An instruction such as “keep it concise” can mean different things in different tasks. An example shows the shape of a result more directly, especially when a format or classification rule matters.
Choose examples that reflect the cases the task will actually encounter. Include an ambiguous case when it helps explain a boundary. Make sample values clearly distinguishable from facts the model should carry into the answer.
Check whether the output follows the intended rule or merely copies surface details from the example. A useful example supports a clear instruction; it should not hide the reason one result is acceptable and another is not.
- Use representative examples.
- Include a boundary case when useful.
- Check the rule rather than surface resemblance.
A small working example.
Imagine supplying an example that leaves an unknown field empty. The demonstration explains that guessing is outside the intended task.
A note to keep beside it.
Keep the origin of a claim close to the claim itself. A generated explanation and an independently checked source play different roles in understanding a result.
Follow a related question
List the assumptions behind a prediction.
A thoughtful view of what comes nextWrite the question beside the idea.
Capture the question firstKeep learning
Related background to continue exploring this subject.
Google: prompt design strategies Google: an introduction to language models
