Two drafts of a prompt exist for an internal tool that explains error messages to junior engineers. One opens "You are a patient senior engineer explaining to a new graduate." The other lists six rules about vocabulary, length, and assumed knowledge. Both perform similarly in testing. Which consideration should decide between them?

The principle — A persona and a rule list are two ways of specifying the same thing at different resolutions. The persona compresses audience, register and depth into a phrase; the rule list enumerates them. When both work, the deciding question is what you will need to do with the specification later.

Why the key is correct — Anthropic describes setting a role as what focuses a model's behaviour and tone for your use case, and Google Cloud lists persona among prompt components as who or what the model is acting as, sitting alongside constraints and response format as separate components. That separation is the clue. A persona is compact and reads naturally, but it cannot be checked: there is no way to test whether an output was "patient senior engineer" enough. Six explicit rules can each be verified, which matters as soon as you need to evaluate outputs systematically, explain a regression, or hand the prompt to someone else. Google's guidance to provide objective constraints rather than subjective qualifiers points the same way whenever compliance has to be assessed. If the requirements are genuinely atmospheric, the persona is the better tool and the shorter one; if they are things someone will audit, enumerate them.

Why the others are wrong — Whether the job title is real does not determine how well a persona conveys audience and depth. The model's knowledge cannot be the deciding factor when both drafts already perform similarly, and a persona would not supply knowledge in any case. And placement in the system layer is a rule that applies equally to whichever draft you pick, so it cannot separate them.

Remember this — Personas are for register. Explicit rules are for anything you will have to verify.

Sources — Anthropic, Prompting best practices; Google Cloud, Overview of prompting strategies.