Zero-shot vs few-shot prompting: what's the difference?

The difference is examples. Zero-shot prompting gives the model an instruction with no examples — you just ask. Few-shot prompting includes a handful of worked examples in the prompt so the model can copy the pattern.

Zero-shot is simpler and cheaper; few-shot buys you more control over format and edge cases when the task is fiddly.

What do they look like side by side?

A zero-shot prompt is just "Classify this review as positive or negative: ..." A few-shot version adds two or three labeled reviews first, so the model sees exactly what output you expect before it hits the real one.

When does each one win?

Reach for zero-shot when the task is common and the instruction is clear — summarizing, rewriting, answering a straightforward question. It's less to write and costs fewer tokens. Reach for few-shot when you need a specific output shape, a consistent style, or you're handling an unusual task the model keeps getting slightly wrong.

Showing beats telling: two good examples often fix a formatting problem that paragraphs of instructions couldn't.

Do modern models still need examples?

Often not. Today's instruction-tuned models are trained to follow plain requests, so they handle many tasks well zero-shot that used to require examples. That makes zero-shot the sensible default — start there, and only add examples if the output isn't consistent enough.

The trade-off is straightforward: every example you paste in costs tokens and money on each call, so use few-shot when it earns its keep, not by habit.