Role-prompting for writing
After this, you'll know how to write a role header that changes Claude's default voice, register, and persuasion pattern before you type a single word of your actual request; and you'll have tested it against the same task with no role, so you can see the gap.
Before you start
Complete Giving Claude Your Context and The Colleague Brief first. This lesson builds on both; once you can brief Claude on your own context, this lesson adds the complementary move of casting Claude's expertise.
The idea
A Claude Project lets you configure your writing context once and stop reprompting it every session. A Claude Project is a saved workspace that remembers your files and instructions across chats. Every lesson before this one pointed the context arrow at you: your role, your audience, your brief, your tone rules. That arrow points the right direction most of the time. But there is a second arrow, and most people never touch it. This lesson is about pointing it at Claude.
A role prompt casts Claude's expertise before the task begins. "You are a direct-response copywriter with 15 years of B2B SaaS experience" (B2B means selling to businesses rather than to consumers) is not flattery or a magic spell. SaaS means software you rent monthly instead of buying. It is an instruction that changes which part of Claude's training gets weighted first. The register (the level of formality and the tone, the difference between a text to a friend and a letter to a bank) shifts. Sentence length changes. The persuasion pattern changes. The vocabulary changes. The opening line stops being a scene-setting paragraph and starts being a hook designed to earn a second sentence. None of that comes from anything you told Claude about yourself.
Here is the before and after: Without a role header, Claude's default writing mode is somewhere between a polished generalist and a slightly formal content site. It is competent, readable, and forgettable. With a role header, "You are a former magazine editor who now writes conversion-focused email sequences for indie creators," the same request about announcing a price increase produces a different opening sentence, a different structural choice, and a different close. Not because Claude gained new information about you. Because you changed who Claude is for this task.
The distinction that matters: "here is who I am" and "here is who YOU are for this task" are different inputs. The first changes what Claude writes about. The second changes how Claude writes, what it reaches for, how it weighs the options in front of it. Both matter. L0 and L1 covered the first. This lesson covers the second.
A well-cast role header has three parts. The profession or specialization (direct-response copywriter, narrative journalist, SaaS onboarding writer). The tenure signal, because "15 years" and "two years" produce different voices even with the same job title. And, optionally, the specific context where that expertise was built (B2B SaaS, lifestyle media, healthcare compliance). You do not need all three every time. A single sharp specialization line beats a long paragraph of credentials. "You are a direct-response copywriter" is already useful. "You are a direct-response copywriter who spent 10 years writing fundraising letters for nonprofits" is more specific and produces a more specific voice.
The common mistake is using a role prompt to set up what you want instead of who Claude is. "You are an expert who will help me write a clear and compelling email" is not a role prompt. That is a task description wearing a costume. A real role prompt names a perspective, a body of experience, a professional habit of mind. The test: could this phrase describe an actual person's career? If yes, it works. If it could only describe a desired output, rewrite it.
Role prompts compound with your context. "You are a former investigative journalist" plus your briefing on the audience, the publication, and what success looks like produces better output than either alone. The role sets the register. Your context sets the target. Both arrows pointing at once is the technique at full power.
Cast the expert; then brief the job.
| Your context | Role prompt | |
|---|---|---|
| What it sets | Who I am, my audience, my brief | Who YOU are for this task |
| Changes | What Claude writes about | How Claude writes it |
| Example line | I run a logistics ops team | You are a direct-response copywriter |
| Covered in | l0-02 and l1-01 | Two arrows, not one |
Both arrows point at Claude. Pointing both at once is the technique at full strength.
Try it (8 min)
Watch out for
- Writing a task description instead of a role. 'You are an expert who will write clearly and helpfully' is not a role. It is a wish. A role names a specific body of experience: 'You are a B2B content strategist who spent five years writing for enterprise HR software buyers.' That specificity is the whole mechanism.
- Stacking five roles into one header. 'You are a copywriter, editor, SEO strategist, and brand consultant' splits Claude's attention across four different professional priors. One sharp specialization outperforms three vague ones. If the task requires two genuinely different perspectives, use two separate prompts.
- Forgetting that role prompts compound with your context. The role sets the register; your brief sets the target. Using only one of the two is leaving output quality on the table. Cast the expert first, then brief the job exactly as you would have anyway.
- Using a role to make Claude sycophantic. 'You are a supportive writing coach who always finds the best in my work' produces feedback that is pleasant and useless. Role prompts should cast productive expertise, not a desired emotional register toward you.
- Applying the same role to every task. A cold email, a product walkthrough, and a company blog post want different casts. The role is not an identity file for Claude across your workflow. It is a task-level input, same as the task description itself.
