Masking Prompt Content from Users
I'm not sure if this is on your roadmap (or if it even makes sense). But as I build out our instance, I would love for the ability to "hide" prompt content the same way we're able to do for characters. So for example they can click on the prompt title (which will have a detailed description) and that instruction gets sent to the agent but the details of that prompt isn't visible to the user.
I love the fact that I can maintain my character programming private as IP, it would be fabulous to be able to do the same for my prompts.
Log in to comment and vote
Comments6
Nov 4, 2023
Closing this as resolved.
Tony Dinh
Sep 20, 2023
Why can't you use AI characters to hide the prompt? It's designed for that exact purpose.
Do note that all LLM are also subject to prompt injection no matter how good you hide it.
Bronze Fox
Sep 20, 2023
Tony Dinh: I use the characters and prompts in conjunction. So a user may take a character (eg a content strategist) and then apply a prompt that directs them to write an email campaign. Ideally it would be great for the user to do both without the programming available. (so an option on the admin panel to "hide" prompt content" that can be toggled on or off. It makes it a bit cleaner so that characters are only characters and prompts are only prompts. And fully appreciate the prompt injection risk.
Tony Dinh
Sep 20, 2023
I see, thanks for sharing.
In this case, I think you can put the majority of your instructions in the character’s system instruction. Keeping the prompt minimal like “Use your knowledge and skills to write an email about {subject}”
Prompts are designed as “message templates” and have variables that allow the users to fill in their own values. It will be very confusing if the user can't see the prompt text they selected.
Bronze Fox
Sep 20, 2023
Tony Dinh: Ah got it, ok that makes sense. In that case, one consideration is to allow users to change the title of "Characters" as it appears in the UI to something admin generated might be useful (per Stef's use case). So this would allow the admin to have "Proprietary Prompts" and "User Prompts" or whatever. Otherwise users might get confused bc they're expecting a persona whereas the admin might be specifying a workflow.
My thinking was actually that the "show prompt" "don't show" prompt would be a toggle. So the admin can decide if it's hidden or not. But I totally get how that would be confusing and not appropriate for your vision on the UX.
Out of curiosity, how do the training system prompt and character prompt get injected into the chat. Do they basically get added like a two part system prompt - I'm thinking in an OpenAI playground context. Would the analogy be like I put the training system prompt and my character prompt in that box one after the other?
Sapphire Horizon
Sep 15, 2023
This not only makes sense, but it's mission-critical for us: As we're developing an "expert system" for life coaching, the uniqueness of our prompts is a significant part of our competitive edge. However, the current setup potentially allows competitors to access our system, trigger prompts, and replicate them. This is a serious concern, especially considering our substantial investment of over $20,000 in relevant trademarks and our commitment to safeguarding our intellectual property. I believe it's crucial to address this to ensure the protection of our proprietary information. I look forward to seeing how this concern will be handled in the roadmap for CUSTOM users.