Few-shot examples

• You lack fine-tuning budget but need specific formatting or tone.
• Edge cases keep failing and you can capture them in examples.
• You’re seeding a new locale and want quick quality before data accrues.

Examples boost quality and reduce retries, but they inflate token cost and latency. PMs must pick the minimal set that moves metrics, and ensure examples stay current. They’re a lever for rapid iteration before committing to fine-tunes.

Curate 3–6 examples that mirror your top user intents and failure modes. Keep them short, anonymized, and labeled with the rule they demonstrate. Rotate examples quarterly and run targeted evals to confirm each still pulls its weight. In 2026, generate synthetic variants automatically, then human-review the best ones before shipping.

A feature that drafts PRDs adds four examples: two good summaries, one refusal, one edge-case with missing data. Accuracy on completeness checklist rises from 68% to 83% while latency stays under 1 s because total tokens stay below 800.

• Adding too many examples, causing the model to copy text verbatim or exceed token limits.
• Leaving outdated UI terms in examples, which confuses users after redesigns.
• Failing to anonymize data, creating privacy risk in logs.