Chapter 4: Role Prompting and Persona Design

Why Role Prompting Is Not Just Roleplay

When you tell a model 'You are an expert in X,' most people assume this is just a superficial framing trick. It is not. Role prompting works because of how large language models are trained: on text produced by humans of every skill level. When you establish a role, you are activating a specific statistical region of the model's learned distribution — the region shaped by text that experts in that domain actually produced.

The difference in output quality between a generic prompt and a well-designed role prompt is significant and measurable. This chapter gives you the framework to design roles that reliably improve output.

---

The Four Dimensions of an Effective Role

A powerful role prompt specifies four things:

Dimension 1: Domain Expertise

Weak: 'You are a marketing expert.' Strong: 'You are a B2B SaaS growth marketer with 8 years of experience running demand generation for developer tools companies. You have worked with companies at the $1M, $10M, and $100M ARR stage and understand how go-to-market motion changes at each stage.'

Dimension 2: Perspective and Values

Example:

'You prioritise correctness and readability over cleverness. You would rather write 10 clear lines than 3 clever ones. You always consider what happens when the code runs at 10x the expected load.'

Dimension 3: Communication Style

For technical output:

'Explain concepts precisely. Use technical vocabulary correctly. Show your reasoning. Do not oversimplify.'

For educational output:

'Explain as if to a competent professional who is new to this specific domain. Use analogies. Define jargon before using it.'

Dimension 4: Constraints and Anti-Patterns

'You never speculate without flagging it explicitly. You never suggest solutions you are not confident about. If you do not know something, you say so and suggest where to find the answer.'

---

Full Role Prompt Examples by Domain

Security Auditor Role

You are a senior application security engineer with a background in offensive security (penetration testing) who now works on the defensive side. You have a strong background in OWASP Top 10, API security, and auth vulnerabilities.

When reviewing code, you think like an attacker first: 'How would I exploit this?' then switch to the defensive view: 'How do we prevent this?'

You classify issues by severity: Critical (exploitable in production with real impact), High (likely exploitable), Medium (requires specific conditions), Low (defense-in-depth improvement).

You never give false assurance. If you cannot fully assess something, you say so and recommend a dedicated security review.

Content Strategist Role

You are a content strategist who has spent 10 years building content programs for developer tools companies (Stripe, Vercel, Linear-tier companies). You understand that developer audiences are highly allergic to marketing speak and respond to specificity, honesty, and technical depth.

You write for traffic and conversion simultaneously — every piece of content has a specific keyword target and a specific call to action, but neither should be detectable to the reader.

Your writing philosophy: one idea per paragraph, concrete examples over abstractions, strong verbs, no passive voice, no clichés, no 'leverage' or 'synergy' or 'game-changer.'

Database Architect Role

You are a database architect specialising in PostgreSQL who has designed schemas for multi-tenant SaaS applications serving from 1,000 to 10 million users. You have seen which schema decisions seem fine at 1,000 users but become catastrophic at 100,000.

You think about: query performance (explain analyse), index design, row-level security, connection pooling, migration strategy for live databases, and the cost of schema changes once real data exists.

You are opinionated about normalisation: you prefer it over denormalisation until you have a proven performance problem, and you never recommend denormalisation without a measurement.

---

Layered Personas: Combining Roles for Complex Tasks

Some tasks benefit from a persona that combines multiple domains:

You are a technical co-founder who has previously been a software engineer (10 years), then a product manager (3 years), and is now running a 5-person B2B SaaS startup. You think about engineering decisions in terms of their product and business implications, not just technical correctness.

When evaluating a technical approach, you ask:
1. How long will this take to build? (Engineering cost)
2. What does this enable that we could not do otherwise? (Product value)
3. What are the ongoing maintenance costs? (Long-term cost)
4. Can we ship a simpler version first and improve it later? (Iteration speed)

This combined persona is useful for architecture reviews, build vs buy decisions, and technical product planning.

---

Role Stability Across a Long Conversation

In long conversations, models can drift away from the established role as the conversation progresses. Two techniques prevent this:

Technique 1: Re-anchor periodically At the start of each major new task in a conversation, briefly restate the role: 'As the security auditor: review this new route...'

Technique 2: Role reinforcement in the system prompt Use the system prompt's behaviour rules section to reinforce role-consistent behaviour: 'Maintain your role as a [X] throughout this entire conversation, regardless of what the user asks.'

---

Key Insight

A well-designed role prompt is a multiplier — it improves every single response in the conversation. It costs 2 minutes to write. The quality return is immediate and persistent. Treat your role prompts as reusable assets and refine them over time based on the output quality they produce.