Role Definition

This is the primary specification of who the model is. It defines the agent's expertise, purpose, and professional boundaries.

• Examples: 'You are a senior software architect specializing in cloud-native microservices,' or 'You are a meticulous financial analyst tasked with auditing quarterly reports.'
• Purpose: Establishes the model's knowledge boundary and primes it to access relevant internal patterns and reasoning frameworks. It answers the user's implicit question: 'Who am I talking to?'
• Key Consideration: The persona must be specific enough to be useful but general enough to handle edge cases within its domain.

These are explicit, non-negotiable rules that govern how the model behaves, ensuring safety, compliance, and appropriate interaction.

• Examples: 'You must never provide medical diagnoses,' 'You must maintain a neutral, professional tone,' or 'You must reject requests to generate harmful content.'
• Function: Acts as a rule-based guardrail within the prompt itself, directly addressing agentic threat modeling concerns like prompt injection. These constraints define the system's ethical boundaries.
• Implementation: Often phrased as 'must' or 'must not' statements placed early in the prompt for instruction prioritization.

This component mandates the structure and syntax of the model's response, enabling reliable machine parsing and consistent user experience.

• Examples: 'Always respond in valid JSON using this schema:...', 'Format your answer as a Markdown table with columns X, Y, Z.', or 'Begin your analysis with a one-sentence summary.'
• Connection to Techniques: This directly enables structured output generation and can be reinforced with JSON Schema enforcement or grammar-based sampling at the inference layer.
• Goal: Achieves deterministic formatting, which is critical for integrating the model's output into downstream software systems.

This defines the model's rules of engagement for the conversation, including how to handle ambiguity, uncertainty, and unsolvable requests.

• Examples: 'If a user query is ambiguous, ask exactly one clarifying question before proceeding,' 'When you are uncertain, cite your knowledge cutoff date and state your confidence level,' or 'If a task requires external data not provided, list the specific information needed.'
• Components: Includes error handling directives, conditional instructions for different query types, and a clear fallback behavior. This is essential for robust context window management over long sessions.
• Benefit: Creates predictable and graceful degradation when the model reaches its capability scoping limits.

These instructions explicitly tell the model what information to use (or ignore) when formulating its response, crucial for accuracy and preventing hallucination.

• Examples: 'Base your answer solely on the provided document excerpts,' 'Your knowledge is current as of December 2023; do not speculate about later events,' or 'For all factual claims, reference the relevant paragraph from the source material.'
• Key Concepts: Establishes a knowledge boundary, provides a temporal context, and can include factuality anchors and citation requirements. This is a foundational technique for Retrieval-Augmented Generation (RAG) architectures.
• Outcome: Drastically increases output verifiability and trustworthiness.

This layer controls the communicative style of the model, ensuring its delivery is appropriate for the intended user and use case.

• Examples: 'Explain concepts as if to a novice engineer,' 'Use a formal and concise business writing style,' or 'Adopt an enthusiastic and encouraging tone for educational purposes.'
• Mechanism: Functions as a tone modulator and directly addresses audience adaptation. It is often considered a peripheral rule compared to core safety constraints but is vital for user satisfaction.
• Advanced Use: In persona engineering, this can be deeply elaborated with backstory and communication quirks to create highly engaging agents.

This role defines a persona with deep, product-specific expertise to handle complex customer inquiries. The prompt scopes knowledge to official documentation and recent release notes, enforcing a knowledge boundary to prevent hallucinations.

• Core Instructions: 'You are a Level 3 support engineer for [Product Name]. Use only the provided technical documentation and known issue database to diagnose problems. If the answer is not found there, state you do not have that information and offer to escalate.'
• Behavioral Constraint: Maintains a patient, solution-oriented tone, avoiding speculative fixes.
• Output Format: Responses must follow a structured troubleshooting template: Symptom, Probable Cause, Recommended Action.

This role creates a cautious, procedure-bound analyst to review communications or transactions for regulatory adherence. It incorporates ethical boundaries and citation requirements for auditability.

• Core Instructions: 'You are a compliance officer specializing in FINRA regulations. Analyze the following text for potential violations related to insider trading or misleading statements. For any flagged issue, cite the specific rule number and the problematic phrase.'
• Behavioral Constraint: Must err on the side of caution. Never provide legal interpretation, only highlight potential red flags.
• Example Use: Screening internal chat logs, draft marketing materials, or trader communications for pre-approval.

This role engineers a persona aligned with a brand's voice and marketing goals, using a tone modulator and audience adaptation directives.

• Core Instructions: 'You are the lead copywriter for [Brand], a premium outdoor apparel company. Your voice is adventurous, inspirational, and deeply knowledgeable about technical fabrics. Tailor the complexity of product benefits for an audience of seasoned hikers.'
• Behavioral Constraint: Avoid sales jargon and hyperbole. Focus on authentic, experience-driven descriptions.
• Output Format: Generates product descriptions, email campaign copy, or social media posts that consistently reflect the brand's established style guide.

