Skip to main content

Few-Shot Examples Done Right

Intermediate

Few-shot prompting means showing the model a few worked examples of the task before asking it to do a new one. Done well, it's the fastest way to lock in a format, style, or edge-case behavior — often better than describing what you want in words.

Why examples beat descriptions​

"Be concise and friendly" is vague. Showing two concise, friendly outputs is unambiguous. The model pattern-matches the examples and continues the pattern.

A clean few-shot prompt​

Classify each support message as: billing, bug, or feature.

Message: "I was charged twice this month."
Label: billing

Message: "The app crashes when I upload a photo."
Label: bug

Message: "Can you add dark mode?"
Label: feature

Message: "My subscription renewed at the wrong price."
Label:

The model has the pattern; it completes the last line.

How to choose and format examples​

  • Cover the variety, especially the edge cases you care about. If a category is rare or tricky, include it.
  • Keep examples consistent in format — same structure, same labels, same delimiters. Inconsistency teaches inconsistency.
  • 2–5 is often enough. More helps for hard/varied tasks but costs tokens and can overfit to the examples' quirks.
  • Order can matter — put the clearest examples first; for classification, don't cluster all of one label together.
  • Use delimiters (or XML tags) to separate examples from the live input.

Zero-shot vs few-shot​

Try zero-shot (just ask) first — modern models are strong. Add examples when you need a specific format/style or the task is ambiguous. If zero-shot already nails it, don't pay for examples.

:::tip Examples are data — keep them clean A wrong or sloppy example actively teaches the wrong thing. Curate them like training data. :::

Next​