A calibrated role

Why does telling a system who to be change what it knows how to say?

Ask a language model to explain inflation and you get one answer. Prefix the same question with you are a central bank economist and the answer changes — different vocabulary, different assumptions about what needs saying, a different level of hedging. Prefix it with you are explaining to a ten-year-old and it changes again.

The obvious reading is that the role gave the system knowledge it did not have. But nothing was added: the weights are identical in all three runs, and no reference material was supplied. So something else is happening, and it is worth being precise about what — because the difference between "the role adds knowledge" and "the role selects from what is already there" decides whether role prompting is a useful tool or a superstition.

Begin with what the system is actually doing. It is producing text conditioned on everything in the context so far — extending a document, in effect. So the question becomes: given a passage that opens you are a central bank economist, what kind of text plausibly follows in the enormous body of writing this model learned from?

Follow that through and the answer stops being mysterious. Text that follows such a framing tends to use particular terms, to assume the reader knows what a nominal rate is, to hedge in particular ways, to structure an argument in a recognisable order, and — crucially — to omit the explanations that a general-audience piece would include. The role is not a switch that enables an expert module. It is a strong hint about which region of an enormous space of possible texts we are in.

Now ask a sharper question. If the role only steers style, why does the content look different too? Because style and content are not as separable as we assume. Consider what the economist framing omits: it leaves out the definitional groundwork, and that space gets spent on second-order detail instead. Consider what the ten-year-old framing omits: technical vocabulary, and it spends that budget on a concrete example. In both cases the selection of which true things to say has changed, along with what is assumed rather than argued. A role sets the audience, and the audience determines what counts as worth saying.

So what does a role genuinely control? Vocabulary and register. What is assumed as shared knowledge. What is explained versus asserted. How much hedging appears. The default shape of the answer — a list, an argument, a story. Which of several defensible framings of a question is taken. That is a substantial list, and it is why role prompting is genuinely useful.

And what does it not control? The facts. Here the evidence is worth taking seriously rather than assuming the flattering answer. Work published in 2024 examined adding expert personas to system prompts and found no consistent improvement on factual question-answering tasks — the persona changed the presentation without reliably changing correctness. Which is what the reasoning above predicts: nothing about the phrase "you are an expert" retrieves information the parameters do not encode. It selects a register, and a register can be confident without being right.

That points at the real hazard. An expert register raises the fluency and assurance of the output, including when the output is wrong. Style is the main cue people use to judge credibility, so a role prompt can make a bad answer harder to catch, and that risk lands hardest in exactly the domains where the role sounds most impressive.

Which suggests where the actual leverage is. Notice that in all these examples the useful work is being done by information about the communicative situation, not by the identity. "You are a doctor" is vague about what should change. "Explain this to a patient who has just been diagnosed, avoiding jargon, in under two hundred words" specifies audience, register, and length directly — and those are the levers the persona was only gesturing at. The word "calibrated" in the name of this idea is doing the real work: a role is worth having when it pins down the situation precisely enough to determine what is said, and worth little when it is decoration.

One honest caveat: the research picture here is not settled. Personas do appear to matter for tone, safety behaviour, and some open-ended generation tasks, and results vary across models and task types. The claim I would defend is narrow — that a role reliably moves form and unreliably moves factual accuracy — rather than that personas never help.

Three refinements.

First, the misconception this is most often used to correct: role prompting is not a capability unlock. It cannot make a model know a drug interaction it never encoded, and treating an expert persona as a substitute for retrieval, tools, or verification is the mistake that turns a formatting technique into a safety problem.

Second, it does not follow that roles are useless. Getting the register right is a real and hard part of most writing tasks, and a single well-chosen framing does it more reliably than a paragraph of style instructions. The gain is real; it is simply a gain in fit rather than in truth.

Third, prefer specifying the situation over the identity. Audience, purpose, level of prior knowledge, length, and the shape of the output are the variables that actually determine the text. A persona is a compressed way of implying all of them at once, which is why it is convenient — and also why it is imprecise, since you are trusting the model's stereotype of a role to contain the settings you meant.

• Audience specification as an alternative to personas, and whether it outperforms them.
• Why omission is such a strong cue for perceived expertise.
• How system prompts differ from in-context roles in how strongly they steer.

Every term the collection defines is gathered in the glossary.