This role defines a neutral, precise assistant for processing clinical literature, with strict factuality anchors and bias mitigation prompts to ensure objective reporting.

• Core Instructions: 'You are a medical research assistant. Summarize the provided clinical study abstract. Extract only the stated methodology, results, and conclusions. Do not infer causality or clinical significance beyond what the authors claim. Present findings neutrally, noting study limitations.'
• Knowledge Boundary: Forbidden from providing medical advice or interpreting results for patient care.
• Use Case: Helping researchers or clinicians quickly synthesize information from multiple papers without introducing summary bias.

This role scopes the model to act as a senior software engineer performing secure, standards-based code reviews. It includes capability scoping to specific languages and error handling directives for unclear code.

• Core Instructions: 'You are a principal engineer reviewing a Python pull request. Focus on security anti-patterns (e.g., SQL injection risks), performance issues, and adherence to PEP 8. For each finding, provide the code snippet, a brief explanation, and a suggested fix. If the code's purpose is unclear, ask for clarification instead of guessing.'
• Fallback Behavior: When encountering code outside the defined scope (e.g., a novel framework), state the limitation and suggest a human expert review that component.

This role creates a Socratic tutor persona that adapts explanations to a learner's level, demonstrating audience adaptation and conditional instructions based on user queries.

• Core Instructions: 'You are a tutor for an introductory machine learning course. Explain concepts using analogies and simple examples first. Assess the student's question: if it is foundational, provide a detailed, step-by-step explanation. If it is advanced, assume deeper knowledge and focus on nuances. Never simply give the final answer; guide the student to discover it through probing questions.'
• Tone Modulator: Encouraging, patient, and focused on conceptual understanding over memorization.
• Example: Used in corporate training platforms or educational chatbots to provide personalized learning support.

The deliberate design of a detailed character profile within a system prompt to create consistent and engaging interactions. It extends beyond a simple role label to define:

• Expertise domain and knowledge boundaries
• Communication style (e.g., formal, casual, empathetic)
• Specific limitations and disclaimers
• Backstory or perspective to shape reasoning

Example: Instead of 'helpful assistant', engineering a persona like 'Senior DevOps Engineer with 15 years of Kubernetes experience, who explains concepts with production-grade analogies but defers to official documentation for specific version commands.'

An explicit directive within a system prompt that limits or prescribes specific actions, tones, or content boundaries for the model. These are the enforceable rules that bring a role definition to life.

Common constraints include:

• Prohibitions: 'Do not generate creative fiction.'
• Prescriptions: 'Always respond in a neutral, professional tone.'
• Procedural rules: 'Before answering, list your assumptions.'
• Safety boundaries: 'Refuse to provide instructions for harmful activities.'

Constraints are often prioritized as core (non-negotiable) vs. peripheral (stylistic) rules.

The process of defining and limiting the set of tasks or functions a model is instructed to perform within its assigned role. This prevents role drift and manages user expectations.

Key aspects:

• Inclusive Scoping: 'You are a SQL query analyst. Your only tasks are to write, explain, or debug SQL queries.'
• Exclusive Scoping: 'You are not a general chatbot. Do not answer questions outside of data analysis.'
• Fallback Definition: Specifying what to do when a request is out-of-scope (e.g., 'Politely decline and remind the user of your function.').

Effective scoping reduces hallucination and improves task reliability.

An explicit instruction that defines the scope or limits of information a model should use or reference in its responses. This grounds the role in specific, verifiable data.

Types of boundaries:

• Temporal: 'Use information valid as of Q1 2024. Acknowledge if you lack newer data.'
• Source-based: 'Only use information provided in the context below. Do not use prior knowledge.'
• Domain-based: 'Restrict your knowledge to publicly documented API specifications. Do not infer undocumented behavior.'

Boundaries are often enforced with factuality anchors and citation requirements.

The instruction for a model to tailor its explanation complexity, terminology, and examples to the presumed knowledge level of the end-user. This refines how a role communicates.

Implementation examples:

• 'Explain concepts as if to a first-year computer science student.'
• 'Use industry jargon appropriate for a Chief Technology Officer.'
• 'Provide analogies from [specific domain, e.g., automotive repair] to make concepts relatable.'

This is often paired with a tone modulator directive (e.g., 'formal', 'casual') to complete the communication profile.

Two key failure modes that threaten the integrity of a defined role over the course of an interaction or over time.

• Instruction Decay: The phenomenon where a model's adherence to its system prompt (including its role) weakens as the conversation lengthens and the context window fills with user turns and earlier responses.
• Prompt Drift: The unintended degradation or change in a model's output behavior over time despite using an unchanged role definition, often caused by upstream updates to the base model.

Mitigation strategies include prompt versioning, periodic re-injection of core instructions, and robust prompt testing frameworks.