Prompt roles: a job description beats a job title
"You are a senior expert" barely moves a modern model — it's a credential sticker, not an instruction. A role earns its place when it encodes behavior, constraints and audience: what to do, what not to do, who you're talking to.
For a while I opened prompts with the responsible-sounding incantation: “You are a senior engineer with 20 years of experience. Refactor this function.” It felt like I was setting the model up to do its best work. The output came back clean, confident — and no better than when I dropped the first sentence entirely.
That sentence wasn’t an instruction. It was a credential sticker. And the model doesn’t get more capable because I told it it was.
Why the title does little
There’s a real memory behind the habit. On older, smaller models, a persona prefix moved the needle — it nudged a weak model toward a register it otherwise wandered away from. So “you are an expert X” became muscle memory, and the muscle outlived the reason.
On a modern Claude (3.5 onward, and the 4.x family) the effect on objective accuracy is marginal to nil. A 2024 study on personas in system prompts put numbers on it: adding a persona didn’t reliably improve measurable performance, and on some tasks it slightly hurt. The intuition is simple — the model already knows how to answer a coding question well. A job title in front of the task doesn’t add a skill it was missing. It just spends tokens.
What a role actually buys you
A role earns its place when it stops being a label and starts encoding a job description — three things, specifically:
- Behavior — what to do, and what not to do.
- Constraints and stance — the posture it judges from (“you have no stake in this”).
- Audience — who the answer is for, which sets the level of detail and the register.
The blog already has a role that works, and it works for exactly these reasons. The writer/reviewer pattern leans on a reviewer agent whose prompt opens:
You are a code reviewer. You did not write this code and you have
no stake in it. Your job is to find problems, not to approve.
The word “reviewer” isn’t what does the work. The work is in the rest: a stance (no stake), a behavior (find problems), and an explicit anti-goal (not to approve). Strip that to just “You are a senior reviewer” and you’re back to a sticker.
Bad vs good
The cargo-culted version:
You are a world-class senior Go engineer. Refactor this function.
The version that actually constrains the work:
Refactor this function. Keep the public signature. No new deps.
Optimize for readability over cleverness. Explain any behavior
change in one line.
The second prompt has no role at all and lands better — because the gains live in instructions, constraints, and output format, not in a self-image. The role only re-enters when it’s pulling its weight on tone or audience:
Explain this stack trace to a backend dev with no Go experience —
assume they know HTTP but not goroutines.
That last one isn’t a credential. It’s a calibration: it tells the model who’s listening, and the explanation reshapes itself accordingly.
When a role helps (and when it’s noise)
| Use a role for… | Skip it for… |
|---|---|
| Tone & voice (skeptical reviewer, plain-language teacher) | Raw accuracy — “you are an expert” ≠ correctness |
| Audience calibration (“explain to someone who…”) | Factual / coding tasks where clear instructions suffice |
| Behavioral framing in reusable agents & subagents | One-off prompts where the instruction already says it all |
Impact
- Prompts get shorter and more honest. The effort moves to the part that matters — the instruction — instead of a decorative preamble that reads like work but isn’t.
- Measurable tasks improve for the right reason. Accuracy climbs from context, examples, and an explicit output format. When it climbs, you know why, and you can reproduce it.
- The roles you do keep are the ones worth keeping. In an agent or subagent definition, a behavioral role gets versioned and reused — like the
diff-reviewer— instead of being re-improvised in every chat.
Decisions
- Separate the credential role from the behavioral role. “You are a senior expert” is the first; “you have no stake, find problems, don’t approve” is the second. That line is the whole difference between cargo cult and technique.
- Spend the elaborate roles where they’re reused. A committed agent prompt earns a careful persona. A throwaway one-off doesn’t — just write the instruction.
- To raise accuracy, invest in instructions, context, examples, and output format — not in a persona. That’s where the gains actually are.
Limitations
- This isn’t “roles don’t work.” They work for tone, audience, and framing. What deflates is the role as an accuracy booster.
- The exact effect is model- and task-dependent. On a small or local model, a persona can still earn more of its keep than it does on a frontier Claude.
- A good role won’t rescue a bad instruction. If the task itself is underspecified, no amount of persona fixes it — the role is seasoning, not the meal.