Zero-shot

Zero-shot means asking a model to perform a task using only an instruction, with no worked examples included in the prompt.

Think of a new cook on their first day. "Make the dal less spicy" with nothing else is zero-shot — they have to rely on general experience. Showing them one bowl you liked is one-shot. Showing them three is few-shot. The cook's skill has not changed in any of these; what changed is how much of your particular taste you handed over.

Zero-shot :  "Classify this review as positive or negative: ..."
One-shot  :  "Example: 'battery dies fast' → negative.  Now classify: ..."
Few-shot  :  three or four examples, then the real one

Older models needed examples for almost everything. Instruction-tuned models handle a large share of ordinary tasks zero-shot, which is why most prompts you write today start there.

When to add examples anyway#

Examples earn their context window cost in three situations, and it is worth recognising them rather than guessing.

The output format has to be exact — a specific JSON shape, a fixed set of fields, no surrounding commentary. Two examples pin this down far better than a paragraph of description. The labels are yours, not the world's, such as sorting tickets into your company's own 12 categories, where the category names carry meaning only your team knows. Or the domain has its own conventions, where a couple of examples show a style faster than any explanation.

Beyond that, if you find yourself needing twenty examples in every prompt, the examples have stopped being a prompt and started being a training set — that is the point where fine-tuning becomes cheaper per request and more reliable.

One related use of the term: "zero-shot classification" also names a specific technique where a model scores text against label names it was never trained on, which lets you add a new category without collecting a single labelled example for it.

Where to go next#

• Full lesson: Prompt engineering
• Related terms: llm, fine-tuning, context-window, agent