Is it actually possible to build a model-agnostic persistent text layer that keeps AI behavior stable?

About this article

Is it actually possible to define a persistent, model-agnostic text-based layer (loaded with the model each time) that keeps an AI system behaviorally consistent across time? I don’t mean just a typical system prompt, but something more structured that constrains how the system resolves conflicts, prioritizes things, and makes decisions even under things like context drift, conflicting instructions, or prompt injection. Right now it feels like most consistency comes from training or the model...