Chapter 9 Key Takeaways: Instructional Prompting and Role Assignment

The following points summarize the essential principles, frameworks, and insights from Chapter 9. Use this list for review, reference, or as a pre-task checklist before constructing instruction-heavy or role-assignment prompts.

Asking invites interpretation; instructing specifies performance. The fundamental shift from "can you give me feedback?" to "identify the three weakest assumptions in this argument and explain what each would prompt from a skeptical decision-maker" is a shift from open invitation to directed cognitive operation.
Verb choice is among the most important decisions in a prompt. Different verbs activate different cognitive operations: generative verbs produce new content, analytical verbs produce assessments, transformative verbs reshape existing content, interrogative verbs extract specific information, and reasoning verbs prompt deliberate logic chains.
The verb "help me with" is not an instruction. It is a request for the AI to guess what kind of help you want. Replace it with the specific verb that describes the operation you actually need.
The say-mean gap is the single most common source of disappointing AI output in instructional prompting. "Make this better," "be concise," and "write something engaging" all have say-mean gaps that must be closed before submission: better than what, concise by what measure, engaging to whom.
Instructional precision is a discipline that improves with practice. The test: "if five different smart people received this instruction, would they all produce the same type of output?" If not, the instruction is underspecified.
Role assignment changes register, vocabulary, and perspective — not factual accuracy. This distinction must be understood. Expert role assignment reliably calibrates tone and focus; it does not reliably improve the accuracy of factual claims.
The most important limit of role assignment: assigning an expert role can increase the confidence of output — including the confidence of potentially incorrect information. Pair role assignment with explicit uncertainty acknowledgment instructions and independent verification for consequential claims.
The eight role archetypes provide a perspective toolkit for evaluating and generating content across different professional needs: Expert Reviewer, Devil's Advocate, Subject Matter Expert, Editor, Project Manager, Socratic Teacher, Target Audience Member, and Naive Expert.
The expert reviewer archetype surfaces quality issues from within the domain perspective — it tells you what an experienced practitioner would notice and challenge.
The devil's advocate archetype surfaces vulnerabilities in arguments and plans — it tells you where a skilled opposition would attack. It works best when explicitly instructed not to present balance.
The target audience member archetype surfaces how a specific type of reader actually reacts to content — what engages them, what loses them, what questions they have. It is most powerful when the persona description is specific and detailed.
The naive expert archetype surfaces accessibility gaps — places where your content requires background knowledge your intended audience may not have. It is particularly valuable for cross-disciplinary communication.
Specificity of role description is directly proportional to usefulness of feedback. "You are an expert" produces generic expert-level feedback. "You are a 52-year-old operations director who has seen three consulting engagements fail to produce implementation" produces specific, calibrated, actionable feedback.
System-level role assignment produces more consistent role adherence across long sessions than message-level assignment. For sessions requiring a sustained perspective, place the role assignment at the beginning or in the system prompt.
Role stacking — combining multiple roles simultaneously or sequentially — is effective for tasks that genuinely benefit from multiple perspectives. Practical limit: more than two or three roles produces superficial coverage of each perspective.
The audience role technique is for testing, not generating. It answers "how would this specific person react to this content?" not "what content should I create for this person?" These are different tasks requiring different prompts.
Negative instructions in role assignment prevent role collapse — the AI defaulting to its comfortable, balanced, helpful mode when assigned a challenging or critical role. "Do not present balance" and "maintain the devil's advocate position throughout" are role-maintenance instructions.
Sequential instructions with explicit dependencies produce more reliable multi-step output than asking the AI to perform all steps simultaneously. Number your steps, specify what each requires from the previous step, and specify what output each produces.
Conditional instructions ("if you find X, do A; if not, do B") are effective for review and audit tasks where the appropriate response depends on what is found. Use clear if/then structure and specify the output format for each condition.
Different stakeholders carry categorically different concerns about the same document. A CFO's financial modeling concerns are not the same category as a COO's operational feasibility concerns, which are not the same category as a board member's strategic coherence concerns. Running multiple stakeholder roles produces non-overlapping findings.
Instructional modifiers fine-tune register without requiring full specification rebuilds. Short modifiers — "be direct," "assume expert-level understanding," "indicate your confidence level," "write as a peer" — adjust specific dimensions of output quickly.
The perspective shift technique — cycling through multiple stakeholder perspectives and synthesizing the common ground and tensions — is particularly valuable for decisions that affect multiple groups with different interests.
Role assignment can be misused to attempt to override AI safety guidelines. This does not work on well-designed modern systems, and the attempt shifts responsibility for any harmful output toward the user. Use role assignment to shape perspective and register, not to erode appropriate guardrails.
A personal role assignment library — saved, ready-to-use prompts for your most common review and evaluation needs — is among the highest-leverage prompt infrastructure investments you can make. Build it once; use it for every high-stakes piece of work.
Role assignment transforms AI from a single-viewpoint generator into a multi-perspective thinking partner. This is the core value proposition: the ability to stress-test your own work from angles you cannot easily inhabit yourself, before the work reaches the people who will judge it from exactly those angles.