There’s a prompting technique that feels slightly silly the first time you try it, and then you never stop using it. Tell the model to be a specific expert — not just “act helpful” but “you are a contract lawyer reviewing this clause for a small business owner” — and the output shifts in a way that adding adjectives to your prompt never quite matches.
The mechanism isn’t mysterious. A model without a role anchor defaults to the statistically average response to your question. That average tends to be hedged, broad, and aimed vaguely at everyone. Assigning a specific expert role narrows the probability space the model draws from — it starts pulling from the patterns, vocabulary, and priorities associated with that domain rather than blending across everything at once.
What actually changes when you assign a role
The most obvious change is vocabulary and framing. Ask “explain what EBITDA means” and you get a definition. Ask the same question as “you’re a CFO explaining this to a founder who just got their first term sheet” and you get something that leads with why it matters in that specific context, not what the acronym stands for.
The second change is what the model chooses to flag. A general “review this email” catches spelling and tone. Add “you are a senior account manager at a consulting firm, this email is going to a client who already pushed back on the invoice last month” and the model starts flagging phrasing that could read as defensive or re-escalate the dispute. The role is literally cueing which class of concerns to surface.
The third change is harder to describe but real: the model adopts a kind of professional responsibility toward the task. Ask it to “be a skeptical editor who has seen too many weak pitches” and it will actually push back, not just suggest improvements. Without that frame it’ll often lead with what’s good and soften the criticism. Same model, same underlying capability. The role changes what it reaches for.
Roles that are specific enough to be useful
Vague roles don’t help much. “Be an expert” is nearly useless — expert at what, for whom, in what context. The more you constrain the role, the more the technique delivers.
Roles that tend to work well:
- “You are a UX researcher who has run usability tests for e-commerce checkout flows.” Paste your onboarding copy and it looks for friction points a real researcher would flag, not generic suggestions.
- “You are a senior hiring manager at a mid-size software company who reads 200 CVs a month.” Now it’s asking “would this make me stop scrolling” rather than “is this technically well-formatted.”
- “You are a skeptical journalist who covers AI hype.” Paste your product description and ask what questions they’d ask. Genuinely uncomfortable in a useful way.
The pattern is: role + specific context for that role + a realistic relationship to the thing you’re asking about. All three matter. A lawyer is too broad. A contract lawyer is better. A contract lawyer reviewing an NDA clause for a freelancer signing with a US client when you’re based in the UK — that’s the level of specificity that gets you something you couldn’t have written yourself.
Where it breaks down
The model doesn’t actually have the expert’s knowledge — it has the patterns of how experts in that domain tend to communicate and reason. For reviewing a marketing email or stress-testing a business plan, that’s usually sufficient. For “be a cardiologist reviewing these symptoms,” the gap between communication patterns and actual clinical judgment is wide enough to matter. This technique is not a substitute for a real professional when the stakes are real.
Second: the model will stay in character even when it shouldn’t. A real expert would sometimes say “I’d need to look that up” or “this is outside my area.” Assigned a role, the model often skips that and generates a plausible-sounding answer in the expert’s voice instead. Confidence within a role can mask uncertainty better than usual. Worth asking explicitly: “Are there parts of this where you’d normally refer me to someone else?”
Third: some roles actively make the model worse. “Be a hard-nosed critic who never gives positive feedback” sounds like you want honesty — what you often get is performative harshness that misses what actually works. The role primes it to find fault whether or not fault exists. Better to ask for “specific weaknesses and why they’d matter to [type of reader]” than to prime it toward a posture.
A concrete way to try this today
Take something you’ve written in the last week — an email, a project summary, a proposal, anything — and paste it with this frame: “You are a [role appropriate to the audience]. Here’s something I wrote. What would you be skeptical about? What’s missing that you’d expect to see?”
The skeptical frame matters. Most prompts implicitly ask the model to be helpful and affirming. This one asks it to be professionally critical, which is the thing a real expert in a meeting would actually do.
You’ll probably get one or two things back that you already knew, and one thing that reframes what you thought was a strength. The third one is what you’re fishing for.
One thing worth sitting with: the better you are at designing the role, the more useful this gets — which means the technique rewards people who already think carefully about their audience. If you’re not sure who the right expert would be for your task, that’s actually a useful question to ask the model first, before you assign any role at all.
