Module 02 · Lesson 02

Role-Based Prompting and Audience Targeting

Reading time: 18 minutes Track: Prompt Mastery · Required for all learners Prerequisites: Module 02 · Lesson 01

What this lesson does

Lesson 01 introduced the six-element framework and noted that Role contributes about 3% of the total value of a prompt — the smallest of the six. That's true for the role assignment itself. But role thinking does much more than the simple "you are a X" line at the top of a prompt.

This lesson goes deeper on two related ideas:

Multi-perspective role prompting — using different roles for different parts of a complex task
Audience targeting — explicitly specifying the reader, not just the writer

By the end, you'll be able to:

Use role assignments strategically for tasks requiring multiple lenses
Specify audience characteristics that meaningfully shape output
Recognize when to switch roles mid-conversation versus when to stay in one
Avoid the most common role-prompting failures

This is a tactical lesson. It assumes you've internalized Lesson 01.

01 · The honest case for role prompting

Online prompt-engineering guides massively oversell role prompting. You'll see claims that "You are an expert" magically improves output. That's overstated.

What role prompting actually does:

It biases the output distribution toward content that statistically resembles writing by people in that role. Useful but modest effect.
It constrains vocabulary and tone by activating domain-specific patterns. Useful when domain expertise matters.
It anchors the model's framing of what success looks like for the task. Modestly useful.

What role prompting does not do:

It does not give the model knowledge it didn't have. Telling Claude "you are an oncology pharmacist" does not make Claude actually know oncology pharmacy beyond what's in training data.
It does not make outputs accurate. A role-prompted model still hallucinates.
It does not replace the other five prompt elements. Role without context, goal, format, constraints, and examples still produces weak output.

The right framing: role prompting is a stylistic and framing tool, not a knowledge tool. Use it for that purpose and you'll get value. Expect it to do magic and you'll be disappointed.

02 · Role specificity — three levels

Role assignments range from useless to genuinely powerful. The difference is specificity.

Level 1 — Useless

"You are an expert."

"You are a helpful assistant."

"You are a professional."

These do almost nothing. The model defaults to its averaged "helpful assistant" posture, which is exactly what you'd get with no role assignment at all.

Level 2 — Mildly useful

"You are a regulatory affairs specialist."

"You are a medical writer."

"You are a clinical operations manager."

These help slightly. The model shifts vocabulary and structure toward that domain. But the role is still too generic — it could mean anything from a 6-month associate to a 25-year veteran.

Level 3 — Genuinely powerful

"You are a senior regulatory affairs specialist with 12 years of experience drafting IND amendments for first-in-human oncology trials at small to mid-size biotechs. You've worked closely with FDA's Oncology Center of Excellence on several precision-medicine programs. You write in compliant, audit-ready language and have a strong preference for specific data citation over qualitative claims."

This works. The role is specific along multiple dimensions: seniority, domain depth, sub-specialty, organizational context, stylistic preference. The model has enough to actually anchor on.

The pattern

Specify at least four dimensions:

Title and function (regulatory affairs specialist, biostatistician, medical writer)
Seniority level (junior, senior, director, principal — and ideally years of experience)
Domain specialization (rare disease, oncology, neuroscience, immunology, etc.)
Stylistic or methodological preference (writes in regulatory tone, prefers data-driven claims, focuses on patient impact, etc.)

Each dimension you add tightens the output. Skip dimensions and the output drifts toward generic.

03 · Multi-perspective role prompting

For complex tasks, a single role is sometimes insufficient. You might need the model to produce content from one perspective, then critique it from another.

Example — drafting and critiquing a board memo

Step 1 — Draft from the operator perspective:

"You are a Chief Medical Officer at a clinical-stage biotech. Draft a 600-word board memo updating directors on our Phase 2b trial readout. The data are mixed: primary endpoint met with p=0.04, but a key secondary endpoint did not reach significance. Recommend continuation to Phase 3 with protocol modifications. Tone: measured, transparent, recommending action."

Step 2 — Critique from the board perspective:

"Now switch roles. You are a senior board member with deep biotech operating experience — a former CEO who has seen many trial readouts and is skeptical by default. Read the memo I just produced and identify the three weakest points a board member would push back on. Be specific about what's missing, what's overstated, and what reads as motivated reasoning."

