Few-shot prompting

Few-shot prompting means putting a few input→output examples in the prompt before the real input. Models are very good at continuing a pattern. An example shows tone, format, length and edge-case handling more precisely than a paragraph of instructions.

Classify each support message as billing, bug, feature_request or other.
Answer with the label only.

<example>
Message: I was charged twice this month.
Label: billing
</example>
<example>
Message: The export button does nothing on Safari.
Label: bug
</example>
<example>
Message: Could you add dark mode?
Label: feature_request
</example>

Message: {{message}}
Label:

With chat APIs you can also give examples as earlier user/assistant turns. Tagged blocks inside one prompt, like <example> above, work just as well and are easier to manage.

#Choosing examples

• Cover the variety: different lengths, phrasing, and the tricky cases. That might be a message that looks like billing but is a bug, or the "none of these" case.
• Vary the answers. If every example is labelled bug, the model will lean towards bug.
• Make them exactly right. The model copies mistakes, formatting quirks and length as faithfully as anything else.
• Three to five is usually enough. Beyond that, adding examples helps less, while cost and the chance of over-fitting to their surface features go up.
• Pick them dynamically for wide-ranging inputs: embed your example bank and include the examples most similar to the current input (embeddings).

#When to use it

• Format and style are hard to describe but easy to show: a particular summary voice, a labeling convention, a SQL dialect.
• Classification and extraction with fuzzy boundaries between categories.
• Not when the output must match a strict schema. Use structured outputs for that, perhaps with one example to show how optional fields are filled.

Keep the examples near the instructions and above the variable input, so they're part of a stable prefix that prompt caching can reuse. Check that they help by running evals with and without them.