Prompting
Few-shot prompting
Few-shot prompting supplies a handful of worked examples in the prompt so the model infers the pattern rather than being told it.
Also written: in-context learning
Two or three well-chosen examples routinely outperform several paragraphs of instruction, particularly for output format compliance. Showing the shape is more reliable than describing it.
They are also the most expensive part of a prompt per unit of instruction, and they are resent on every request — which makes the trade worth making deliberately.
In practice
One example gets copied as a template. Two or more communicate a pattern. The example that earns its tokens is a hard case, not a third easy one — and if an example and an instruction disagree, the example wins silently, which is why a stale example is worse than a missing one.
Common questions
How many examples should I include?
Two to five for most tasks. One gets copied as a template rather than read as a pattern. Past five, returns fall off quickly while every request keeps paying for the tokens.
What makes a good example?
A hard case, formatted exactly as the instructions describe. Easy examples confirm the model already knew what to do, and where an example contradicts an instruction, the example wins silently.