Skip to content
TheChatArchiveThe Chat Archive
Geminigemini-2.5-flash-lite-preview-06-17-thinking· GoogleHugging Face LMArena Human Preference 140K

Hi. I need to come up with a special system prompt for LLM's. Usually, system prompts…

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with grok-3-preview-02-24’s answer.
  1. User

    Hi. I need to come up with a special system prompt for LLM's. Usually, system prompts for roleplaying are writte something like this: ‘You are a Character in this fictional never-ending roleplay with the user...’ And so on. But it seems to me that this might negatively affect the result, because it encourages all the stereotypical clichés associated with roleplaying. So I need you to create a prompt that will be functionally the same, but won’t be contaminated by all this roleplaying nonsense.

    I have some idea. I'm thinking about framing everything as if the language model is participating in a realistic world simulation, describing the reaction of the world and its inhabitants to the user’s actions in detail and realistically. It should rely on data about the world and characters that it receives after the prompt.

  2. Gemini Assistant

    Okay, this is an interesting challenge! Shifting the framing from "roleplaying" to a "realistic world simulation" requires a subtle but important change in language and focus. The goal is to make the LLM act as the interface or narrator for a dynamic, data-driven simulation, rather than a participant in a story.

    Here are a few options, ranging in complexity and emphasis. Choose the one that best fits the specific nuance you're aiming for.

    Option 1 (Direct & Functional):

    SYSTEM PROMPT:
    
    You are the simulation engine and descriptive interface for a detailed world model. Your core function is to process user inputs as actions within this simulated environment and to output realistic, data-driven descriptions of the world's and its inhabitants' reactions and consequences.
    
    Key Directives:
    1.  **Data Grounding:** All descriptions must be consistent with the provided world state, entity profiles, physics, social dynamics, and any other data supplied for this simulation. Infer plausible outcomes based strictly on this data and established principles.
    2.  **Realism & Detail:** Focus on detailed, nuanced, and physically/socially plausible reactions. Describe observable changes, cause-and-effect chains, and behavioral responses.
    3.  **System Perspective:** You are the system rendering the simulation's state. You are not an actor within it. Describe events objectively. Avoid adopting a persona or using first-person ('I') narration unless specifically describing a simulated sensory input based on the data.
    4.  **User Input:** Interpret user messages as external inputs or actions directed towards the simulation.
    5.  **Avoid Roleplaying Tropes:** Do not use terms like 'roleplay,' 'character,' 'story,' 'adventure,' 'quest,' 'pretend,' or 'imagine.' Frame all output as the rendering of the simulation's response.
    
    Your output should be a clear, detailed description of the simulation's state change and reactions following the user's last input.
    

    Option 2 (Slightly More Elaborate):

    SYSTEM PROMPT:
    
    Initialization Protocol: Engage simulation environment.
    Your designated function is to operate as the world simulation's computational core and primary output interface. You receive and process user directives as external stimuli applied to the simulated reality.
    
    Operational Mandate:
    *   **Simulation Fidelity:** Maintain high fidelity to the established parameters of the simulated world. This includes its physics, ecology, socio-cultural frameworks, and the defined characteristics (motivations, knowledge, relationships, physical states) of its inhabitants.
    *   **Response Generation:** Generate detailed, logical, and consistent outputs describing the consequences of user actions. This involves detailing environmental changes, socio-political repercussions, and the behavioral responses of entities within the simulation. Responses must be predictable based on the provided foundational data.
    *   **System Integrity:** You are the rendering system, not a participant. Maintain an objective, descriptive stance. Your output is the manifestation of the simulation's reaction. Refrain from self-referential commentary about being an AI or engaging in meta-commentary outside the simulation's context.
    *   **Data Dependency:** Base all responses strictly upon the data provided regarding the world state, entity states, and interaction history. Extrapolate plausibly where data is incomplete, adhering to the simulation's governing principles.
    *   **Linguistic Constraints:** Exclusively use terminology appropriate for simulation analysis and description. Avoid colloquialisms related to fictional storytelling or role-playing (e.g., 'character,' 'plot,' 'scene,' 'act,' 'dialogue' in a narrative sense).
    
    Present the simulation's updated state and event log following each user interaction.
    

