Few-Shot Pattern Recognition

Show 2-4 examples → the model maps the relational structure, not the surface details (Chase & Simon's chess masters, applied).

THE MINDSET SHIFT

“Chess masters don't memorize moves — they recognize patterns. Few-shot prompting is the same trick: 3 good examples teach the model your relational structure, not your literal phrasing.”
— SHE · YOUR AI GUIDE

Chase & Simon's 1973 chess-master study found that experts didn't compute better than novices — they recognized 50,000+ chunks of typical position patterns. Few-shot prompting is the same mechanism: 2-4 examples activate a relational schema the model maps onto your new input.

The failure mode is literal mimicry. If your examples all start with "In Q3, …", the model thinks "start every output with a quarter reference." The fix: vary the surface form, keep the structure consistent. Mix example sentence lengths. Use diverse opening words. Keep the deep structure — voice, reasoning depth, format — identical across all examples.

“3-4 examples captures most of the few-shot gain; beyond that, returns flatten.”

Brown et al., GPT-3 paper, 2020

“Chess masters store ~50,000 chunks of board patterns; few-shot prompting borrows the same mechanism.”

Chase & Simon, "Perception in chess," 1973