When you are working with AI, especially when writing custom instructions, YAML schemas, or JSON contracts, you should avoid giving the system any responsibility for judgment or decision-making.
AI does not handle discretion the way humans do. If you phrase an instruction as “use judgment,” “when appropriate,” or “unless necessary,” you are handing the model a gray zone. In gray zones, it will drift. Over time, it will reintroduce exactly the behavior you were attempting to prohibit, usually in the name of clarity, helpfulness, or style.
Effective instructions for AI need hard edges. They need to be operational, binary, and decidable. Either a behavior is permitted or it is forbidden. Either a structure is allowed or it is banned. The more explicit the boundary, the more reliable the output.
If you want consistent results, do not ask the AI to decide whether a rule applies. Encode the rule as absolute, and take responsibility for exceptions yourself. Humans handle judgment. AI handles execution.