• workshop_dense_100#slide 69

Role-based prompting assigns a named institutional role or persona to the model so that vocabulary, tone, scope, and exclusions are aligned to a duty of care. In the prompt-engineering paper, examples include a clinician, a credit analyst, and a policy brief author; the implementation section also gives 'senior clinical scribe' and 'credit analyst'. The important point in RAIDT is that the role is not cosmetic. It specifies audience-appropriate explanation, escalation norms, uncertainty language, and what the model should not do.

Why it matters is visible across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). A well-specified role improves Responsibility by making limits explicit and by surfacing red flags or abstentions when evidence is thin. It improves Interpretability because explanations are tailored to the actual reader rather than to a developer. Because RAIDT treats run as the unit of governance, the role template should be versioned in a prompt registry and tied to prompt hashes, run logs, and reviewer forms inside a run-level evidence pack. That is how role prompting enters the audit trail.

The papers are equally clear that role prompting does not itself supply provenance; RAG and other controls are still needed when claims must be source-tethered. Accordingly, the score profile for role-based prompting is usually uneven rather than uniformly high. On the RAIDT anchors 1=missing / 3=partial / 5=audit-ready, role prompts can lift a run above missing governance cues, but they are most defensible when combined with citations, version lineage, and adjudicated review. That is why RAIDT treats prompt design as one of the influence methods as governance interventions.

In clinical note summarisation, the system prompt can assign the model the role 'senior clinical scribe'. The prompt instructs it to summarise de-identified consultation notes into a fixed handover format, surface medication or symptom red flags, state uncertainty explicitly, and avoid speculative diagnosis. That role matters because clinicians need cautious, audience-appropriate language rather than developer-oriented exposition. It also creates a concrete review target: reviewers can ask whether the output stayed within scribe duties, used the mandated caveat, and escalated where evidence was incomplete.

If the organisation logs that role prompt, its version, the note identifier, the output hash, and any retrieval context used, the run can be re-examined later. The result is a cleaner handover between prompting and governance: the role shapes behaviour, and the evidence shows exactly how that behaviour was produced.

• 04-RAIDT_Prompt_Eng_V2
• 06-RAIDT_RAG_V1