A role prompt tells the model who to be for this task; a system prompt sets standing rules that apply across tasks. You can use both, but they do different jobs: a role shapes tone and audience, a system prompt holds the reusable policy. OpenAI’s prompt engineering guide explains that instructions provided through the instructions parameter, or through a developer message, take priority over the user input. It compares those standing instructions to a function definition, and the user message to the arguments. A typical developer message, it says, includes identity, instructions, examples and context. Anthropic’s prompting best practices recommend giving Claude a role in the system prompt, noting that even a single sentence can focus behavior and tone. Google’s Gemini 3 prompting guidance says to place essential behavioral constraints, role definitions (persona) and output-format requirements in the system instruction or at the very beginning of the user prompt.
For eBook and workbook projects, Interactive eBook Builder is a $59 one-time Notion AI template with Spec Builder, Drafter, Polisher and SEO BOT already set up, so standing rules and task-specific roles can live in different stages instead of being rewritten into every chat. Notion AI access and usage are separate from the purchase. The examples below are teaching prompts written for this article, not paid Builder prompts and not recorded model outputs. If you work mainly in Notion, our guide to system prompts for an AI writing assistant in Notion covers the Instructions page. If you are trying to decide whether “act as an expert” is doing real work, see Should I tell AI to act as an expert?. This article compares the two techniques directly.
What is a role prompt?
A role prompt assigns a persona or job: “You are a patient teacher for first-time sellers,” or “You are an editor checking this chapter against the notes.” Used well, it tells the model who it is helping and how it should sound. Used poorly, it is a costume—“You are a world-class New York Times bestselling author”—that promises authority the model does not gain from the sentence.
Anthropic’s example is deliberately modest: a system prompt that says “You are a helpful coding assistant specializing in Python.” That is a focus, not a claim of credentials. For a book, the useful version is equally modest and specific:
You are helping a first-time freelance designer who reads in short sessions
and wants calm, practical language. Explain decisions in plain words. Do
not perform expertise you have not been given in the notes.
The role here does three things: it names the reader, it sets a tone, and it refuses a fake expert stance. It does not replace the notes, the format or the limits.
What is a system prompt?
A system prompt is the standing brief that should not have to be repeated in every request. In the OpenAI API, that often lives in an instructions or developer message. In Claude’s API, it lives in the system parameter. In many chat products, a close equivalent is a custom instruction, a project instruction, or, in Notion, an Instructions page. The label varies; the job does not: these are the rules that should still be true on the next chapter.
OpenAI’s suggested shape is a good checklist even if you never touch the API:
- Identity: purpose, communication style and high-level goals.
- Instructions: what the model should do, what it should never do, and any format rules.
- Examples: optional input/output samples.
- Context: standing facts the model will need often.
Google’s long-context and Gemini 3 notes add a placement tip: put critical behavioral constraints, the persona and the output format in the system instruction or at the start, and, when the source material is long, put that material first and the specific question last. For a book, that usually means: standing policy in the system prompt, chapter notes in the request, question at the end.
How do the two techniques differ?
They overlap, which is why the terms get mixed together. A role can be placed inside a system prompt. A system prompt can exist with no persona at all. The useful distinction is scope.
| Question | Role prompt | System prompt |
|---|---|---|
| What is it for? | The stance, audience and tone of this job | The reusable policy for the project |
| How often does it change? | When the page’s job changes | Rarely; only when the book’s rules change |
| What does it usually contain? | Who you are helping and how to sound | Voice, format, sources, limits, and what to do with uncertainty |
| What can it not do? | Make unsupported claims true | Replace the notes for this chapter |
| Where should it live? | In the system prompt, or at the start of the request | In the product’s standing-instructions field when you have one |
A practical split for an eBook looks like this:
- System prompt: “Write for new freelance designers in plain, calm language. Stay inside the supplied notes. If a fact is missing, write 'Need from me' instead of inventing it. Do not add medical, legal or income claims. Use the book’s heading pattern and leave workbook exercises in the specified format.”
- Role for a planning pass: “You are a chapter planner. Your job is to resolve conflicts in the notes and return an outline, not to draft the chapter.”
- Role for an editorial pass: “You are a consistency editor. Flag wording that breaks the standing rules or drifts from the notes. Do not rewrite for style unless a rule is broken.”
