Prompt engineering that scales past the demo

A prompt is code without a type system. It works on the three examples you tried and breaks on the fourth, which is why production prompts are built as templates with explicit slots, not paragraphs of prose.

Separate the instruction, the context and the output contract. The instruction rarely changes, the context changes per request, and the output contract should be strict enough to parse.

Build an evaluation set of twenty real inputs before you tune anything. Every change is then a measurable win or loss instead of a vibe.

Version prompts the way you version code, and keep the failing examples. Regressions in prompt work are silent — nothing throws.