Describing exactly what you want is hard. Showing it is easy. Few-shot prompting means including a few examples of input and desired output in the prompt so the model can copy the pattern.

The technique became famous with the 2020 GPT-3 paper, Language Models are Few-Shot Learners, which showed that large models could perform new tasks from just a few examples, without any retraining.

Zero-shot vs. few-shot

  • Zero-shot: “Classify the sentiment of this review.”
  • Few-shot: the same instruction, followed by three labelled reviews, then the new one.

Few-shot prompts produce more consistent formats and handle edge cases better.

A template

Classify each customer message as BILLING, TECHNICAL or OTHER.

Message: “I was charged twice this month.” → BILLING Message: “The app crashes when I upload a photo.” → TECHNICAL Message: “Do you have an office in Berlin?” → OTHER

Message: “[new message]” →

Rules for good examples

  1. Cover the variety you expect, including tricky cases.
  2. Balance the labels — if most examples share one answer, the model leans toward it.
  3. Keep the format identical across examples.
  4. Use realistic inputs, not idealized ones.
  5. Three to five examples are usually enough; more costs tokens without always helping.

When to use it

  • Extracting fields into a fixed structure
  • Matching a brand’s tone of voice
  • Classification and tagging
  • Converting between formats

When it is not enough

If you need hundreds of examples to get reliable results, consider fine-tuning a model instead, or combining prompts with validation code that checks the output format.

A worked example: classifying customer feedback

Suppose you want to sort feedback into categories. A few-shot prompt might look like this:

Classify each message as Billing, Bug, Feature request or Praise. Answer with the category only.

Message: “I was charged twice this month.” → Billing Message: “The app crashes when I upload a photo.” → Bug Message: “Please add a dark mode.” → Feature request Message: “Your support team was fantastic!” → Praise

Message: “Export to PDF stopped working after the update.” →

The examples show the model the exact categories, the output format (a single label) and the boundary between similar cases — here, a broken feature is a Bug, not a Feature request.

Choosing examples well

  • Cover each category at least once.
  • Include tricky cases near category boundaries.
  • Vary length and style so the model doesn’t copy superficial features.
  • Keep the order mixed so the model doesn’t learn to favor the last label.
  • Test on new examples you didn’t include in the prompt to check it generalizes.

The research behind it

The term “few-shot learning” in language models became widely known with OpenAI’s 2020 GPT-3 paper, Language Models are Few-Shot Learners, which showed that large models could perform new tasks from a handful of examples in the prompt, without updating their weights. This ability — often called in-context learning — is now a basic tool of prompt design.

Common mistakes

  • Using examples that are all very similar, so the model can’t generalize.
  • Including examples with errors, which the model will copy.
  • Mixing formats between examples.
  • Forgetting to show the “edge case” that matters most to you.

Few-shot plus instructions

Examples work best combined with a short instruction stating the task and output format. The instruction tells the model what to do; the examples show how.

What to watch next

  • Automatic selection of the best examples for each query.
  • Long-context models that can use many more examples.
  • Comparisons between few-shot prompting and fine-tuning for production tasks.
  • Structured output features that reduce the need for format examples.

Key takeaways

  • Few-shot prompting shows the model examples of the input and desired output.
  • Examples teach format, categories and boundaries better than descriptions alone.
  • Choose diverse examples, including tricky cases, and test on new inputs.
  • GPT-3 popularized the idea of learning tasks from examples in the prompt.
  • For large-scale or highly specific tasks, fine-tuning may work better.

Frequently asked questions

How many examples should I use?

Often three to five are enough. Add more for complex tasks, watching cost and context length.

What is zero-shot prompting?

Asking the model to do a task with instructions only, without examples.

When should I fine-tune instead?

When you need consistent results on a narrow task at scale, or the behavior can’t be captured with a few examples.