Skip to content
TheChatArchiveThe Chat Archive
Googlegemma-3-27b-it· GoogleHugging Face LMArena Human Preference 140K

can we add some words in prompt to don't use information from previous prompts or…

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with Claudeclaude-3-5-haiku-20241022’s answer.
  1. User

    can we add some words in prompt to don't use information from previous prompts or information

  2. Google Assistant

    You're right to think about preventing "memory" or influence from previous interactions! While LLMs like Granite don't have perfect memory, they can be subtly influenced by the context of the conversation. Here's how we can add instructions to minimize that, along with a revised prompt:

    Write a business impact statement in 80-100 words. Start with the affected system name and main issue. Focus only on current business process disruptions, not technical configuration problems. **Do not use information from any previous prompts or conversations; base your response solely on the information provided in this prompt.**
    
    Use present tense and active voice. Connect impacts with "causing," "leading to," or "resulting in." Name specific business applications and their direct impacts when known. 
    
    Do not include: technical details, generic statements, urgency language, resolutions, or future impacts. Write as one flowing paragraph for executives.
    
    Input details: [paste incident details here]
    

    Explanation of the addition:

    • "Do not use information from any previous prompts or conversations; base your response solely on the information provided in this prompt." This is a direct instruction to ignore any context outside of the current input.

    Important Considerations:

    • LLM Behavior: This isn't a foolproof guarantee. LLMs are probabilistic, and some residual influence is possible.
    • Prompt Length: Adding too many negative constraints can sometimes reduce performance. We've kept it concise.
    • Testing: The best way to verify this is to test it. Run the prompt multiple times with the same input and see if the responses are consistent. If you see variations that suggest it's drawing on external information, you might need to strengthen the instruction further.

    Further Refinements (if needed):

    If you still see unwanted influence, you could try:

    • Repeating the instruction: "Remember, base your response only on the provided input details."
    • Framing it positively: "Focus exclusively on the information provided in the 'Input details' section."

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