Zero-shot

In machine learning, 'zero-shot' originally meant a model could classify or handle categories it had never seen during training, generalizing from related knowledge. In LLM prompting it means you just describe what you want and ask, without providing examples. You're relying entirely on what the model already knows from pretraining.

Frontier models like GPT-5 or Claude are remarkably capable zero-shot on common tasks: write a summary, fix this bug, translate this text, answer this question. Zero-shot prompting is the default mode for most AI usage. The question of when to shift to one-shot or few-shot prompting comes up when outputs are inconsistent, the format needs to be very specific, or the task is unusual enough that the model doesn't have a strong prior on what you want.

Understanding zero-shot also helps you think about model capability: a capability is truly useful at zero-shot when the model does it reliably without coaching. Many claims about model abilities should be qualified by 'zero-shot' vs. 'with a carefully engineered prompt.' A model that does something well zero-shot is more robustly capable than one that only performs well with extensive prompting.