When Two Good Rules Contradict Each Other, the Real Problem Is No Decision Was Made

A developer discovered that their AI agent was behaving inconsistently because two separate instruction files contained conflicting writing guidelines — one demanding terse output, the other requiring readable, well-connected sentences. For months, the agent alternated between compressed-but-unreadable and readable-but-padded responses depending on which rule it weighted more in a given session. Repeated edits to the rules failed to fix the problem because the rules themselves were not at fault. The root cause was that no decision had ever been made about which behaviour should take priority when the two rules conflicted. The author argues that writing a specification and making a decision are fundamentally different tasks — a spec can be fully written yet completely undecided, and the two states look identical until something tries to execute.