"Act as an expert" doesn't reliably make AI more accurate, but a short role can still help you control tone and audience. The honest answer has two halves: for facts, roles are a coin flip; for style, they work.
Below are the findings of one large study, the three jobs a role can do, a side-by-side example, and a better template than "You are a world-class [job title]".
Does "act as an expert" make the answer more accurate?
Not on average, according to the largest published test I could find. In "When 'A Helpful Assistant' Is Not Really Helpful" (Zheng et al., 2023), the researchers built a list of 162 roles covering 6 types of interpersonal relationship and 8 domains of expertise. They ran them across 4 popular families of language models and 2,410 factual questions.
| What they tested | Result |
|---|---|
| Adding a persona vs. no persona | No improvement in performance across the range of questions |
| Gender, type and domain of the persona | All influenced accuracy, so the choice isn't neutral |
| Picking the best persona per question after the fact | Significantly improved accuracy |
| Predicting the best persona automatically | Often no better than random selection |
The authors' summary is that a persona can help in some settings, but "the effect of each persona can be largely random." In plain terms: you can't reliably know in advance whether "act as a senior editor" will help or hurt a given question, and the study could not find a dependable way to predict it either.
Two caveats. The test measured factual question answering, not writing quality, and models change fast. Treat it as evidence against the belief that expertise is a magic word, not as a rule for every task.
What can a role actually do for my writing?
It can set the voice, the reading level and the point of view. Those are style jobs, and style is what a role is good at. Anthropic's prompting documentation recommends giving the model a role in the system prompt and says even a single sentence makes a difference to its behavior and tone. OpenAI's developer guide similarly puts an "Identity" section first in a developer message, describing purpose and communication style.
| Job | Works well? | Example |
|---|---|---|
| Set tone | Yes | "Write like a patient tutor talking to a nervous beginner." |
| Set audience level | Yes | "Explain it to a smart 15-year-old." |
| Limit scope | Often | "You are a copyeditor. Only fix errors; do not rewrite." |
| Add knowledge | No (per the study above) | "You are a tax expert." (still needs sources) |
| Guarantee accuracy | No | "You are a world-class physician." (still verify) |
Notice the scope row. "Copyeditor" works not because the model gains copyediting skill but because the word narrows what it will do. That is a boundary, not expertise, and the boundary is worth more than the title.
What should I write instead of a job title?
Describe the reader, the outcome and the quality bar. A job title tells the model who to pretend to be. A description tells it who it's helping and what success means. Compare:
Version A: You are a world-class personal finance expert with 20 years of experience. Write about budgeting.
Version B: Write for freelancers with irregular income who dislike spreadsheets. After reading, they should be able to set up one "salary" transfer today. Plain words, warm tone, no jargon. About 600 words. If you are unsure of a number, say so instead of guessing.
Version B has no credentials, and it contains much more information. It names the reader, the end result, the tone, the length and what to do under uncertainty. Version A leaves all of those to chance and spends its words on a claim of authority.
Anthropic's docs make a similar point when they suggest showing your prompt to a colleague with minimal context and asking whether they'd be confused. Would a human freelancer know what to write from Version A? Probably not.
This is the same thinking behind a good book brief. In Interactive eBook Builder, the Configuration Hub asks for your title, target reader and preferences before any writing starts, so the "who is this for and what should they get" answers are written down once and reused, instead of being compressed into a persona line each time you open a chat.
How do I handle facts if a role doesn't make the model an expert?
Give it the facts, and give it permission to say it doesn't know. Anthropic's guide on reducing hallucinations suggests allowing the model to say "I don't know", asking it to quote supporting passages from provided documents, and checking claims against sources, and it adds that these techniques reduce hallucinations significantly but don't eliminate them. A role doesn't appear on that list.
So for a nonfiction chapter, the order of effort is: your own verified notes first, a short role or audience line second, and a human fact-check last. Our guide to stopping AI from making up facts in nonfiction ebooks covers the full routine.
How long should a role be?
One sentence, two at most. A paragraph of credentials ("PhD, 25 years, bestselling author, trusted by Fortune 500 firms") adds length without adding information the model can use, and longer prompts aren't automatically better. We covered the evidence on how long a prompt should be, including a study where accuracy fell from 0.92 to 0.68 as irrelevant text grew to 3,000 tokens. Credentials are the kind of text you'd cut first.
Where a role belongs depends on where it will live. In a one-off chat, put it first. In a reusable assistant, put it in the system prompt and keep it stable; how to write system prompts for a reusable assistant shows a full layout. If you switch tools, remember that one model's response to a role can differ from another's, which is covered in whether you can reuse a prompt across ChatGPT, Claude and Gemini.
A role template I'd use for chapter drafting (my suggestion, so treat it as opinion rather than a tested result):
You are helping me draft a chapter of a practical guide.
Reader: [who, in one line]. Goal after reading: [one action].
Voice: [three adjectives], second person, short sentences.
Quality bar: concrete examples over generalities; flag any claim you can't support.
That is a role in all but name. It sets who is helping, who is being helped and what good looks like. For step-by-step voice setup, see how to give AI your brand voice.
Where do roles still fit in a longer workflow?
Use them as job labels for separate steps, not as claims of genius. A planner, a drafter and a polisher are three different jobs, and naming them keeps each pass in its lane. That is how Interactive eBook Builder is organized: Agent A builds the Spec and Chapter Plan, Agent B drafts from the plan, and Agent C polishes. Each is defined by its task and instructions rather than by a boast of expertise, and the instructions are editable if you want to change the standard.
Before you hit send
- Is the role one sentence, and does it describe a job (edit, tutor, planner) rather than a status (world-class)?
- Does the prompt name the reader and the outcome?
- Did you provide the facts instead of expecting expertise to supply them?
- Did you allow "I'm not sure" as an answer?
- Will you check any claim that matters?
The takeaway: a role is a tone-setting tool, not a knowledge upgrade. Spend your words on the reader and the facts.