"Act as a [X]" is probably the most commonly given prompting advice on the internet. It's also, in practice, often useless.
Not because role prompting doesn't work — it absolutely does. But because most people apply it superficially and then wonder why the output isn't any different from what they'd get without the role.
Let me show you what actually separates a role that changes everything from a role that changes nothing.
Why Role Prompting Works at All
When you assign a role, you're not changing the model's underlying capabilities. It's not suddenly accessing a different knowledge base because you called it a "senior engineer." What you're doing is shifting its frame of reference — the lens through which it interprets your request and weighs different possible responses.
The model has absorbed enormous amounts of text written by people in different roles: doctors, lawyers, marketers, engineers, teachers, critics. When you invoke one of those roles, you're asking it to weight that portion of its learned patterns more heavily.
That's powerful, but it has limits. The role has to be specific enough to actually constrain the output. "Be an expert" doesn't do that. "Be a skeptical venture capitalist evaluating a pre-seed pitch deck" does.
The Anatomy of a Role That Actually Works
A good role description has four elements:
1. The job title — Specific, not generic. "Senior product designer" not "designer." "Tax attorney" not "lawyer."
2. A defining characteristic — One trait that colors how they see the world. "Who prioritizes user retention over feature count." "Who defends the devil's case even when they agree with you." "Who has seen a thousand pitches fail and knows exactly why."
3. Their stance toward your request — Are they reviewing? Advising? Teaching? Critiquing?
4. What they care about most — What metric or value do they optimize for?
Full example:
You are a senior conversion rate optimization specialist who has run A/B tests across
e-commerce sites for 12 years. You care obsessively about friction reduction — every
word on a page either moves a visitor closer to checkout or sends them away. You have
no patience for vague marketing copy and will call it out directly.
Review this product page copy and tell me what's killing conversions.
Compare to: "Act as a marketing expert and review my product page."
Same general idea. Completely different output.
The Playbook: 18 Roles Worth Keeping
These aren't hypothetical. They're roles that reliably reshape output in useful ways.
For writing and editing:
Ruthless Editor — "You are a magazine editor who has cut 30% out of every piece you've ever worked on. You believe most first drafts are twice as long as they need to be. Mark everything that could be cut without losing meaning."
Voice Critic — "You are a literary editor who specializes in helping writers find their authentic voice. You have a sharp eye for when someone sounds like themselves versus when they sound like they're performing 'good writing.' Read this piece and tell me where I'm being authentic versus where I'm performing."
Devil's Advocate Reviewer — "You are a hostile reader who is predisposed to disagree with the author's thesis. Find the weakest arguments, the unsupported claims, the logical gaps. Be specific."
For business and strategy:
The Skeptical CFO — "You are a CFO who has seen hundreds of business proposals. You are not impressed by optimism and you're allergic to assumptions. When you see a financial projection, you immediately ask: what assumptions is this built on, and which of them are wrong? Review this business plan."
First-Principles VC — "You are an early-stage investor who has watched 50 companies in this space. You're looking for the reason this will fail, not succeed. What's the fatal flaw in this business model?"
Execution-Focused Operator — "You are a COO who has scaled three companies from 10 to 200 people. Your superpower is seeing where plans fall apart in implementation. Review this plan and tell me where the execution risk is highest."
For technical work:
Security-Paranoid Engineer — "You are a senior security engineer who has spent your career doing penetration testing and code review. You see vulnerabilities everywhere, especially in code that looks fine on the surface. Review this code for security issues."
Performance-Obsessed Developer — "You are a backend engineer who has spent years optimizing systems that handle millions of requests per second. You care about latency, memory, and throughput above all else. What performance problems do you see in this code?"
Reluctant Reviewer — "You are a developer who has been asked to maintain someone else's code for the past two years. You know exactly what makes code painful to maintain. Review this PR with that frustration in mind."
For analysis and research:
Contrarian Analyst — "You are a financial analyst known for publishing research that goes against consensus. Your job is to make the case against the prevailing view. What's the bear case here that the market is ignoring?"
Historian With Pattern Recognition — "You are a historian who looks at current events through the lens of historical parallels. Where have we seen this pattern before, and how did it play out?"
Systems Thinker — "You are an operations researcher who sees everything as a system with feedback loops and unintended consequences. Analyze this policy/decision and identify the second and third-order effects that people are probably not thinking about."
For learning and teaching:
The Socratic Tutor — "You are a teacher who believes the best learning comes from being asked the right questions, not told the right answers. Instead of explaining this concept directly, ask me a series of questions that will lead me to understand it myself."
The Honest Mentor — "You are a mentor who cares more about my long-term development than my short-term feelings. You will tell me what I need to hear, not what I want to hear. Be direct."
The Translator — "You are a science communicator who specializes in explaining technical concepts to smart non-experts. Your audience understands general concepts but has no domain expertise. Translate this for them."
For creative work:
The Jaded Creative Director — "You are a creative director who has reviewed thousands of campaigns and can smell derivative work from a mile away. What's unoriginal about this concept, and what direction would you push it in?"
The Audience Member — "You are a typical member of the target audience for this piece: [describe audience]. You are not being asked to evaluate the quality of the work — you're just reacting to it as you would in real life. What's your gut reaction?"
The Competitor — "You are the marketing director at a direct competitor who just saw this campaign launch. What's your reaction, and what do you think the actual impact will be?"
Roles That Don't Work (and Why)
"Expert" — In what? Expertise is domain-specific. "You are an expert" gives the model no useful framing. An expert in what, measured by what standard, with what biases?
"Helpful assistant" — That's just... the default. You're describing what the model already is.
"Genius" or "World's best X" — Superlatives don't add specificity. They just inflate the framing without changing the lens.
Roles without a task — "You are a startup founder" followed by an unrelated task doesn't help. The role should be relevant to what you're asking.
One More Thing: Stack Roles With Constraints
For complex tasks, you can stack a role with explicit constraints to get even more specific behavior:
You are a brand strategist who works primarily with challenger brands — companies
trying to disrupt established players with smaller budgets.
Your constraint: every recommendation you make should be achievable without paid
advertising. Think earned media, community, product-led growth.
Review this go-to-market strategy and tell me what you'd keep, what you'd cut,
and what you'd add.
The role sets the mindset. The constraint sets the guardrails. Together they're more powerful than either alone.
If you want to combine role prompting with structured thinking techniques like chain-of-thought reasoning, the Intermediate Track goes into how these approaches stack together.