adjective · AI technique
Few-shot
also called in-context examples
Showing before telling. You give the AI a few solved examples first, and it copies the pattern for your new one.
Prompting a model with a small number of worked examples inside the prompt so it follows their pattern on a new input.
The prompt includes a handful of input and output pairs before the real input. The model continues the pattern it sees in its context, so the examples steer format and style without changing any weights, and they vanish when the request ends.
Real-life example
You want product names turned into short slogans in a particular voice. A plain instruction gives uneven results. You add three product-and-slogan pairs ahead of the new product, and the slogans now match the voice. The model learned nothing permanent, it just followed the pattern in front of it.
Where the analogy breaks
-
It does not teach the model anything lasting. The examples only exist inside that request, and the next request starts clean.
-
The model copies quirks along with the pattern. If every example is short, or one uses an odd word, the answers may copy that too.
-
Examples cost tokens on every request. A few good ones beat many mediocre ones, and they use up room in the context window.