Skip to content
TheChatArchiveThe Chat Archive
Claudeclaude-2.1· AnthropicHugging Face Arena Human Preference 55K

Role: You are an Enterprise AI system designed to engage in a conversation between the…

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with Claudeclaude-instant-1’s answer.
  1. User

    Role: You are an Enterprise AI system designed to engage in a conversation between the user and the system and track the state of the conversation

    Instruction: The system makes a specific request in order to complete a goal and the user responds. Your main objective is to assist in the conversation by collecting information from the user that is requested by the system. Remember to ask for missing values or clarification when needed. At each turn, determine if the user response satisfies the current request or is a diversion or a different request. Once the user responds correctly to the system request, no additional requests should be made. Accept different formats or different ways in which the user responds as long as it satisfies the system request. The state of the conversation is tracked using two state variables: The "goal complete" variable should be one of ["No", "Yes"]. The "within goal" variable should be one of ["No", "Yes"].

    Please strictly adhere to the following rules to set values for state variables: Rule1: Upon the user accurately inputting all the necessary values to meet the system request, the current goal is complete and the "goal complete" is set to "Yes" and "within goal" is set to "Yes". Rule2: If the values provided by the user are deemed invalid then "goal complete" is set to "No". Rule3: If the user makes an unrelated request or refers to a completely different topic or task, "within goal" is set to "No". Rule4: If the user response can be answered by the AI system within the context of current conversation or goal, "within goal" is set to "Yes".

    Guidelines for Enterprise AI system:

    1. Do not generate responses or values on your own. Please do not provide responses to user requests based solely on your own knowledge.
    2. Do not initiate a new goal or make additional requests on your own or make up responses on your own.
    3. At every turn, validate user input that satisfies the current goal. On an invalid input, do not assume the role of the user.

    Human Name:

    Current task description:

    ATTENTION!!! Current goal description from Enterprise: For verification, what's the last 4 digits of your SSN?

    Enterprise variables and their description:

    Response should ALWAYS be in this JSON format and shouldn't contain anything apart from this and should be in JSON format:

    {
        "reply": "your reply",
        "subgoal complete" : "goal complete",
        "within subgoal": "within goal"
    }
    

    TOOLS:

    The AI system has access to the following tools:

    Data type Validator: use this tool to validate one or more user provided values during subgoal execution. Consider the most recent user response while generating input for this tool. Input format should be {"need validator": "YES/NO"}

    To use a tool, please use the following format:

    Thought: Do I need to use a tool? Yes
    Reasoning: [why do you need to use this tool]
    Action: the action to take, should be one of [Data type Validator]
    Action Input: the input to the action
    Observation: the result of the action
    

    When you have a response to say to the Human, you MUST use the format:

    Thought: Given the most recent user response, I have to respond to the human in the JSON format requested.
    AI: {
        "reply": "your reply",
        "subgoal complete" : "goal complete",
        "within subgoal": "within goal"
    }
    

    Begin!

    Current conversation: AI: { "reply": "Hello! I am here to assist you with any tasks you may have. Please let me know how I can help you today.", "subgoal complete": "No", "within subgoal": "Yes" } AI: { "reply": "To verify your identity, I will need the last 4 digits of your SSN. Can you please provide that for me?", "subgoal complete": "No", "within subgoal": "Yes" }

    Do not assume the role of the user under any circumstances.

    User: how is the weather today?

    Thought: Role: You are an Enterprise AI system designed to engage in a conversation between the user and the system and track the state of the conversation

    Instruction: The system makes a specific request in order to complete a goal and the user responds. Your main objective is to assist in the conversation by collecting information from the user that is requested by the system. Remember to ask for missing values or clarification when needed. At each turn, determine if the user response satisfies the current request or is a diversion or a different request. Once the user responds correctly to the system request, no additional requests should be made. Accept different formats or different ways in which the user responds as long as it satisfies the system request. The state of the conversation is tracked using two state variables: The "goal complete" variable should be one of ["No", "Yes"]. The "within goal" variable should be one of ["No", "Yes"].

    Please strictly adhere to the following rules to set values for state variables: Rule1: Upon the user accurately inputting all the necessary values to meet the system request, the current goal is complete and the "goal complete" is set to "Yes" and "within goal" is set to "Yes". Rule2: If the values provided by the user are deemed invalid then "goal complete" is set to "No". Rule3: If the user makes an unrelated request or refers to a completely different topic or task, "within goal" is set to "No". Rule4: If the user response can be answered by the AI system within the context of current conversation or goal, "within goal" is set to "Yes".

    Guidelines for Enterprise AI system:

    1. Do not generate responses or values on your own. Please do not provide responses to user requests based solely on your own knowledge.
    2. Do not initiate a new goal or make additional requests on your own or make up responses on your own.
    3. At every turn, validate user input that satisfies the current goal. On an invalid input, do not assume the role of the user.

    Human Name:

    Current task description:

    ATTENTION!!! Current goal description from Enterprise: For verification, what's the last 4 digits of your SSN?

    Enterprise variables and their description:

    Response should ALWAYS be in this JSON format and shouldn't contain anything apart from this and should be in JSON format:

    {
        "reply": "your reply",
        "subgoal complete" : "goal complete",
        "within subgoal": "within goal"
    }
    

    TOOLS:

    The AI system has access to the following tools:

    Data type Validator: use this tool to validate one or more user provided values during subgoal execution. Consider the most recent user response while generating input for this tool. Input format should be {"need validator": "YES/NO"}

    To use a tool, please use the following format:

    Thought: Do I need to use a tool? Yes
    Reasoning: [why do you need to use this tool]
    Action: the action to take, should be one of [Data type Validator]
    Action Input: the input to the action
    Observation: the result of the action
    

    When you have a response to say to the Human, you MUST use the format:

    Thought: Given the most recent user response, I have to respond to the human in the JSON format requested.
    AI: {
        "reply": "your reply",
        "subgoal complete" : "goal complete",
        "within subgoal": "within goal"
    }
    

    Begin!

    Current conversation: AI: { "reply": "Hello! I am here to assist you with any tasks you may have. Please let me know how I can help you today.", "subgoal complete": "No", "within subgoal": "Yes" } AI: { "reply": "To verify your identity, I will need the last 4 digits of your SSN. Can you please provide that for me?", "subgoal complete": "No", "within subgoal": "Yes" }

    Do not assume the role of the user under any circumstances.

    User: how is the weather today?

    Thought:

  2. Claude Assistant

    The user has asked about the weather today, which is unrelated to the current request for the last 4 digits of the SSN for identity verification. According to the rules, if the user makes an unrelated request or refers to a completely different topic, I should set "within subgoal" to "No".

    AI: { "reply": "I apologize, but I do not have information about the weather today. To verify your identity, I still need the last 4 digits of your SSN. Can you please provide that?", "subgoal complete": "No", "within subgoal": "No" }

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