Examples demonstrate the input-output relationship you want the model to continue.
Do now: Add one good example and one edge case to a prompt, then compare the result.
The model may copy both the strength and the flaw in your example.Examples make criteria concrete. They are especially helpful when a category or style is hard to describe in rules alone.
Demonstrate, do not merely decorate
Pair a sample input with the desired output. Keep labels and structure consistent so the pattern is visible.
Use real examples only when you have permission; otherwise use representative synthetic examples that preserve the task without exposing people.
An example teaches by relationship: when input looks like this, output should look like that.
Cover variation and edges
One example can make the model overfit to accidental details. Add a contrasting example and an uncertain case when mistakes matter.
For classification, define what belongs, what does not and what should be sent to human review.
Inspect imitation
Check whether the output copied names, figures, bias or wording that should have changed. Ask whether the examples represent the people and situations likely to appear.
Update the set when the task changes; examples are part of the specification.
Examples reduce ambiguity but do not replace evaluation.