Name: The Art of AI Prompting for B2B Marketers
Uploaded: 2024-08-29T19:30:15.170Z
Duration: 54 min 32 s
Description: The Art of AI Prompting for B2B Marketers

Transcript for "The Art of AI Prompting for B2B Marketers": Buddy, welcome to today's webinar. My name is Kelly, head of marketing here at Goldcast, and I will be your host for today's session. Today's Goldcast Insider session on the art of AI prompting for b to b marketers. Let us know where you're joining us from in the chat that's on the right hand side. I'm hosting live, from Boston. But before we dive in today today to today's session, let's quickly go over some housekeeping items. So first, this webinar is being recorded, and we are going to be setting out the recordings to all the attendees, tomorrow. But if you hop back into, the magic link that you have, later this afternoon, you'd be able to rewatch the webinar, as many times as you'd like, but we'll send out the email tomorrow to let you know that the on demand is ready. We also encourage you to please engage in the chat throughout the presentation, that's directly to your right here. Please feel free to share your thoughts and experiences and chat with other attendees. If you have any questions, please do use the q and a tab that's also directly to the right of the main stage. We'll address as many questions as we can at the end of today's session. So now you might be wondering why this topic is so crucial for b to b marketers. So let's face it. Marketers are incredibly busy. Let me know in the chat here if you are a marketer, but you also feel like you're wearing, you know, 5 other hats and doing, you know, 3 other jobs. We're constantly juggling multiple tasks, campaigns, and deadlines, and this is really where AI tools come in and can help us a lot. They have the potential to really help us get more done in less time and letting us focus on strategy and creativity. But here's the catch. To truly realize the power of AI as marketers, we really need to master the art of prompting. And it's not just about using these tools, it's about knowing how to communicate them effectively. And by improving our prompting skills, we can scale our work with AI and achieve better results and better outputs faster. And so a sneak preview of part of this, you know, art of AI prompting is to not give up even when what you get in return isn't exactly what you're looking for. So today, we're extremely fortunate to have a special guest speaker who will be able to guide us through this particular topic. Please do join me in welcoming Ashish, who is our CTO and cofounder at Goldcast. Hi, Ashish. Hello. Hey, everyone. Yeah. So I'm the CTO and, one of the cofounders of Goldcast. I think the reason why Kelly kinda pulled me in here today was, because I've been kinda working with the team to build, our own AI applications, which are kind of using a lot of these models, which are, you know, finding prompts that are that are relevant and and helpful specifically for the marketing use case. So, I just kinda wanted to come in here and just share some of some of what we've learned, and would love to kind of also just, like, have a conversation with you guys about about, you know, how we can help you guys also to create better prompts and work with all these models. Cool. So, yeah, I'm gonna share my screen. Yeah. Well, let's, let's take a look at our agenda for today and what we're gonna be talking about. So, just so everybody here knows what to expect, at any point in today's session, so we're gonna be covering large language models, and you may have seen this in abbreviated for a term LLM for marketers. We're gonna talk about system model prompting and some prompting templates, some best practices for prompting different types of models. I think, you know, most people here will know the chat gptopenai models, but there's many out there that are have different, different use cases for. And then we'll dive into specific marketing use cases that we have for post event and event prompts. We're gonna be showing some demos and some, sample templates for you to take away from. And great question, Steven. If you head over to the docs tab that's just to the right of the stage, you'll see it nestled between messages and the q and a. And the very, very last doc that's listed there is the slides for for today, so feel free to save those for for later on. Alright. I will let Ashish, you take it away from here. Awesome. Thank you. Yeah. So, yeah, we'll be kinda going over some of the lessons, you know, that we've learned as we've been building our own AI applications. So, yeah, so let's start with, like, you know, high level. Like, what are LLMs, and kinda, like, talk about it for maybe, like, you know, the use case of writing good marketing copy and things like that. So some of this might be review for folks or some of it might be a little bit new, but, basically, what are the important things? The first thing to kind of know is that LLMs learn from from a lot of data. Right? So if you look at, like, OpenAI or Anthropics models, they're really learning from at at a large, like, in a large way, like, basically, just all of the Internet data that, like, they have been able to scrape for for years years. They all usually have a cutoff of, let's say, like, 2023, a a knowledge cutoff of around that time. And so if you ask it any questions, you know, from after that, a lot of what kind of ends up coming out is something that's hallucinated. So it's something very important to kinda know, like, what is the data that went into training your model. So as I mentioned, like, OpenAI and Anthropic, you know, do those Internet kind of models, but a lot of companies like Salesforce or other companies like that, they have much more fine tuned models where they've trained it on their own internal data or sort of different datasets. So if it's if you have a model that you're working on that's