What does it reliably change?

Register, emphasis and caution. A named reviewer surfaces objections; a named explainer assumes less. Those are real and useful shifts in how an answer is shaped, and they are shifts in presentation rather than in what the underlying material contains.

When does it help most?

When the request has not otherwise said who the answer is for or what standard applies. A role fills that gap efficiently. Where the request already states the audience and the standard explicitly, the additional effect is small, because the role was supplying information that is now present anyway.

Confidence rising faster than accuracy. A domain title produces the vocabulary and assurance of expertise without the knowledge, which makes a wrong answer harder to spot precisely in the areas where a reader is least able to check it. The technique is safest where the reader could have written the answer themselves, and least safe where they could not.

Glossary

Role prompting

Role prompting tells a system what part to play, such as a cautious reviewer, which sets the register and the standard of what comes back rather than what it knows.

In plain terms

Telling it who to be. Answer as a careful editor, or a sceptical reviewer, or somebody explaining to a beginner. It reliably changes how the answer reads: what it emphasises, how cautious it is, how much it assumes. It does not make it know more than it did.

Why it matters

Because it is the most repeated piece of prompting advice and the one most often oversold. Knowing what it genuinely does, which is shaping register and emphasis, stops teams expecting expertise from a sentence and helps them use it where it actually works.

How it works

It sets register and priorities rather than knowledge. Naming a cautious reviewer produces hedged, qualified output that surfaces objections; naming an enthusiastic explainer produces the opposite. Both draw on the same underlying material. What changed is which parts of it get foregrounded and in what voice.

The effect is largest where the request is otherwise underspecified. A vague question answered without a role gets a generic answer; the same question with a role gets one shaped by that role, which frequently reads as a large improvement. Where the request already states the standard and audience explicitly, adding a role changes much less.

It is the same mechanism as naming a part in a multi-participant arrangement, which is worth knowing because the limits transfer. Telling something it is a checker makes it attend to checking; it does not give it a different vantage point from the thing it is checking. Both are instructions about attention.

Claimed expertise does not arrive with the title. Naming a domain expert changes the vocabulary, the confidence and the shape of the answer, all of which read as expertise, while the underlying material is unchanged. That combination, more confident presentation without more knowledge, is the specific risk the technique carries.

What the title changes

The reason this technique is repeated so widely is that the left column is genuinely valuable and immediately visible. Ask for a plain explanation and you get one; ask a cautious reviewer and the objections arrive first. Nobody has to be persuaded that it works, because the difference is obvious on the first try. What is not visible on the first try is that nothing in the right column moved. The answer is more assured, better organised and phrased the way a specialist would phrase it, all of which are the signals a reader uses to judge whether somebody knows what they are talking about. In a domain the reader knows well, the gap shows up quickly and the technique is used sensibly. In a domain they do not, the same shift in presentation is indistinguishable from competence, which is precisely the situation in which somebody reaches for a named expert in the first place.

Seen in the wild

An assistant asked to answer as a sceptical reviewer, where the output becomes markedly more hedged and objection-first without drawing on anything new.
Claude
A framework where each participant is given a role, a goal and tools, which is this technique used structurally rather than conversationally.
CrewAI
A general assistant whose documentation notes it can make unsupported assumptions, which is the failure a confident role makes harder to spot.
ChatGPT

Common misconceptions

People assume

Naming an expert produces expert answers.

In fact

It produces expert-sounding answers from the same underlying material. Vocabulary, confidence and structure all shift; what is known does not. That gap matters most in domains where a reader cannot tell the difference, which is exactly where people reach for the technique.

People assume

It is the highest-leverage thing to change.

In fact

It is the most repeated and rarely the largest. Where the request is underspecified it helps considerably, and where the material in front of the request is wrong it changes nothing at all. What is supplied usually moves the answer more than who it is told to be.

Telling them apart

Role prompting vs Agent roles

Role prompting

Setting register and emphasis for one request.

Agent roles

Standing parts in a multi-participant arrangement, with scope and tools.

The same instruction, used conversationally or structurally. Neither creates the competence the title suggests.

Questions

What does it reliably change?: Register, emphasis and caution. A named reviewer surfaces objections; a named explainer assumes less. Those are real and useful shifts in how an answer is shaped, and they are shifts in presentation rather than in what the underlying material contains.
When does it help most?: When the request has not otherwise said who the answer is for or what standard applies. A role fills that gap efficiently. Where the request already states the audience and the standard explicitly, the additional effect is small, because the role was supplying information that is now present anyway.
What is the risk?: Confidence rising faster than accuracy. A domain title produces the vocabulary and assurance of expertise without the knowledge, which makes a wrong answer harder to spot precisely in the areas where a reader is least able to check it. The technique is safest where the reader could have written the answer themselves, and least safe where they could not.

Key takeaways

It sets register and priorities, not knowledge.
The effect is largest where the request was otherwise underspecified.
The same mechanism as naming a part in a multi-participant setup, with the same limits.
Confidence rises with the title; accuracy does not, which is the risk it carries.

Tools that use this

Claude
A sceptical-reviewer framing producing hedged, objection-first output.
CrewAI
Roles, goals and tools per participant, the structural form of the same idea.
ChatGPT
Documented unsupported assumptions, which confidence makes harder to spot.