Technology
Few-shot prompting
Few-shot prompting includes 2-5 input/output examples in the prompt so the LLM learns the pattern by demonstration. Materially improves accuracy on classification, extraction, and format-sensitive tasks vs. zero-shot.
More detail
The trade-off: examples consume tokens (cost + latency). Sweet spot is 3-5 diverse examples covering the edge cases. Beyond 5 examples, returns diminish and you're better off fine-tuning a smaller model. Aiprosol's 200-prompt vault uses few-shot patterns for the classification + extraction prompts where format matters.