Role Prompting: 5 Recipes That Ship (2026)

> Quick answer (2026): Role prompting means telling the model who to be before you tell it what to do, like "You are a senior SQL reviewer." It reliably shapes tone, format, and register. It does not reliably make the model smarter. Use roles to control voice and audience; use few-shot examples, chain of thought, or tools to control accuracy. Below are five paste-ready role prompts, each with the failure mode nobody prints.

Both Anthropic's system-prompt guidance and OpenAI's prompting guide tell you to set a role. What they skip: a role changes how the model writes far more than how well it reasons. Assign roles on purpose, for the thing they actually do.

Recipe 1: The domain expert

Set expertise to raise precision and vocabulary.

You are a senior Postgres DBA with 15 years on high-write systems.
Review this schema. Flag anything that will not scale past 10M rows.

Why it works: the role pulls domain vocabulary and the right level of detail to the front. You get "partial index" and "bloat," not "consider indexing."

Fails when: you expect the persona to add knowledge the model lacks. It will not. A confident "senior DBA" hallucinates just as confidently, and now it sounds authoritative. Verify the claims; the title is not evidence.

Recipe 2: The audience, not the expert

Set who the answer is for, not who writes it.

Explain this deploy failure to a backend engineer who has never used Kubernetes.
No k8s jargon without a one-line definition.

Why it works: audience roles control register better than expert roles. "Explain to a CFO" and "explain to a junior" produce genuinely different, useful outputs from the same facts.

Fails when: the audience is so simple the model drops detail you needed. "Explain to a five year old" is cute and lossy. Pick the least-expert reader who still needs the real answer.

Recipe 3: The strict format contract

Use the role as a behavioral contract, not a personality.

You are a JSON API. You return only valid JSON matching this shape:
{"verdict": "pass|fail", "reasons": string[]}
No prose. No markdown fences. No apology.

Why it works: framing the model as a machine suppresses the chatty preamble instruct-tuned models love. "You are an API" beats "please respond in JSON" for compliance.

Fails when: the role alone still drifts on edge cases. A role is not a schema. For anything you parse, pair it with real structured output, the way the structured-output recipes do, so a bad response fails loudly instead of silently.

Recipe 4: The adversary

Assign a hostile reviewer to surface what a helpful one hides.

You are a skeptical security reviewer who assumes this code is exploitable.
Find the three most likely ways it breaks in production. Be specific.

Why it works: the default assistant is agreeable. An adversarial role gives it permission to disagree, which surfaces real failure modes a "helpful assistant" smooths over.

Fails when: it manufactures nitpicks to satisfy the role. Ask for the top three, ranked, with a concrete trigger for each. A cap forces prioritization instead of theater.

Recipe 5: The voice, on purpose

Use a persona for style transfer, and know that is all it is.

Rewrite this changelog as a terse senior engineer. Short sentences.
No adjectives. State what changed and what breaks. Nothing else.

Why it works: personas are excellent at tone. Style, length, and formality transfer cleanly. This is the thing role prompting is genuinely best at.

Fails when: you expect the terse-senior voice to also be more correct. It changes the wrapper, not the content. Voice and accuracy are separate dials.

When role prompting is theater

The consensus says "you are an expert" makes the model perform better. The evidence is thinner than the advice.

Multiple 2024 and 2025 studies found that adding an expert persona to the system prompt does not reliably improve accuracy on reasoning and knowledge benchmarks, and sometimes nudges it down. PromptHub's roundup of the persona research is a fair summary: personas move tone reliably and accuracy barely.

So spend roles where they pay:

• Voice, register, and audience: yes. This is what roles do best.
• Format compliance: yes, as a contract, backed by a schema.
• Accuracy and reasoning: no. Reach for few-shot examples, chain of thought, or a tool call instead.

On modern instruction-tuned and reasoning models the persona effect is smaller still. Assign the role for what it does. Do not pay it to do what it cannot.

Cost to test: $0.01 to run all five roles once on a small 2026 model.