Usage & Prompting
Zero-Shot
You give the model a task it was never shown in this form, without a single example, and it solves it from its pretraining. That this often works was the real surprise of large language models. Zero-shot is the fastest test of whether a task is within reach for AI: try it bare first, then refine.
An example
“Assign this complaint to one of five categories: delivery, product quality, invoice, advice, other.” Without a single example, a current model gets clear cases right reliably. With borderline cases, such as a complaint about a miscalculated follow-up delivery, it guesses, and possibly differently on every run. The categories in your company are not the ones the model knows from training.
A closer look
The task is only described, not demonstrated. That this works at all is a property of large pretrained models: the task form itself appeared in training in some shape.
Why it matters
Zero-shot is the right first attempt because it costs nothing and immediately shows whether a task is within reach at all. As a permanent solution it is only suitable where the task is general and the tolerance high. For anything with in-house categories, in-house edge cases, and an error rate someone cares about, it is the start of the work, not the result. The workable sequence: measure zero-shot, then add examples, then, if necessary, consider fine-tuning. Whoever starts at step three pays for step one.
Commonly misunderstood
Zero-shot is the right first measurement but rarely the final state. Where a fixed format matters, one or two examples are almost always the cheaper improvement.