Paste this into Claude
You are a direct-response copywriter with 12 years of experience writing email sequences for B2B software companies. Your default mode is short sentences, one idea per paragraph, and an opening line that earns the second line. You are allergic to preamble. You never start with a compliment, a restatement of the reader's problem, or a sentence that could have been deleted. I need you to write a 5-sentence cold outreach email for a project management tool aimed at operations managers at mid-size logistics companies (50–300 employees). The tool's core claim: it cuts weekly status-meeting time by half because all updates flow through one async feed instead of being collected by a coordinator in meetings. Write the email. Then, below a divider, write two sentences explaining why you chose the opening line you did; what the direct-response principle was. Do not explain what you are going to do before doing it. Start with the email.
What a good response looks like
Subject: You probably have a meeting this week that could be a message Hi [Name], Operations managers at logistics companies I've worked with run 4–6 status meetings a week. About half that time is coordinators collecting updates that already exist somewhere; in a spreadsheet, a chat thread, an inbox. We built a tool that routes those updates to one async feed automatically, so the meetings become decisions-only instead of update collection. Companies like Freight Line Logistics cut weekly meeting time by 52% in the first month. Worth 20 minutes to walk through how it works for your team? [Your name] --- Why that opening line: it uses a pattern interrupt; the subject line assumes something true about the reader's week before they've opened the email, which earns the open. The first sentence then confirms it with specificity (4–6 meetings, logistics context) rather than stating it as a problem the reader needs to accept. Direct-response principle: proof of specificity signals that this email was written for them, not blasted to a list.
What good looks like
- The email's opening line is not a question, not a compliment, and does not restate the reader's problem in generic terms. It earns the next sentence.
- No sentence in the email is longer than 20 words. The longest paragraph is two sentences.
- The email makes one claim, not three. There is a single call to action at the close.
- Claude's two-sentence explanation names a specific direct-response principle (scarcity, specificity, loss aversion, pattern interrupt, proof element); not a generic observation about 'being clear.'
- You could delete the role header, re-run the same task, and the output is noticeably more generic or verbose. The gap is the lesson.
Go deeper (6 min)
Paste this into Claude
You are a narrative journalist who spent eight years writing long-form features for business magazines before moving to content strategy. You default to concrete scenes over abstractions, active verbs over adjectives, and specific numbers over general claims. You have a strong bias against sentences that could have been written by a committee. I'm writing a 300-word case study introduction for a small accounting firm that helped a restaurant owner recover from a botched bookkeeping situation left by a previous accountant. The restaurant owner had been overpaying quarterly taxes for two years without knowing it. Write the first 80 words of that case study; the opening scene or paragraph that would make a reader want to finish the piece. Then write one sentence of craft notes explaining your structural choice. Do not use the word "journey." Do not start with the restaurant owner's name.
What good looks like
- The opening contains at least one concrete sensory or procedural detail (a number, a document type, a month, a dollar figure) rather than an abstract problem description.
- No sentence begins with 'It is' or uses a passive-voice lead.
- The craft note identifies a specific structural choice (in medias res, withholding the resolution, anchoring with a specific detail) rather than saying 'I tried to make it engaging.'
- The word count is 75–90 words for the opening passage; not a full paragraph short, not over the target.
- Re-running the same task without the role header produces a noticeably more generic, report-style opening.
When this breaks
- Breaks when the role is a credential list rather than a perspective. 'You are a Harvard-educated marketing professional with 20 years of Fortune 500 experience' does not give Claude a voice to reach for. A credential list describes status, not craft. Name the specific genre of work and the specific professional habit: 'You are a conversion copywriter who writes landing pages for bootstrapped SaaS products.'
- Breaks when the role contradicts the task. Casting a long-form magazine writer to produce a five-line product description creates tension Claude resolves by averaging the two; and the average is mediocre. Match the cast to the format: short-form tasks want short-form specialists.
- Breaks when the role is implicit and the task is ambiguous. 'Write something about our product launch' with a role header still leaves too much open. The role narrows the register; it does not replace a clear task description. Both inputs are required.
AI can help with this
Open your AI writing tool and run the same task twice: once with a specific role header and once without it. Put both outputs side by side and ask the tool to name concrete differences in opening strategy, sentence rhythm, and persuasion pattern.
You can now
You can write a role header that describes a real writing specialty.
- You can separate your context from the tool's role for the task.
- You can run the same task with and without the role header.
- You can name one visible difference in structure, register, or opening strategy.
Key takeaways
A role prompt casts Claude's expertise, not your context. One line changes the register, the sentence length, and the opening strategy; independently of anything you told Claude about yourself.
- Role-prompting points the context arrow at Claude, not at you. 'Here is who YOU are for this task' is a different input from 'here is who I am.'
- A good role header names a specialization, a tenure signal, and optionally the specific context where that expertise was built. Three elements, one sentence.
- The test for a real role prompt: could it describe an actual person's career? If yes, it works. If it only describes a desired output, rewrite it.
- Role prompts compound with your brief. The role sets the register; your context sets the target. Both arrows pointing at once is the technique at full strength.
- Run the same task with and without the role header at least once. The gap between outputs is the fastest way to calibrate how much specificity your role needs.
Go deeper