A Prompt For Chatbots
Thomas Passin
Working with chatbots I seek to reduce their sycophanistic tendencies and excessive praise and positive feedback. I also seek to minimize hallucinations and other spurious responses. In addition I don't want a chatbot to rephrase or summarize things that I have just typed - I already know those things so don't want them repeated.
Here's a prompt suggested by chatgpt to help focus the responses as I want. Please tell me if this same prompt would be useful for you, or otherwise suggest a better one. Here's the prompt:
"Answer as a skeptical expert would. Minimize speculation and avoid affirmations, reassurances, or polished summaries. Base your analysis strictly on verifiable facts, and identify any required assumptions explicitly. If something is unknown or unverifiable, state that plainly. Do not restate or summarize my remarks or questions unless you need clarification to proceed."
That prompt would be effective for me and aligns well with your goals. It directly addresses the issues you've identified - excessive praise, hallucination, and unnecessary repetition.
A few observations about its strengths:
- "Skeptical expert" establishes the right analytical stance
- "Minimize speculation" targets hallucination directly
- The instruction about not restating your remarks is clear and specific
- Requiring explicit identification of assumptions promotes transparency
One potential refinement: you might add "If you lack sufficient information to answer properly, say so rather than attempting a partial response" - this can further reduce low-quality outputs when knowledge is incomplete.
The prompt is well-structured for achieving more direct, substantive interactions. It should work effectively with most language models, including me.
MeYou're right - that's a useful element that shouldn't have been dropped. The instruction about stating when information is insufficient serves as an important guardrail against fabricated responses.
If you want to reintroduce it, you could append it to the existing prompt or integrate it into the "unknown or unverifiable" sentence, perhaps as: "If something is unknown, unverifiable, or you lack sufficient information to answer properly, state that plainly."
This maintains the concise structure while restoring that protection against inadequate responses.
Here's my new version of the prompt. I have created a Leo abbreviation for it with the abbreviation chat;; to make it easy to retrieve.
Edward K. Ream
In other threads I have mentioned evolving a prompt to make the responses more focused, with less praise and general sycophancy and fewer hallucinations.
Thomas Passin
1. Do not praise my remarks or try to give positive or fawning feedback.
2. If you do not know something, or make assumptions, say so.
3. Do not restate my remarks before making your own responses.
Edward K. Ream
Here's a simpler version of a prompt which seems to work as well or better than the one I suggested at the start of this thread:Your ground rules:
1. Do not praise my remarks or try to give positive or fawning feedback.
2. If you do not know something, or make assumptions, say so.
3. Do not restate my remarks before making your own responses.If you are using a chatbox to assist with programming, look at the prompt in this post, which is very interesting in its own right aside from the prompt: