How to Give Claude a Role Worth Following
"Act as an expert marketer" is one of the most common pieces of prompting advice, and on its own it does almost nothing. A label without a stance doesn't tell Claude what an expert marketer would actually notice, question, or refuse to let slide, so the output barely changes from not including the label at all.
Warning
A role only works if it comes with a stance. "Act as an editor" changes nothing. "Act as an editor who cuts anything the data doesn't support" changes everything.
A role isn't a costume Claude puts on. It's a filter for what gets noticed, prioritized, or pushed back on.
What a role is actually for
A well-built role does two things a plain request doesn't: it tells Claude what to prioritize when there's a tradeoff, and it gives Claude permission to disagree with you instead of defaulting to agreeable. Neither of those comes from naming a job title. They come from describing how that role would actually approach the material.
Review this paragraph as a skeptical editor who cuts any claim the data doesn't directly support: "Our new onboarding flow increased signups by 40% and completely transformed how customers experience the product."
”That role produces something specific: it will likely flag "completely transformed" as an unsupported claim riding on the back of a real number, and suggest cutting it. "Act as an editor" alone would probably just tighten the sentence and leave the overreach untouched, since nothing told it that overreach was the thing to hunt for.
Two ways to use a role
Evaluating something that already exists, where the role's job is to catch what you'd miss reading your own work
Generating something new, where the role's job is to shape tone, priorities, and what gets included or left out from the start
For evaluation, the strongest roles are ones with a specific bias built in: a skeptical editor, a first-time user who's never seen the product before, a reviewer who only cares whether the argument holds up under the weakest counterexample. The bias is the point. It's what makes the role notice something a neutral read would skip past.
You're a first-time user who has never seen this app before. Read these onboarding instructions and tell me every point where you'd be confused or unsure what to do next.
”For generation, the role shapes what the output leads with. Ask Claude to draft a project update "as a project manager focused on what leadership needs to decide this week" and you get something structured around decisions and risks. Ask it to draft the same update "as a project manager focused on team morale," and you get something that leads with wins and effort, even from the same raw notes.
Inside Claude Tutorial
A well-built role transfers across almost every kind of request.
Once you've written one role with a real stance, the pattern applies everywhere: reviewing a document, drafting one, or stress-testing an idea. The app has a full lesson dedicated to just this.
An edge case: when a role fights the facts
A role with too strong a bias can start distorting output rather than sharpening it. Asking Claude to review something "as your biggest fan" or "as your harshest critic" can push it toward telling you what that persona would say rather than what's actually true about the work. If a role starts producing praise or criticism that doesn't track with the substance in front of it, that's a sign to dial back to a role defined by a specific job to do, "check for unsupported claims," rather than a specific emotional stance, "be really harsh."
Continue reading
- Writing Prompts That Get You a Usable First Draft: the base framework a role sits on top of.
- Using Examples to Show Claude Exactly What You Want: another way to shape output precisely, worth pairing with a role for anything format-sensitive.