A team's support assistant is told in its opening message to "act like a friendly senior support agent", but over a long conversation it drifts back to a neutral, generic voice. What does the guidance identify as the fault?

The principle — Where an instruction lives determines how long it lasts. Standing instructions and the immediate request are different layers, and a role belongs to the standing one.

Why the key is correct — Google's prompt-health checklist lists a missing role definition as a distinct fault, saying that if you are going to ask the model to act in a specific role, that role should be defined in the system instructions. Anthropic makes the same placement point, describing a role set in the system prompt as what focuses behaviour and tone. A persona introduced as ordinary conversation is one message among many and competes with everything said afterwards; the same words in the system layer apply to every turn.

Why the others are wrong — A persona is not a knowledge source, so missing experience is not the mechanism. Specificity strengthens a role rather than undermining it. And a system instruction persists by design, which is precisely why restating it each turn is a symptom rather than a fix.

Remember this — If a behaviour should hold for the whole conversation, it belongs in the standing layer.

Sources — Google Cloud, Overview of prompting strategies.