Few-Shot Prompting: Teach the Format with Examples
Few-shot prompting shows a model examples of the input-output pattern you want. It is useful when describing a style or classification rule takes more words than demonstrating it. Choose examples that reveal the rule, including boundaries where a reasonable person might otherwise hesitate.
Why it matters
Examples shape expectations. If every sample contains the same topic, length and answer, the model may copy those accidental features rather than learning the distinction you care about.
A practical workflow
- Write the actual task and name the categories or output fields before adding any examples.
- Select a few varied samples that demonstrate the desired rule without including private information.
- Add an ambiguous case with an explicit explanation of why it receives its particular label.
- Test new inputs that were not used in the examples and inspect where the pattern breaks.
An example
To sort workshop feedback into logistics, teaching and equipment, include one comment about each category and one mixed comment. Define whether mixed comments receive multiple labels. Then test a new sentence about unclear instructions for connecting a laptop.
A mistake to avoid
Do not treat successful imitation as proof of factual understanding. A model can reproduce a neat format while assigning the wrong category. Examples also should not contain mistakes that you expect it to correct automatically.
Try it and check the result
Hold back several comments and label them yourself first. Compare model labels with that reference, paying particular attention to mixed and unfamiliar cases rather than only the easiest examples.
Reference and further reading
GitHub's prompt engineering guidance.
The workflows and examples in this guide are illustrative. Product controls and availability can change; check the linked official guidance for your setup.
Continue with Write Better AI Prompts with a Simple Task Brief or explore all blog topics.