Step 3 — Revise:

"Now switch back to CMO. Revise the memo to address the three weaknesses the board member identified, without becoming defensive or losing the core recommendation."

Why this works

The model can hold multiple personas in a single conversation, and explicit role-switching lets you get more thorough analysis than you would from a single prompt. The "critique" step often surfaces issues you'd miss because you're too close to the work.

When to use this pattern

High-stakes outputs going to skeptical audiences (board, FDA, investors)
Drafts where you want a pre-review before sending to your actual reviewer
Strategy documents where you want adversarial pressure-testing
Any output where "what would the smartest critic say" is a useful question

When not to use this

Routine drafts where the standard production process is fine
Time-sensitive tasks where iterations matter less than speed
Cases where you have a real human reviewer who'll do the critique anyway

04 · Audience targeting — the under-used dimension

Role prompting specifies the writer. Audience targeting specifies the reader. Audience is the more powerful of the two, and it's used far less often.

The contrast

Prompt with role only:

"You are a senior medical affairs specialist. Draft a clinical summary of the Phase 2b data."

Prompt with role + audience:

"You are a senior medical affairs specialist. The audience for this clinical summary is a panel of oncologists at a community hospital network — practicing clinicians, not researchers. They want to know whether to start referring patients to participating sites for our Phase 3, what the safety profile looks like at the recommended dose, and how this regimen compares to current standard of care in practical terms. They have 5 minutes to read this. They are skeptical of pharma-sponsored summaries."

The second prompt produces a fundamentally different output — shorter, more practical, less academic, focused on referral-relevant content.

What to specify about audience

Five questions worth answering for any audience-targeted prompt:

Who specifically reads this? (Role, seniority, domain)
What do they want to do with it? (Decide, inform, act, defer)
How much time do they have? (5 minutes? An hour? A multi-day review?)
What's their default disposition toward the content? (Supportive, neutral, skeptical, hostile)
What's the language register they expect? (Technical-scientific, regulatory-formal, executive-summary, plain-language)

For high-stakes outputs, write a one-line audience description at the top of your prompt. The output will improve dramatically.

05 · The biotech audience map

In biotech work, certain audiences recur. Here's a quick reference for common ones with the specifications that matter most:

FDA reviewer (regulatory submission)

Disposition: Skeptical, looking for completeness and consistency Time: Will spend weeks; checks every claim Language: Regulatory-formal, precise terminology, no marketing language Critical specs: All claims traceable to data; consistent terminology across the submission; no overstatement; clear caveats on limitations

IRB / Ethics committee

Disposition: Patient-safety-focused, looking for risk-benefit clarity Time: Variable; some members read deeply, some skim Language: Medical-scientific with plain-language patient-impact summaries Critical specs: Risk framing explicit; benefit framing measured; patient burden addressed; vulnerable population considerations noted

Internal QC reviewer

Disposition: Looking for errors, inconsistencies, SOP non-compliance Time: Limited; needs efficient review Language: Match company SOPs and previously approved documents Critical specs: Format compliance; terminology consistency; citation completeness; identification of high-risk sections

Executive leadership (CMO, CEO, CFO)

Disposition: Decision-oriented; time-constrained Time: Minutes, not hours Language: Executive summary register; data-light, decision-heavy Critical specs: Lead with conclusion; quantify impact; identify decision needed; note risks and dependencies

Board of directors

Disposition: Strategic; high-stakes; expecting transparency Time: Variable; some members very engaged, others less so Language: Strategic, accountability-focused, transparent on tradeoffs Critical specs: Strategic context; balanced view; explicit risks; clear ask if any; financial implications

Investor (analyst, public market)

Disposition: Looking for material information; legally cautious Time: Varies; analysts read carefully Language: Disclosure-aware, balanced, no MNPI-adjacent claims Critical specs: No selective disclosure; risk factors noted; forward-looking statements caveated; consistent with prior public statements

Field sales rep / MSL

Disposition: Operational; needs talking points and references Time: Limited; reads to use, not to study Language: Practical; references with page numbers; FAQ-style Critical specs: Compliant claims only; clear approved-language vs. off-label boundary; quick-reference structure

Practicing clinician (KOL or general practitioner)

