Zero-shot vs few-shot prompting: when do you use each?

I start by thinking about how much structure the task needs. For a common task with a simple format, I use zero-shot prompting: I describe the task and give any necessary context, but no demonstrations. That is quicker to maintain and avoids spending context window space on examples. If the model already understands the labels and the expected output is easy to recognize, zero-shot is usually the cleaner choice.

I switch to few-shot prompting when the format is unusual, the labels are subtle, or I need consistent style. I put a few input and output examples directly in the prompt, usually with a clear separation between examples and the final query. I select examples that cover the hard cases, vary the inputs, and make the labels balanced. I also test whether the examples are actually changing behavior, rather than assuming that more examples always help.

I see one failure mode as surface-pattern copying. If every example has a positive label, the model may treat positive as the default, even when the input calls for another label. I would catch that with a small set of edge cases, including examples with the opposite label, and adjust the examples if needed. In short, I use zero-shot for straightforward tasks and few-shot when examples communicate the missing pattern. I also treat examples as data that needs evaluation, not as decoration.