What is the difference between zero-shot and few-shot prompting?
Zero-shot gives only the instruction. Few-shot adds a handful of worked input and output examples in the prompt so the model copies the format and style. Few-shot often improves consistency without any training.
Explain it simply
Telling a new intern "classify these emails" is zero-shot. Showing three emails with the label you want for each, and then handing over a fourth, is few-shot.
Examples show things instructions struggle to say: exact label names, tone, length, edge cases. The model spots the pattern in the prompt and continues it, which is called in-context learning. No weights are updated, and the effect lasts only for that call.
Choose examples that are varied, correct and close to real inputs. Bad or one-sided examples mislead the model, for example if all your examples are "positive", it may lean positive. Examples also use tokens, so a long list adds cost and latency.
Example
Classify the message as BILLING, DELIVERY or OTHER.
Message: "Charged twice for my order" -> BILLING
Message: "Parcel hasn't arrived in 6 days" -> DELIVERY
Message: "Do you have a store in Pune?" -> OTHER
Message: "Refund not received yet" ->What to say in the interview
"Zero-shot relies on the instruction alone; few-shot adds examples in the prompt to fix format, labels and tone through in-context learning. I pick diverse, correct examples and check cost, since examples add tokens."
Common mistakes
- Using examples that all share the same label.
- Confusing few-shot prompting with fine-tuning.