Few-shot prompting

You place several input-and-output pairs in the prompt before the real input. The model has learned during training to continue patterns, so a consistent set of examples steers format, tone and edge-case handling far more precisely than a description usually can. Nothing is trained or saved — the examples work only because they are in the context. Improvement is steepest over the first two or three examples and usually flattens after about five, so the cost of a sixth is nearly always higher than its benefit.

This is why it is also called in-context learning. The model has not learned anything permanent; it has been shown, within this one request, what good looks like.

Use it when the output shape is easier to demonstrate than to describe. Classification with fuzzy category boundaries, a specific writing voice, structured extraction with awkward edge cases — all are quicker to fix with three good examples than with three paragraphs of rules.

It is also the step to try before fine-tuning. Fine-tuning costs money, time and a dataset; few-shot costs tokens and can be changed in seconds. Reach for it first and only move on if it genuinely is not enough.

Unbalanced examples. If four of your five examples have the same label, the model picks that up as a hint about likely answers and skews towards it. Cover your categories evenly, and put genuinely hard cases in rather than easy ones.

The second is adding more examples than help. Gains usually flatten after a handful, while every example costs tokens on every single request — turning a prompt into a dataset makes it expensive and slow without making it better.