Key answer
Define what an example should demonstrate, then cover differences found in real tasks. Representative does not mean prettiest, easiest or exclusively the hardest exceptions.
Full explanation
Liu et al. compared demonstration selection for GPT-3 sentiment classification, table-to-text and open-domain QA, finding semantically similar retrieval more effective than random choice. This supports task relevance, not one universal combination. Sections 2–3 Anthropic recommends realistic examples, edge cases and sufficient variation to reduce unintended patterns. Official guide
List factors that change handling, then select a few complementary cases. Customer-service classification may include ordinary questions, missing information, two requests in one sentence, negation and human judgment. Each example needs a trustworthy expected answer and key criterion.
If three cases already show one rule, another similar case may add little. A contrast where one changed condition should change the answer can clarify boundaries: eligibility to apply and an approved application must not both be described as a completed refund.
These examples teach customer-service reply rules. Check whether they cover only smooth cases and identify missing insufficient-information, ineligible or easily confused cases. Explain which boundary each proposed example clarifies; do not invent real customer records.
Separate teaching examples from acceptance cases. Explain rules in prompts, then test transfer on unseen cases. Evaluate everyday averages against the real distribution and report important exceptions separately. Easy-case scores must not conceal critical errors; all-extreme tests do not represent daily performance either.
References
1. Liu et al. (2022), What Makes Good In-Context Examples for GPT-3?, DeeLIO, Sections 2–3
2. Anthropic, Prompting best practices, Use examples, official guide