Few-Shot vs Zero-Shot: How Many Examples Your Prompt Actually Needs

Zero-shot (just an instruction, no examples) works well when the task is common and the desired output format is unambiguous — most everyday writing, summarization, and Q&A tasks fall here, and adding examples just adds tokens for no benefit. Few-shot (instruction plus 2-5 examples) earns its cost when: - The output format is unusual or hard to describe in words (a specific structured format, a particular tone, a non-obvious pattern). - The task has edge cases that are easier to show than to explain — include one example that demonstrates the tricky case, not just the easy one. - You've tried zero-shot and the model keeps making the same category of mistake — one well-chosen example often fixes it faster than more instruction text. Diminishing returns set in fast: 2-3 well-chosen examples usually beat 10 mediocre ones, and too many examples can make the model overfit to their surface pattern instead of the underlying task. Prefer diverse examples over similar ones — three examples that all look alike teach the model less than three that cover different cases.