What is zero-shot?
Zero-shot is the plainest way to ask: you give the model an instruction and nothing else, no worked examples, and let it answer from what it already learned in training. “Summarise this in one sentence.” “Classify this message as urgent or not.” If the model has seen enough of that kind of task during training, it just does it. The prompt carries the request and only the request.
It is the natural default, and often the right one. For common tasks a capable model handles zero-shot cleanly, and you have spent no examples to get there. That keeps the prompt short, which leaves more of the context window for the actual input and the answer.
When does zero-shot start to wobble?
Zero-shot leans entirely on the model’s training, so it wobbles exactly where that training is thin or your needs are specific. The answer may be correct but in the wrong format, or it may drift from one run to the next because nothing in the prompt pinned down the shape you wanted.
That is the moment to add examples and move to few-shot. The trade is simple: a zero-shot prompt is shorter and faster to write, while a few-shot prompt spends context on examples to buy consistency. Start zero-shot, and reach for examples only when the plain instruction stops being enough.