Instruction Best Practices: Precision Beats Clarity

Controlled experiments reveal that precise, narrow scoping in LLM instructions dramatically outperforms broad categorical scoping—broad scopes actually perform worse than incorrectly scoped instructions. Additionally, naming exact constructs (e.g., 'unittest.mock') yields roughly 10x better compliance than describing categories (e.g., 'mock objects'). The key takeaway is that if a scope wouldn't work as a grep pattern, it should be rewritten or removed entirely.