Role Prompting

Role prompting, also known as persona prompting, is a technique in the field of prompt engineering for large language models (LLMs) in which the model is explicitly assigned a specific role, persona, or expert identity before performing a task. In other words, the prompt text or system instructions specify that the model is, for example, a "teacher," "historian," or "pirate," thereby setting the style, tone, and behavior of the generated response.

This approach is widely used in conversational systems. For example, ChatGPT's standard system prompt, "You are a helpful assistant," effectively sets the model's basic role in the conversation.

Purpose and Application

The goal of role prompting is to guide the model toward a specific style and focus for its response, making it more relevant and contextually appropriate for the task. Assigning a role allows the model to adopt a corresponding tonality and vocabulary. This is particularly useful in creative and open-ended tasks, where a role helps make the text more vivid and stylistically diverse.

The technique is also used in professional scenarios. For example, a technical support chatbot can be assigned the role of a polite service representative, while in multi-agent systems, each agent is given its own persona (e.g., "manager," "developer") to facilitate effective collaboration. The flexibility of modern LLMs allows them to take on virtually any role—from a fictional character to a niche specialist—and generate responses that closely match the assigned persona.

Impact on Quality and Accuracy

The effectiveness of role prompting for improving the objective accuracy of responses remains a subject of active research, and the results are often contradictory.

Contradictory Results

On one hand, some studies demonstrate improved results. For example, a study by Zheng et al. (2023) found that if the model's role is thematically aligned with the task, its performance can increase. A paper by Kong et al. (2024) claims that role prompting can improve the model's zero-shot reasoning with a well-chosen persona.

On the other hand, large-scale studies present a more complex picture. In a systematic experiment with 162 different roles, no significant gain in accuracy was found compared to a neutral scenario. Moreover, on average, role prompting even slightly reduced the accuracy of the responses.

The work of Kim et al. (2024) describes role intervention as a "double-edged sword": their experiment showed that with a role instruction, the GPT-4 model correctly answered some questions it had previously gotten wrong (≈15.8% improvement), but a nearly equal share of tasks were, on the contrary, "broken" by the role (≈13.8% degradation). This highlights that adding a role does not in itself guarantee a quality increase and can unpredictably affect the model's behavior.

An undeniable advantage of role prompting is the control it provides over the style and format of the response. Even if the role does not boost factual accuracy, it allows for responses with a specified tonality (friendly, formal, mentoring), which makes the output more coherent and engaging for the user.

Recommendations for Creating Roles

Research offers several practical tips for the effective application of role prompting:

• Role Selection: It is recommended to choose neutral social roles (e.g., "colleague," "mentor"), avoiding overly intimate or highly specialized personas.
• Formulation: It is better to assign the role directly from the model's perspective ("You are an X") rather than through complex imaginary scenarios ("Imagine you are..."). Direct persona assignment has proven more effective.
• Two-Step Approach: For complex queries, it is suggested to divide the task into two steps: first, provide the model with instructions about its role and context, and only then ask the main question. This allows the model to "get into character" first, which improves the stability of the result.

Limitations and Risks

The use of role prompting is associated with several systemic risks.

Reinforcement of Stereotypes and Bias

LLMs inadvertently adopt stereotypes ingrained in their training data. Assigning a role associated with a specific profession, gender, or nationality can activate and amplify these stereotypes in the responses. A study by Gupta et al. (2023) showed that adding social attributes (age, ethnicity) to a role can bias the model's output and significantly reduce reasoning accuracy. Another analysis (Deshpande et al., 2023) found that certain personas lead to increased toxicity in generated responses.

Security Vulnerabilities

Malicious actors can use role prompting to bypass moderation safeguards (a technique known as jailbreaking). Research has shown that assigning a special persona (e.g., a character without ethical constraints) makes it easier to prompt an LLM into performing forbidden actions. Shah et al. (2023) successfully used the cooperation of several "role-playing" agents to bypass model restrictions, demonstrating a systemic vulnerability.

Thus, when assigning a role to a model, it is important to consider the potential biases and undesirable effects embedded in that role and to apply this method with caution.

Links

• Role Prompting: Guide LLMs with Persona-Based Tasks — a guide to role prompting on Learn Prompting.
• Use role prompting with Watsonx and Granite — a practical guide from IBM.

Literature

• Kim, J.; Yang, N.; Jung, K. (2024). Persona is a Double-edged Sword: Mitigating the Negative Impact of Role-playing Prompts in Zero-shot Reasoning Tasks. arXiv:2408.08631.
• Kong, A. et al. (2023). Better Zero-Shot Reasoning with Role-Play Prompting. arXiv:2308.07702.
• Zheng, M. et al. (2023). When “A Helpful Assistant” Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models. arXiv:2311.10054.
• Shah, R. et al. (2023). Scalable and Transferable Black-Box Jailbreaks for Language Models via Persona Modulation. arXiv:2311.03348.
• Deshpande, P. et al. (2024). Evaluating Persona-Prompted LLM Responses in Power-Disparate Health Communication. arXiv:2503.01532.
• Xiong, F. et al. (2025). The Influence of Persona Assignment on Stereotypes and Safeguards in Chinese Large Language Models. arXiv:2506.04975.
• Li, Y. et al. (2025). System Prompts as a Mechanism of Bias in Large Language Models. arXiv:2505.21091.
• Wang, L. et al. (2025). Persona-Assigned Large Language Models Exhibit Human-Like Bias and Toxicity. arXiv:2506.20020.
• Grover, K. et al. (2023). In-Context Impersonation Reveals Large Language Models’ Ability to Simulate Human Personas. PDF.
• Verma, V. et al. (2024). Systematic Survey of Prompt Engineering in Large Language Models. arXiv:2402.07927.
• Gupta, S. et al. (2024). Exploring the Impact of Personality Traits on LLM Bias and Toxicity. arXiv:2502.12566.
• Safdari, M. et al. (2024). LLMs are Vulnerable to Malicious Prompts Disguised as Scientific Language. arXiv:2501.14073.