Few-Shot Example Standard

Version: 1.0.0
Last updated: 2026-07-16

Purpose

Use curated examples to clarify ambiguous labels, boundaries, and output structure.

Why

Examples can improve in-context task performance [1], but copied or unrepresentative examples create bias and accidental policy. Examples supplement explicit instructions and schema validation; they do not replace them.

How

Choose examples from the production distribution.
Cover one ordinary case, one boundary case, and one valid refusal or abstention.
Remove secrets and tenant data.
Keep labels and explanations consistent with the rubric.
Test the prompt with and without examples; retain only measured gains.

<examples>
  <example>
    <input>Cancel order 42; it shipped yesterday.</input>
    <output>{"action":"escalate","reason_code":"already_shipped"}</output>
  </example>
  <example>
    <input>Cancel order 43; status unavailable.</input>
    <output>{"action":"needs_review","reason_code":"missing_status"}</output>
  </example>
</examples>

When

Use when label boundaries or exact formatting remain ambiguous after clear instructions. Avoid examples for volatile facts that belong in retrieved context.

Tradeoffs

Benefit	Cost
More consistent interpretation	More tokens and maintenance
Boundary cases become explicit	Example bias
Easier output imitation	Can overfit superficial wording

Anti-Patterns

Fixed “2–3 examples” rules without evaluation.
Examples containing production PII or secrets.
Only positive/easy cases.
Contradictions between examples and written policy.

Enterprise Considerations

Version examples as evaluation-sensitive assets. Track source, consent, data classification, label owner, and review date; prohibit cross-tenant examples.

Checklist

Examples represent production and edge cases
No regulated or tenant-specific data is embedded
Examples agree with policy and schema
Incremental quality gain is measured
Example set is versioned and reviewed

References

Brown et al., “Language Models are Few-Shot Learners,” NeurIPS 2020, https://arxiv.org/abs/2005.14165

Changelog

1.0.0 (2026-07-16): Initial few-shot curation standard.