Not every question needs an expert. Give AI a role for the wrong kind of task and the answer gets worse, not better, a strange enough finding that four separate write-ups covered it this year.
A role changes how AI writes, not what it knows. Giving it a persona shifts tone, structure, and the kind of language it reaches for. It doesn't add facts it didn't already have, which is exactly why it helps some tasks and quietly hurts others.
Where a role earns its place
| Role | Use it for | Example |
|---|---|---|
| Friendly editor | Tightening an email or message before you send it | “Read this as a friendly editor, cut anything that isn't doing work” |
| Skeptical colleague | Stress-testing an idea before you commit to it | “Poke holes in this plan like a colleague who doesn't want to see you waste a quarter” |
| Patient tutor | Explaining something unfamiliar without jargon | “Explain this like I've never touched a spreadsheet” |
| Blunt reviewer | Feedback on creative or written work | “Review this like an editor who won't soften the note” |
| Local expert | Trip planning, avoiding tourist-trap advice | “Answer like someone who actually lives there” |
| Brand voice writer | Matching a specific tone across posts | “Write this the way our LinkedIn posts usually sound” |
| Beginner-friendly coach | Fitness, hobbies, anything with a learning curve | “Coach me like I've never done this before” |
| Senior consultant | Framing a decision, not solving a calculation | “Advise like a consultant weighing trade-offs, not crunching the numbers” |
| Copyeditor | A grammar and style pass, nothing more | “Just tighten the language, don't change the argument” |
| Devil's advocate | Testing whether an argument actually holds up | “Argue the other side as hard as you can” |
Every one of these is a style or perspective job: how something sounds, what it emphasizes, whose instincts it borrows. None of them ask the AI to be more factually correct, and that distinction turns out to matter a lot.
Where habit beats judgment
Most people who add a role to a prompt do it out of habit, not judgment: “act as an expert” gets bolted onto every question, whether the task needs an expert's opinion or just the model's plain read of the facts. That habit is exactly what Wharton's Generative AI Labs tested: expert personas on factual, benchmark-style questions, across six models, science, engineering and law included, produced no consistent accuracy gain, and low-knowledge personas (“act like a child”) actively hurt results. The Register put a number on it, reporting on related USC research: on the MMLU benchmark, the plain version of a factual question, no persona at all, outscored the expert-persona version, 71.6% against 68.0%.
That doesn't mean a role is a bad idea. It means the presence of a role was never the variable that mattered, the decision to use one was. Search Engine Journal and PromptHub land on the same split from a different angle: personas reliably help writing, tone, and roleplay-style tasks, and reliably risk hurting math, coding, and general-knowledge questions. The likely reason, per the Register's reporting, is that an expert prefix nudges the model into instruction-following mode at the expense of factual recall, useful when the job is how to say it, a liability when the job is what's actually true.
A role earns its place exactly when the task is about how something's said, and costs you accuracy exactly when the task is about what's actually true. Nothing about a plain factual question waves a flag that tells you to skip the persona, which is why the habit is so easy to fall into and so easy to get wrong.
Role is letter three of CORDON
Knowing when that decision goes one way or the other is exactly the job Role does inside CORDON, not “add a persona,” but “does this task even call for one.” It's one of six fields, alongside Context (covered in Why AI can't help without context.), that shape the finished prompt without you having to remember which combinations work. Try CORDON and let it decide when a role earns its place, and when it doesn't.