Contents · 7 sections↓
Role prompting asks the model to answer as a specific person: a skeptical customer, a support lead, an editor. Used well, it changes what the model pays attention to. Used as a magic phrase ("you are a world-class expert"), it does little. This guide is for anyone reviewing drafts, testing messaging or getting feedback from a model. It shows one real draft, a price-change email, reviewed with a generic prompt and with two specific roles, and the model's output for each.
What is role prompting?
Role prompting tells the model whose perspective to take before it answers, for example "Act as a customer who has paid for this product for three years." A good role specifies who the person is, what they care about and what they should produce. Research suggests a persona alone rarely makes answers more accurate. The value of a role is perspective: it surfaces concerns a generic answer misses.
Key takeaways:
Roles change the angle, not the IQ. "You are an expert" doesn't add knowledge the model lacks.
Make the role concrete: situation, what they care about, what they don't know.
Give each role a job and a format, or you get a generic review in costume.
Watch for invented facts. A role answering "as the support lead" may make up policies.
[02]
How role prompting works
Models have read text from countless perspectives. A specific role points the model at one of them, which shifts vocabulary, priorities and what it notices. That is why "a freelancer who watches costs" reacts differently to a price increase than "an editor" does.
What roles don't reliably do is improve correctness. When "A Helpful Assistant" Is Not Really Helpful (Zheng et al., 2023) tested many personas in system prompts across factual questions and found that adding a persona did not improve performance overall, and the effect of any given persona was hard to predict. So use roles for perspective, audience and tone, and use context and checks for accuracy.
A role prompt that works has four parts:
Part | Example |
|---|---|
Who | "A freelance designer who has paid for Ledgerly Pro for three years" |
What they care about or know | "Watches costs closely, has never used the mobile app" |
The job | "Read the email as you would the moment it lands" |
The format | "First reaction, what annoys you (quote it), what would make you stay" |
[03]
Worked example: reviewing a price-change email
The draft:
Subject: Exciting changes to your Ledgerly plan
Hi there,
We've been working hard to make Ledgerly better than ever. Over the past year we added recurring invoices, multi-currency support and a brand new mobile app.
To keep investing in the product, the Pro plan will move from $12 to $16 per month starting with your next billing cycle after December 1.
Nothing changes in how you use Ledgerly. If you have questions, just reply to this email.
Cheers,
The Ledgerly teamNo role: "Review this email."
The output (trimmed):
Summary of issues
- Tone: Friendly but a bit generic — could be warmer and show appreciation for existing customers.
- Clarity on timing: “starting with your next billing cycle after December 1” is ambiguous …
- Notice period: Many customers expect 30 days’ notice for price changes. …
…
Rewritten email (concise, clearer)
…
If you want, I can produce alternate tones (more formal, more apologetic, or more promotional) …About 470 words: a checklist, three subject lines and a full rewrite. Some points are useful (the timing is ambiguous), but it is unfocused, and you still don't know how a customer will feel.
Role 1: the long-time customer
Act as a freelance designer who has paid for Ledgerly Pro for three years. You watch costs closely and you have never used the mobile app.
Read the email below as that customer would, the moment it lands in your inbox.
Answer in this format:
- First reaction (one sentence, in your own voice)
- What annoys or confuses you (max 3 bullets, quote the exact phrase)
- What would make you stay (max 2 bullets)
<email>
…draft…
</email>- First reaction: Seriously — another price hike, and I haven't even used the mobile app they keep touting.
- What annoys or confuses you:
- "move from $12 to $16 per month"
- "brand new mobile app"
- "Nothing changes in how you use Ledgerly."
- What would make you stay:
- Honor a legacy price for existing/long-term customers (keep my $12/mo or offer a loyalty discount).
- Provide a prorated refund or extended billing period for customers who prepaid, plus a clear breakdown of exactly what new value I'm paying for.This is the review the generic prompt couldn't give. The subject line says "exciting" about a 33% increase, the justification leans on a feature this customer never used, and "nothing changes" reads badly next to a higher price.
Role 2: the support lead
Act as Ledgerly's support lead. You will answer every reply to this email next week.
Read the draft below and list the questions customers will send because the email doesn't answer them. For each, give the one line we should add to the email.
Format: a table with columns Question | Line to add. Max 5 rows.The output (trimmed):
| Question | Line to add |
|---|---|
| Does this price change affect annual (prepaid) subscribers? | Annual subscribers billed before December 1 will keep their current rate until their next renewal date; the new $16/month rate applies at renewal. |
| When exactly will the new rate take effect on my account? | The new rate will apply starting with your next billing cycle after December 1 (i.e., on your next scheduled invoice date after that day). |
| Will I be charged a prorated amount if the price changes mid-billing cycle? | If your plan changes mid-cycle we will only charge or refund prorated amounts so you won’t be double-billed. |
…
| How can I downgrade, cancel, or get a refund if I don’t want the new price? | You can downgrade or cancel anytime from Settings → Billing in the app, and reach out to support@ledgerly.com for refund requests or help.The questions are exactly right: annual plans, effective date, proration. The "lines to add" are a problem. The model invented policies (how annual plans are handled, prorated refunds) and even a support address, support@ledgerly.com, that the draft never mentions. It played the role so well it made decisions only the real team can make. Fix it in the prompt: "Don't invent policy. Where the answer is a business decision, write DECISION NEEDED instead of a line."
[04]
When to use role prompting (and when not to)
Good uses:
Pre-mortems on messaging: how will a skeptical customer, a new user or a journalist read this?
Reviews from a specific angle: a support lead for missing answers, a lawyer's checklist for risky claims (not as legal advice).
Audience and tone: "explain this as you would to a new hire in finance".
Practice: a role-played customer for training support staff.
Skip it when you need facts or accuracy. "You are a senior tax advisor" won't make tax answers correct; give the model the rules and check the output. And never let a role prompt's output stand in for real people's sign-off.
[05]
Common mistakes
Mistake | Why it fails | Fix |
|---|---|---|
"You are a world-class expert" | Adds no knowledge; results are unpredictable | Describe the person's situation and concerns |
Role with no job or format | You get a generic review in a costume | Give a task and an output shape |
Trusting the role's facts | The model fills in policies and details | "Don't invent facts; mark decisions needed" |
Five roles in one prompt | Views blur together | One role per call, then compare |
Stereotyped personas | Output reflects clichés, not your customers | Base roles on real customer segments or interviews |
[06]
Frequently asked questions
Should the role go in the system prompt or the user prompt?
If the whole conversation uses the role (a support assistant), put it in the system prompt. For one-off reviews like the ones above, the user prompt is fine. See system vs user prompts.
Does "act as" work better than "you are"?
There's no reliable difference. What matters is the detail you give about the person and the task.
Can role prompting replace user research?
No. It is a cheap first pass that catches obvious problems before real people see a draft. Real customers will surprise you in ways a model won't.
How do I combine several perspectives?
Run each role as a separate call with the same draft and format, then merge the findings yourself or in a final call. Separate calls keep the perspectives distinct.
Is role prompting the same as roleplay for fun?
Same mechanism, different goal. Here the role is a tool for a specific output. The same rules apply: concrete character, clear task, clear format.
[07]
Next steps
Previous in the path: few-shot examples.
Next: step-by-step reasoning.
The four parts of any prompt: how to write clear prompts.
Stop roles from inventing facts: grounding.
Tune tone further: polish and style controls.