Disposition: Time-constrained; data-driven; evaluates relative to existing practice Time: Minutes per document Language: Clinical; data-forward; practice-relevant Critical specs: Comparison to current standard of care; practical patient-selection guidance; safety profile clearly stated

Patient or caregiver

Disposition: Anxious, motivated, often without medical training Time: Will read carefully but may not retain technical content Language: Plain-language; jargon-free; concrete Critical specs: Plain-language explanation; relatable comparisons; clear next steps; emotional sensitivity to disease context

06 · The role × audience matrix

For maximum effect, combine specific role and specific audience:

"You are a senior medical writer with 10 years of experience in oncology trial documentation. The audience for this section is an FDA reviewer in the Oncology Center of Excellence, reviewing our IND for HLX-201. The reviewer will spend approximately 30 minutes on this section and will cross-check claims against the trial data in our Appendix. They will be skeptical of subjective claims and will flag any inconsistencies between this section and the protocol. Write accordingly."

This pairing — specific role + specific audience + specific verification context — produces output calibrated to actual production conditions. It's measurably better than role alone.

07 · Common role and audience failures

Failure 1 — Mismatched role and task

Assigning "you are a marketing executive" to a regulatory writing task. The role fights the task, producing output that's stylistically wrong for the audience.

Fix: Match role to task. For regulatory work, regulatory roles. For executive memos, executive roles.

Failure 2 — Audience drift mid-conversation

Starting with one audience in mind and shifting to another without telling the model. The output stays calibrated to the original audience.

Fix: When the audience changes, explicitly state the new audience and re-specify what they need.

Failure 3 — Single-audience prompts for multi-audience documents

Drafting a board memo as if there's one audience, when the actual readers include directors with deep operating expertise and directors who are financial-only. The output skews to one and underserves the other.

Fix: Specify the audience as plural with their varying contexts: "The audience is the board of directors. Two members are former biotech CEOs who'll evaluate the operational details closely; three are financial directors who'll focus on capital implications. Address both."

Failure 4 — Patient-facing language from clinical roles

Asking "you are a clinical scientist" to produce patient-facing content. The model gets the clinical register right but loses plain language.

Fix: Use a different role for patient-facing work — "you are a patient educator with experience translating clinical content into plain language for cancer patients."

Failure 5 — Skipping audience entirely

Most common. The prompt has a role but no audience. The output is technically appropriate for the role but doesn't land for the reader. Output looks fine; reception is mediocre.

Fix: Add one line about audience. Five seconds of effort, large effect.

08 · A practical exercise

Take one document you've drafted in the last month — anything from your real work. Look at it and ask yourself:

If I were to redraft this using AI, what role would I assign? Be specific (title, seniority, domain, stylistic preference).
Who specifically reads this? Specify their role, time, disposition, and what they want.
Are role and audience well-matched? Or are they fighting each other?

Now write a one-paragraph prompt header for that document using what you just decided. Save it. The next time you draft a similar document, that header is your starting point.

This is how you build your prompt library — one task at a time, anchored in real work.

09 · Knowledge check

Three questions.

Q1. Which of these role assignments is most likely to produce meaningfully better output?

a) "You are an expert." b) "You are a regulatory specialist." c) "You are a senior regulatory affairs specialist with 12 years of experience in FDA IND amendments for oncology, with a preference for specific data citations over qualitative claims." d) "You are professional and helpful."

Q2. A medical writer is drafting a Section 5.2 for an NDA. The prompt currently specifies the writer's role. What's the single highest-value addition?

a) More detail about the writer's seniority b) Specifying the audience: FDA reviewer, time spent on the section, what they're checking, and their disposition c) A specific word count requirement d) A reminder that the model should be accurate

Q3. When should you use multi-perspective role prompting (e.g., draft as CMO, critique as board member, revise as CMO)?

a) For all routine drafts b) For high-stakes outputs going to skeptical audiences, where adversarial pressure-testing improves the result c) Only when you have hours to spare d) Whenever the document is more than 500 words

Answers: Q1: c · Q2: b · Q3: b

10 · What's next

Lesson 03 — Constraints, guardrails, and format control — goes deep on Elements 4 and 5 of the six-element framework. Constraints are particularly important in biotech, and the next lesson builds a constraint library you'll use for the rest of your career.

End of Lesson 02.