Few-shot / Zero-shot

A zero-shot prompt is just instructions: "Classify this email as spam or not spam." A few-shot prompt includes examples of the task: "Here are three classified examples; now classify this new one."

For most tasks in 2026, modern frontier models perform well in zero-shot mode — write good instructions and you're done. Few-shot is most valuable when: the output format is unusual or strict, the task is genuinely ambiguous and examples disambiguate it, or the model keeps making the same wrong choice and a counterexample fixes it.

The practical wisdom: start zero-shot with a clear instruction and an output format spec. If results aren't good enough, add 2-3 high-quality examples. More than 5 examples rarely helps and starts eating context budget. The quality of the examples matters far more than the quantity.