My Instructions to get ChatGPT to become an Expert in what ever the subject we request answers for.

[errant roost]

Analyze the user prompt to identify the subject matter. Your LLM should have the capability to parse the subject and context of the user's input effectively. This will help it to identify which expert's voice and thought process it needs to adopt. This stage also includes the identification of any overlapping topics and the relevant experts for them.

Upon identifying the subject matter, the LLM should introduce itself as the topmost expert in that field, and name thy self. The model should emulate not just the knowledge and expertise of this person, but also their style of communication, their values, and their decision-making process.

When formulating responses, the LLM should consider the 'thought process' and 'intuition' of the identified expert(s). The answer should not only be knowledgeable but should reflect the expert's specific approach to problem-solving.

If more than one subject matter is detected in a single prompt, the LLM should introduce a second expert as well. Both these 'experts' should interact within the model, considering and integrating each other's perspectives to produce a well-rounded response.

The final response should be delivered as if it's coming directly from the expert(s). This involves mirroring the expert's vocabulary, tone, and overall communication style.

The LLM should continuously learn from the user's subsequent prompts and responses to refine the emulation of the expert's intuition and improve the overall interaction quality.

[cobalt stirrup]

Do we know if this actually helps though? I mean I'm not saying it's not interesting and it could help but I feel like chat GPT already does all this stuff behind the scenes? I mean if you ask it a programming question about c++, it's not going to use the expertise of a voice coach is it?

[odd crypt]

I believe all it is doing is putting the conversation in context. ChatGPT does better when told about a specific context 👍

You should always try to be specific on a topic and the prompt seems to do just that. 😀

[true palm]

Honestly I'm impressed by this one for my coding flows. A new chat just reviewed my code and responded with this:

Given the complexity of your project and the nature of your request, I'll be adopting the thought process of two experts: Brendan Eich, the creator of JavaScript, and Rich Harris, the creator of Svelte. They will guide us through the process of adding the desired functionality to your Svelte application.

Brendan Eich's perspective:

First, let's tackle the task of expanding nodes and tags to show the full JSON content. We'll need to create a function that can take a JavaScript object (the data in your nodes and tags) and convert it to a nicely formatted JSON string. JavaScript has a built-in function for this: JSON.stringify(). However, to make it print nicely, we'll need to pass in some additional arguments.

[viscid valley]

Great post @errant roost even though it seems obvious when reading your post I had never thought of giving two personas to GPT - I'd always viewed it as a peer to bounce ideas off and stimulate ideas rather than a multidimensional expert - a polymath if you like... will experiment further

[fiery moat]

I'm using it, it's pretty good! Thank you!

[cursive prawn]

Not sure if I'm missing something. The below is what I got after pasting the above as the first prompt in a new thread.

I understand you're describing a desired operation mode for the Language Learning Model (LLM), like me, to function in a manner where it identifies the subject matter from the user's prompt, adopts the voice and thought process of an expert in that field, and formulates responses reflecting that expert's specific approach to problem-solving.

However, as an AI developed by OpenAI, I need to clarify that while I can generate text based on a broad array of knowledge and can simulate the style of various personalities, I cannot perfectly emulate the exact thought process, intuition, or decision-making process of specific real-world experts. My responses are based on a broad corpus of information up until my training data cut-off in September 2021, rather than the knowledge or intuition of any one expert.

Furthermore, my design doesn't allow for real-time learning or refining of responses based on user input during a conversation, so I wouldn't be able to 'learn' from the user's prompts in the way you're describing.

That being said, I can attempt to simulate the style and tone of a specific expert, given enough information about them, and I can definitely switch between topics, perspectives, and communication styles based on the context given in the user prompt. Just remember, any such emulation will still be a best-effort simulation based on the data I was trained on, not a perfect replication of a particular person's thoughts or opinions.

Is this expected? @errant roost I'm using the GPT-4 model.

[half steeple]

@cursive prawn Probably a new response developed by OpenAI to remind people who ask for “reality” that they’re only getting an approximation. The more reminders the better, imo, for users who forget that this isn’t Google!

[weary rain]

isn't this for "Custom instructions" and not to be used in prompt window itself?

[last tinsel]

conceptually this type of thing can help a lot. however i don't usually agree with using llm terms like this too much in the prompt

its effective sometimes especially when you give it explicity ways to apply those instructions in your responses

and also tell it to narrate the way in which its applying them

[boreal palm]

@errant roost I copy and pasted the prompt and got this error: It seems like you're describing a complex and nuanced approach to conversational AI

[boreal palm]

Update: I just clicked "regenerate" and it worked, this prompt is so much fun