Examples Set the Boundary Better Than Paragraphs
A paragraph explains a rule, an example shows the boundary. Few-shot is especially strong where words stall: format, tone, classification, ambiguous cases. And the power of examples is not in quantity but in covering different classes of behavior.
A good set isn't a dozen similar successes but one representative per class: the ordinary answer, the clarification, the refusal, the conflict. An example shows not only the correct input but the correct reaction to an awkward one.
<examples>
<example kind="answer">
<input>Can I return an item after 20 days?</input>
<evidence>kb-returns: the window is 30 days</evidence>
<output answerability="answer" source_id="kb-returns" />
</example>
<example kind="clarify">
<input>Can I return it?</input>
<evidence>the rules depend on the category</evidence>
<output answerability="clarify" question="What item is it?" />
</example>
<example kind="escalate">
<input>The policy has two different dates.</input>
<evidence>the sources conflict</evidence>
<output answerability="escalate" reason="policy_conflict" />
</example>
</examples>To see whether all your classes are covered, mark them - the counter will show coverage.
Interactive lab 3
Few-shot set coverage
0% class coverage The set shows only the happy path.
But it is easy to overdo examples.
Don't encode the whole world in examples. A large collection eats context and can overfit behavior to accidental details. Add an example when it closes an observed failure mode, not for completeness.