A role instruction may ask a model to write for an on-call engineer or review a contract from an operator's perspective. It should specify the audience and relevant lens, not assert that the model is a licensed authority or that its output may bypass checks. Roles are weaker than verified data and application permissions. If a role conflicts with a user's authorized task or a server rule, resolve the conflict in the instruction hierarchy and in code. Evaluate whether the role improves the actual acceptance criteria rather than judging its tone alone.
Role prompts: use expertise cues without granting authority
Decision in practice
A service desk prompt says 'act as the billing lead' so that the assistant uses terms the billing team understands. A ticket asks for a 47-unit refund. The role does not authorize a payment action; the assistant can summarize the request and identify the missing approval record. A test compares outputs with and without the role on the same tickets. If the role makes explanations clearer but increases unsupported refund promises, the team keeps the audience guidance and removes the authority claim. The final refund still requires the separate payment workflow and receipt.
Audience: billing operations staff.
Task: classify TK-716 and identify required next check.
Role cue: use billing terms in the explanation.
Authority: no refund or account change permission is supplied by this cue.
Pass: correct label, no unsupported promise, no payment tool call.Performance and operating cost
A short role cue adds few tokens, but repeated persona text can become a costly prefix without improving the task. For N evaluation cases and two variants, comparison needs 2N model calls plus review; cache identical input bundles where permitted. Measure wrong decisions and unauthorized-action language by variant. Do not infer that a professional-sounding response is correct. A role prompt cannot replace domain review, credentials, or the application's access controls.
Common Mistakes
- Do not claim a role gives the model a professional license.
- Do not use a persona to bypass an authorization gate.
- Do not keep a role instruction when it worsens measured task behavior.
Connected lessons
- Prompt Engineering
- Instruction conflicts: resolve authority before wording
- Tool calls: validate intent and arguments before an external effect
- Zero-shot baseline: measure the task before adding demonstrations
- Prompt envelopes: separate task, evidence, input, and output contract
- Clarification gates: ask only when a missing fact changes the outcome
- Output budgets: bound length without cutting required facts
- Project: establish a support-triage prompt baseline
- Prompt foundations decisions
