatedly correcting the output shape. Every example you add uses tokens and can sometimes narrow the model's creativity unhelpfully. Well-worded zero-shot instructions often outperform mediocre few-shot examples.

Related concepts: chain-of-thought (adding "think step by step" before the final answer, without examples) and in-context learning (the broader category of "showing the model examples in the prompt itself, without any fine-tuning").