The policy stays put. The role changes with the stage. That is also why Spec Builder, Drafter and Polisher in Interactive eBook Builder have separate jobs: each stage can keep the same book policy and a different task role.
A before-and-after example
A weak combined prompt tries to do everything in one costume:
Act as a world-class business expert and write my chapter on revision
limits. Make it authoritative and inspiring.
That prompt does not name the reader, does not supply notes, and uses the role as a substitute for a brief. A stronger version separates the standing rules from the task role, then gives the chapter its own notes:
System or standing instructions:
You help the author of a short eBook for new freelance designers. Use
plain, calm language. Stay inside the notes supplied in each request.
If a fact, rate or legal point is missing, write "Need from me" instead
of inventing it. Do not add income, legal or "industry standard" claims.
Workbook pages must include a one-sentence instruction and writing space.
Task role:
You are the chapter planner for this request. Return decisions and an
outline only, not a full draft.
Request:
Plan the chapter on stating a revision limit before work starts.
Notes:
- The reader says yes to extra revisions because they are afraid of
seeming difficult.
- The chapter should help them put a limit in the kickoff email.
- I have not chosen a default number of revisions; the reader will fill
that in.
Return a short decisions list, a five-heading outline, and one workbook
line. Flag anything that still needs a decision from me.
The improved prompt still has a role, but the role is a job description for this turn. The system prompt holds the policy that should still be there when you later ask for a draft or an edit. If you only have a single chat box and no system field, paste the standing instructions at the top and keep them identical from chapter to chapter.
What a role cannot fix
A role will not make the model more reliable on facts it has not been given. It will not remember a previous chapter unless that chapter, or a short brief from it, is in the prompt. It will not enforce a format as clearly as a stated format and an example will. And a grandiose expert role can hide those gaps, because the prose sounds sure of itself.
If you need consistency across a book, put the rules in the system prompt and put examples in the instructions, as OpenAI describes. If you need a different tone for a workbook page than for a case study, change the role and keep the policy. If a sentence in the draft makes a claim you cannot support, the fix is in the notes and the limits, not in a more impressive title for the model.
For a longer treatment of standing instructions that plan and polish chapters, see how to structure system prompts for chapter planning and editorial polish.
How to keep the split tidy in a real project
Write the system prompt once, on purpose, and stop tweaking it every time a single paragraph goes wrong. Change it when you have seen the same miss across several chapters—invented statistics, a heading pattern that will not stick, a voice that keeps sliding into hype. For a one-off miss, change the request, the notes or the task role.
Keep the system prompt short enough that you can actually read it before you run a chapter. A long policy that you never review becomes a second, hidden manuscript. Identity, a handful of instructions, one short example if you need it, and a rule for missing facts will carry most books.
In the Builder, you can leave those standing rules in the supplied, editable instruction sets and adapt the audience, voice and activities for your niche. SEO BOT can later publish supporting articles to Auto-Blogger, each linking back to your eBook, site and brand, using the same book policy so the blog does not sound like a different author. Review still belongs to you.
FAQ
Can I put the role in the user message instead of the system prompt?
Yes. Anthropic prefers a role in the system prompt, and Google allows either the system instruction or the very beginning of the user prompt. If your tool has a standing-instructions field, use it for the policy and the default role. If it does not, put them at the top of the request and keep the wording stable.
Do I need a different role for every chapter?
Only when the job changes. Teaching chapters, case chapters and workbook pages may need slightly different roles. Two teaching chapters in the same book usually should not. If you change the role every time, you are no longer getting the benefit of a standing brief.
Should the system prompt include my whole book?
No. Standing facts that rarely change can go there, such as the audience and the voice. Chapter-specific notes belong in the request. Long source material should be clearly labeled and, following Google and Anthropic’s long-context advice, placed before the specific question.
Is a system prompt the same thing as a prompt template?
They overlap. A system prompt is the standing part of a template. A full template also has clearly marked blanks for the parts that change, such as this chapter’s notes and the requested output. Our guide to writing a prompt template covers that structure.
Before you hit send
- Have you separated the standing policy from the role for this task?
- Is the role a specific job and audience, or a vague expert costume?
- Are facts, notes and format rules stated outside the role?
- Will this system prompt still be correct on the next chapter?
- If a claim is missing from the notes, have you told the model to flag it instead of inventing it?
- Are you judging the result against the policy and the notes, not against whether the voice sounds authoritative?