Zero-shot means prompting an LLM to perform a task without providing any examples — relying entirely on the model's training. Few-shot means providing 2–5 input/output examples in the prompt before the actual task, giving the model a format to follow.
Zero-shot is simpler and uses fewer tokens; few-shot often produces more consistent output on structured tasks like classification, extraction, or formatting. The choice between them is a practical engineering decision: test both with your specific task and model, measure output quality, and choose accordingly. Many production systems combine both — a zero-shot instruction with a few worked examples embedded in the system prompt.