Roleplay Prompts: Wearing Different Hats

Most people try “pretend you are X” and stop there. You’ll go further. A role is more than a costume; it’s a bundle of constraints that shape what the model notices, how it argues, and what it delivers. In this upgraded guide you’ll turn loose “roleplay” into role engineering—small, reusable components that sharpen tone, focus, and usefulness.

We’ll build a compact mental model, then turn it into copy-paste scaffolds. You’ll see how adding mandates, stance, evidence rules, and format contracts makes answers both clearer and safer. Finally, you’ll run a short lab that compares hats and shows the impact in minutes.

A good role has five parts. Think of them as a Role Card you hand to the model.

• Who: the hat, audience, domain boundary.

• Stance: how to reason (optimistic vs skeptical, planner vs critic).

• Mandate: the job to do and the success criteria.

• Evidence policy: what to cite, what to defer, how to flag uncertainty.

• Format contract: headings or schema the answer must follow.

💡 Insight: Role ≠ voice. Stance and evidence policy do the heavy lifting for quality, while format enforces consistency. Voice can be thin; the other three can’t.

Question: “Is Naples a good weekend destination?”

Below are engineered hats—notice how each adds stance, evidence, and format.

Travel Agent (planner stance)

Historian (analytic stance)

City-Break Critic (comparative stance)

⚠️ Pitfall: “Be X” without a mandate becomes cosplay. Add a job and a yardstick.

1) Reusable System “Policy” (drop-in once per session)

This keeps hats disciplined without being heavy.

2) Role Card Template (fill the braces)

Use this to author consistent hats in seconds.

3) Task Wrapper (the per-question prompt)

TASK: {{QUESTION_OR_INPUT}}
CONTEXT: {{ANY_LIMITS_OR_ASSUMPTIONS}}
OUTPUT: Follow the Role Card’s format exactly. Keep to {{WORD_BUDGET}} where stated.

4) Self-Check & Tighten (automatic second pass)

Ask the model to audit itself against the card, then rewrite.

5) Role Stack (primary + secondary lens)

When one hat isn’t enough, add a lightweight secondary lens.

💡 Insight: A critic+historian stack yields decisive picks that still respect context. A planner+coach stack keeps logistics sharp while minding user constraints.

Keep it small and sharp. Each hat below includes a stance and a signature move.

• Coach (Socratic stance): asks 1–3 questions first; then offers a 3-step plan.

• Auditor (skeptical stance): finds violations against a brief checklist; outputs a risk score.

• Tutor (diagnostic stance): explains in the learner’s words; ends with a tiny quiz and answer key.

• Product Manager (prioritization stance): trades off scope vs time; outputs a MoSCoW list and next 3 actions.

• Safety Officer (precautionary stance): enumerates failure modes; proposes layered mitigations.

Use the Role Card for each. Swap stance and mandate to change behavior without rewriting everything.

There are dozens of knobs. These five deliver the biggest gains fast:

1. Mandate clarity: “Do X by Y standard” beats “Analyze.”

2. Evidence policy: define what must be cited or flagged.

3. Format contract: headings or schema you can eyeball for compliance.

4. Word budgets: force concision (“≤80 words” for summaries).

5. Stance toggles: optimistic ↔ skeptical, planner ↔ critic, teacher ↔ examiner.

💡 Insight: Word budgets and format contracts reduce “answer entropy.” The model fills the boxes you draw.

If results feel theatrical, your stance is too vibe-heavy. Switch to “planner” or “skeptic” and add a mandate with a metric. If the model hedges forever, your mandate lacks a decision rule; add “make a call using these criteria.” If it fabricates, your evidence policy is weak; add “flag uncertainty; do not invent.” If it rambles, your format contract is slack; add headings plus word budgets. If it becomes stiff, relax one lever—usually the budget or the number of sections.

⚠️ Pitfall: Over-stacking roles can create contradictions. Keep one driver and one lens, and specify which breaks ties.

If your downstream tool expects structured data, use a schema instead of headings.

💡 Insight: A schema doesn’t just format output; it prunes the search space, which improves fidelity and reduces hallucination.

Roleplay becomes reliable when you move from “be X” to a Role Card with stance, mandate, evidence policy, and a format contract. These levers reduce ambiguity and channel the model’s attention. You saw how the same question transforms under three hats: the planner commits to logistics, the historian guards context, and the critic makes a call.

The trade-off is flexibility. Strong constraints boost precision but can flatten voice. Loosen word budgets or add a secondary lens when you need nuance. Watch for fabrication: clear evidence rules and uncertainty flags keep you out of trouble.

As you adopt Role Cards, you’ll see sharper tone, fewer rewrites, and outputs that fit downstream workflows. That’s the real impact: less entropy, more decisions, faster.

Next steps

1. Build a 5-hat bank for your domain using the Role Card template; keep each to ~6 lines.

2. Add the Self-Check pass to your default prompt chain; measure acceptance rate before/after.

3. For one task, try a role stack (e.g., critic + historian) and write a one-sentence tie-breaker rule; compare results.