Zero-shot means a model attempts a task cold. You give it an instruction, but no worked examples and no extra training for that task. If you ask a chatbot to sort customer emails into "billing," "bug" and "praise" without showing it a single sorted email, you are using it zero-shot.
How it works
A modern model learns an enormous range of patterns during pretraining. Because it has seen so much language, it can often follow a new instruction by recognizing what you are asking for. It does not learn anything new when you prompt it. It applies what it already has.
The "shot" counts the examples you provide. Zero shots means none. Few-shot prompting means you paste a handful of examples into the prompt to show the format or style you want. Fine-tuning goes further and changes the model itself, which is covered under fine-tuning.
Why it matters to you
Zero-shot is the cheapest and fastest way to try an idea. There is no data to collect and no training run to pay for. You write a clear instruction and see what happens.
It is also a useful test of a model's real breadth. A model that handles tasks well zero-shot has generalized beyond its training, which is the point of a foundation model. Researchers use zero-shot results to compare general models against specialists built for one job.
The tradeoff is reliability. With no examples, the model has to guess your format and your edge cases. If results are inconsistent, add a few examples, tighten the instruction, or consider tuning. Check outputs with evals rather than a few lucky tries.
An example
Say you want product reviews labeled positive or negative. Zero-shot, you write: "Label this review as positive or negative." The model answers from general knowledge. If it keeps tripping on sarcasm, you move to few-shot and add three sarcastic reviews with the right labels. If that still is not enough on thousands of reviews, fine-tuning may pay off.
Zero-shot also shows up outside language. A robot control system or an image model can be called zero-shot when it handles a new environment or object without being trained on it first.
The same idea underlies prompt engineering: most of that craft is getting the best zero-shot or few-shot behavior out of a model that stays unchanged.