More examples is not better. How to pick the two or three that teach the pattern instead of memorizing it.
Few-shot examples are training data you ship in the prompt. Choose them like a teacher, not a hoarder.
Pick examples that span the hard cases: the empty input, the ambiguous one, the one that should be refused. Three sharp examples beat ten similar ones.
If your examples all look alike, the model learns the surface, not the rule.
What actually moves output quality - and the small set of habits that beat clever one-liners every time.
A repeatable four-part structure for prompts that stay reliable as your product and models change.
Copy-ready building blocks - from role priming to output contracts - with annotated before/after examples.
