RFC 2119 for Agents.md: Make Your Agent Rules Compact and Strict (fertsch-roever.com)

A recent workshop led by Adam Bien has highlighted a significant shift in how instructions for AI agents, particularly those in AGENTS.md and SKILL.md files, should be structured. Instead of using verbose and often ambiguous prose, the recommendation is to adopt the precise vocabulary of RFC 2119, which specifies five keywords: MUST, MUST NOT, SHOULD, SHOULD NOT, and MAY. This approach streamlines directives, reducing the risk of misinterpretation by language models (LLMs) while providing clarity on mandatory vs. optional tasks. The implications for the AI/ML community are profound. By using this compact and strict language, developers can ensure that LLMs interpret guidelines correctly and consistently, leading to fewer errors during model training or execution. This method not only enhances comprehension but also simplifies rule management, allowing practitioners to quickly identify mandatory actions versus suggestions. Files written this way have been observed to become shorter and easier to review, ultimately boosting productivity in the development and maintenance of AI systems.