What makes an AI instruction a rule? Structuring normative force for agents
SparkTsaiX · x · 2026-11-17
- Part 2/6 of the Behavior Rule Architecture series: moving an instruction from a prompt into a file makes it reusable, but doesn't automatically make it a Rule.
- "Please avoid changing unrelated code," "you should not," and "you must not" point the same direction but carry different normative force—the model still has to infer whether each sentence is a preference, recommendation, obligation, or prohibition.
- Fix: Normative Natural Language (NNL) reduces structural variation by exposing fixed components: Actor + Normative Expression + Action + Target + Condition; Constraint Normative Language (CNL) tightens it further.
- Thesis: structure before interpretation—less inference required each time an instruction is consumed, more predictable agent behavior.
Related event: Should AI rules live in code or prompts? Rule architecture debate heats up(3 posts)→