Few-Shot Prompting

Few-shot examples are the most effective technique when detailed instructions alone still produce inconsistent, unusably-varied results. Four named applications:

• Ambiguous-case handling — 2 to 4 targeted examples showing the reasoning for why one action was chosen over other plausible alternatives. The reasoning is the actual payload here, not just the final answer shown.
• Format consistency — demonstrating the exact output shape (location, issue, severity, suggested fix) directly, rather than describing it.
• False-positive reduction with generalization — examples distinguishing an acceptable pattern from a genuine issue. The critical property: examples let the model generalize to novel variants of the pattern it hasn't seen before, in a way a hardcoded exception list cannot.
• Varied document structures — showing extraction across differently-structured source documents (inline citations vs. bibliographies, methodology sections vs. embedded details) fixes empty or null extraction on document shapes the model hasn't been shown yet.

The discriminator: examples generalize; enumerated lists only match what was explicitly listed. Whenever a scenario mentions novel, previously-unseen, or continually-evolving patterns, few-shot examples are the fit — a regex filter or a hardcoded exception list only ever catches the specific instances someone thought to enumerate, and new variants slip through by design.

A trap worth naming: reaching for few-shot examples to fix a problem whose actual root cause is a thin tool description (Domain 2, §2.1) adds token overhead without addressing the cause — examples fix inconsistent generation, not poor tool selection.

The four applications at a glance: