Few-Shot Prompting

Reviewed by Proof · Updated

Definition

Few-shot prompting means showing a model a few worked examples in the prompt so it copies the pattern for a new input.

Key takeaways

  • You put two to five example inputs and outputs before the real input.
  • The model picks up the format, style and labels from the examples.
  • It needs no training. The examples live only in that request.
  • Every example adds input tokens, so keep them short.
  • The idea was made widely known by the 2020 GPT-3 paper "Language Models are Few-Shot Learners" (Brown and others).

How to do it

In a chat API, the cleanest way is to write each example as a pair of messages: a user message with the example input, then an assistant message with the ideal output. Then add the real input as the last user message.

Tips:

  • Make examples match the real task. Same length, same kind of input.
  • Cover the edge cases. If some inputs should get "unknown", include one.
  • Vary them. If every example has the same answer, the model may just repeat it.
  • Keep the format exact. The model copies small details, like punctuation and capital letters.

If the model does well with no examples at all, you may not need them. See Zero-Shot Prompting.

Example

This teaches a simple sentiment label with two examples.

curl https://api.luv13.ai/v1/chat/completions \
  -H "Authorization: Bearer $LUV13_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "luv13/glm-5.3-flash",
    "messages": [
      {"role": "system", "content": "Label each review as positive, negative or mixed. Reply with the label only."},
      {"role": "user", "content": "Fast shipping and it works great."},
      {"role": "assistant", "content": "positive"},
      {"role": "user", "content": "Nice screen, but the battery dies by noon."},
      {"role": "assistant", "content": "mixed"},
      {"role": "user", "content": "It broke on day two."}
    ]
  }'