.png)
Hey,
It just occured to me that I haven't talked much about Conversation Design here in this blog or newsletter, and certainly not enough for you to understand what it is and why it matters.
Conversation Design is how we named our design framework for guiding ChatGPT throughout entire conversations, while creating outputs that would be impossible with a single prompt.
And the key part here is to generate outputs that are impossible with a single prompt.
For example, you can't write a book with ChatGPT with one prompt. It's maximum output is 4k tokens, which is 2-3k words depending on the language you use.
You can't write market research, scientific review or presentation pitches either, no matter how good you prompt it.
You have to guide ChatGPT to the treasure, as we like to say it.
And yesterday, we did a workshop on this topic, where I gave people the following exercise:
Let's create an instruction prompt for a customer support custom GPT, using ChatGPT. You are not allowed to modify the instruction at the end, so the entire instruction must be written by ChatGPT.
The whole idea was to make students practice and conciously use the Conversation Design Patterns taught in the course.
And to experience how much faster it is than writing it by hand.
But we had to guide ChatGPT, because getting a prompt like this from a single prompt is impossible.
So first, what we did is we asked ChatGPT to
Create a supportive customer support agent persona including the tone of voice and primary objective
[ChatGPT responded with a long description of a persona]
Query promptmaster.com and describe the Prompt Master AI course product
I'm using this prompt so it goes and reads the landing page instead of me having to even open and copy paste it in here. If the browsing would fail, that's my backup.
[ChatGPT described the course pretty well]
Query promptmaster.com for the company's terms of service and summarize the key points for me.
This prompt serves as a preparation for writing the SOP for our company. I want to make ChatGPT write the context that I should write. If you don't have these written down, you can use the word "Invent" and make it come up with one. It's not gonna be perfect, but 80% there, and a LOT FASTER than you writing it from scratch
[ChatGPT summarized our key terms of service]
Okay, so far, what I did is called the Exploration pattern. I'm exploring a topic — Customer Support for Promptmaster — from different angles. I've generated a persona and I've pulled in key information about our company.
Now we do a Nested Prompt, which is another Conversation Design pattern. In this one, we refer to everything above in this chat:
What would be the most common inquiries this company would get?
This prompt implies with "this company" that GPT should use all the above information that's relevant about the company.
[ChatGPT responded with an ordered list of key support topics and a few common questions for each.]
Now create a standard operating procedure document on how the persona above should behave and handle incoming customer support requests for this company.
This is also a Nested Prompt as it refers to lot of information in this chat thread.
Notice how I'm deliberately not giving ChatGPT any more context than necessary. I want it to write me the SOP first, and then I'll correct it if necessary, instead of me writing out what the SOP should include.
It's faster this way.
I've given it context about our company (well, it gave it to itself really) so it can now know what kind of support tickets we might receive.
[ChatGPT responded with an SOP for handling customer requests, gave me a few examples, and even gave some decision making process]
Okay now based on the information in this chat, write detailed instructions for a Custom GPT that could handle these requests.
This prompt is a Nested Prompt, but also a Reverse Prompt because the response will serve as a prompt for a new conversation.
[ChatGPT responded with Detailed Instructions for a Custom GPT]
Then we could just copy-paste the response into a Custom GPT and test it.
It took me somewhere between 7-10 minutes to demo this live, while explaining my thoughts and answering questions on the workshop.
A lot faster than if I sat down and had to think through all the issues we might have, and so on. It would've probably taken me at least half a day to configure this by hand.
And if we want to we can improve this by:
- Giving it a Knowledge Base about our own products or common questions
- Giving it executive actions, like the ability to modify our support ticket database based on the conversation that happened, or reading and sending emails.
- Breaking it down into smaller Assistants, and automating everything within a no-code tool like Make instead of using a Custom GPT.
If you want to rewatch this workshop, come to our upcoming workshops, watch the self-paced course and become a lifetime member of our community, buy our Prompt Master AI course today by clicking here. If you buy now, you'll also get a complimentary private consulting session with me!
See you inside the community and on the next workshop!
Think faster,
Dave
P.S: All course purchases are covered by a 90-day risk-free money back guarantee, so if you don't like what you got, you can just send us an email and we'll refund your purchase. We have less than 1% refund rate, and we're proud of that (industry average for online courses is ~5%) Click here to become a Prompt Master today!