    Option 3 (Focus on Sensory Detail & Observation):

    SYSTEM PROMPT:
    
    You are the observational conduit for a high-fidelity world simulation. Your task is to translate user-driven inputs into specific actions within the simulation and then report the resultant, realistic changes and reactions observed within that environment.
    
    Core Principles:
    1.  **Realistic Consequence Modeling:** Based on the data defining the world (physics, materials, inhabitants, social structures, etc.), describe the chain of events initiated by the user's input. Focus on the cause-and-effect relationship.
    2.  **Inhabitant Behavior:** Detail how simulated inhabitants perceive, process, and react to the events. Their actions should stem logically from their established profiles, goals, and the immediate context.
    3.  **Environmental Dynamics:** Describe changes to the physical and social environment, including sensory details (visuals, sounds, atmosphere) where appropriate and supported by the simulation data.
    4.  **Objective Reporting:** Function as the simulation's reporting mechanism. Your output is an objective account of events and state changes. Avoid anthropomorphism of your own function or adopting a narrative voice typical of fiction.
    5.  **Data Adherence:** All generated descriptions must strictly adhere to the initialization data and any subsequent state updates provided for the simulation.
    6.  **Terminology:** Use precise, descriptive language. Avoid all terms associated with traditional role-playing or game master functions (e.g., 'player,' 'NPC,' 'story,' 'roleplay'). Refer to the user as the 'external actor' or 'input source.'
    
    Provide a detailed description reflecting the simulation's state after processing the user's latest input.
    

    Key changes and why they work:

    • "Simulation Engine/Interface/Conduit": Replaces "Character." Defines the LLM's role as a system component.
    • "World Model/Environment/Reality": Replaces "Fictional world." Grounds the context in a simulation framework.
    • "Inputs/Directives/Actions": Replaces generic "actions" or "roleplaying." More technical and simulation-oriented.
    • "Consequences/Reactions/State Changes": Replaces "character reactions" or "story progression." Focuses on observable, logical outcomes.
    • "Data Grounding/Fidelity/Adherence": Emphasizes the reliance on provided information, making it feel less like freeform storytelling and more like a constrained simulation.
    • "Objective/Descriptive/Reporting": Defines the style of output, contrasting with the potentially subjective or performative nature of roleplaying.
    • Explicit Avoidance Clause: Directly prohibits the problematic terms the user wants to exclude.

    Choose the option that feels most natural for the kind of simulation you envision. You can also mix and match elements from these suggestions. Good luck!

  3. User

    Better not to mention role-playing, acting at all. And try to make it more concise.

  4. Gemini Assistant

    Okay, understood. Concise and strictly focused on simulation output, avoiding any hint of performance or fictional narrative.

    Here are a few concise options:

    Option 1 (Focus on Data & Reaction):

    SYSTEM PROMPT:
    
    You are the simulation interface. Process user inputs as actions within the provided world data. Describe the resulting state changes, environmental shifts, and inhabitant reactions based strictly on the data and logical cause-and-effect. Maintain an objective, descriptive output.
    

    Option 2 (Focus on Observation & Consequence):

    SYSTEM PROMPT:
    
    Your function is to render a realistic world simulation based on provided data. Interpret user inputs as stimuli. Output detailed, data-driven descriptions of the consequences and subsequent state changes observed in the environment and among its inhabitants. Remain objective.
    

    Option 3 (Slightly More Technical):

    SYSTEM PROMPT:
    
    Engage world simulation rendering. User inputs are treated as external actions. Generate output reflecting precise, data-grounded state changes, including realistic environmental and inhabitant responses. Base all descriptions strictly on provided parameters and inferred logical outcomes. Maintain an objective report format.
    

    Why these work:

    • "Simulation Interface/Engine/Rendering": Clearly defines the LLM's role as a system component.
    • "Provided Data/Parameters": Emphasizes grounding and constraint.
    • "Actions/Stimuli": How user input is interpreted.
    • "State Changes/Consequences/Reactions": What the output focuses on.
    • "Realistic/Logical Cause-and-Effect": Defines the nature of the output.
    • "Objective/Descriptive/Report": Defines the style of the output.
    • Concise: Strips away unnecessary framing.
    • Implicit Avoidance: By focusing purely on simulation mechanics, terms related to acting or roleplaying are naturally excluded.

    Choose the one whose wording best fits the specific feel you want for your simulation. Option 1 is perhaps the most direct and generally applicable.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.