What LLMs 'Know' That You Don't Know They Know: Stop Inventing Prompt DSLs and Ride 50 Years of Unix Pre-Training Gravity

A recent analysis from Dev.to argues that developers should stop creating complex, custom natural-language prompt DSLs for AI agents. Instead, the author suggests leveraging established computer science formalisms—such as Makefiles, Unix init.d structures, and RFC 822 headers—which are deeply embedded in the pre-training data of frontier LLMs. By using these battle-tested primitives, developers can achieve significantly higher reliability and performance. The article presents data from 1,680 controlled trials showing that replacing verbose, custom English instructions with canonical CS formalisms leads to massive improvements in task execution, crash recovery, and state management. Because these standards appear millions of times in the model's training corpus, they act as high-efficiency 'static linkages' that allow the model to perform complex logic without burning fragile working-memory attention on interpreting bespoke, low-frequency prompt syntax.