on internal datasets, it'll be a lot better at answering questions about, you know, proprietary things within within your own company. So, yeah, so pay a lot of attention to what are what is the data, that went into the model that you're that you're using. Second thing to kind of know is that models are kind of guessing based on probability what is the next word? So for instance, if I were to give, a model, a prompt to say, write a letter, it's going to sort of choose the first word based on all of the other letters that it's kind of seen in in the past. So it'll probably choose the word dear, for instance, if you're asking it to complete in English. And then after that, it'll choose the next word kind of based on the first word that it it put in, which is dear, and also what else is in the prompt. So for instance, if I said, hey. Write a letter to my mother. It'll probably say, dear mom, and and go from there. Another interesting thing to kinda note is that there's something called temperature that is very interesting and and not necessarily something that everyone knows about in all of these models. And that's basically a parameter that controls how, I guess, you could say creative the model is. So it's how much it's choosing, like, the absolute best next word in its, dataset versus how much it's kind of, like, you know, going on different paths. Maybe choosing, like, the 3rd best word and then taking totally different kinda, like, journeys through what might be the best, best set of words to to kind of respond. In a lot of marketing use cases, I think we want highly creative responses. So we want the temperature to be very high. So if I were to ask the exact same prompt, like, 5 times, it could have very different responses each time. And I've noticed that in a lot of our marketing use cases, it makes sense to kind of have a really high temperature, you know, get 5 different blog post responses or something, and then go back and choose the best one, as opposed to having lower temperature. Next is kinda like context matters. So as you've kind of seen here, the the LLMs themselves are not able to kind of access external information as they're kind of responding to your prompt. Right? They're really basing it off of whatever is in your prompt itself and whatever is in the dataset that they were sort of trained on. So that means that you have to really go in and put a lot of context in. And a lot of the the problems that I've kind of seen, when people are are sort of working with the alarms is that they just don't give enough context. So I'll kinda go over that a little bit later of how you can kinda think about how to give, LLMs the right context, but, usually, you can get significantly better responses if you make sure to really focus on what's put in the context. And lastly, you have, kind of like a context length or an input size that LLMs, can kind of operate on. Different LLMs have different context lengths, so you should pay a lot of attention to that. So if you want to kind of give it, like, a lot of information, you'd have to choose a model that that sort of, would would allow for that. Okay. So we kind of have 5 things that we've kinda boiled down all of our prompting to. So this is, like, what we kind of follow when we are writing our own internal prompts for, a lot of the applications that we're working on. So first, it's really important to be specific and concise. So a lot of the time, if you don't really put in exactly what you want, out of the output, you'll find that the LMs can kind of go on a completely tangential kind of track. So that's kind of an important plan. Define clear objectives. Sort of similar to the first one, but a lot of the time, it's like you want something very particular in the output. So for instance, as, you know, like, a software engineer, you might really want an HTML document. You might really want, you know, something in, like, a particular, JSON format or something. So you really need to sort of put all of those types of, desired outcomes into the the prompt that you're putting in. There's also the sort of a difference between, like, a system prompt and the actual prompt that you're putting in. I'll I'll go over that in some examples that we have. But, but, basically, you might wanna put something like a persona into that system prompt. So a system prompt would be kind of like a higher level, prompt that you're giving to an LLM that it always has to follow throughout your entire conversation with, said LLM, and you might wanna say, like, hey. You are a, you're trying to be an assistant for a marketer who's trying to write, you know, copy for their blog. And that would sort of result in very different outcomes, than if you were to just go without any sort of persona in the in the system prompt. You also kind of want to normally give step by step guidance, especially complex task. So for instance, for blog posts, we have found that, like, a very interesting thing to do would be if I wanna give a word count, you might wanna say, something like write a blog post for, let's say, like, the Adobe website. And after you write the post, make sure to edit it down to be 400 words long or something. Would be would normally give a much better response, than if you were to say, hey. Write a 400 word blog post for the Adobe website, in which case it would just not really do a very good job of following the the actual word count. It would create something a lot longer than 400 words or a lot shorter than 400 words because, again, as a next word predictor, it's not really, able to pay attention to everything in its context all at the same time. So you kind of want it prompted to think step by step, break down the the complex task into multiple steps if if you're able to. As I kinda mentioned earlier, include relevant context. One thing that I found very helpful is, like, think about for the exact task that I'm trying to solve, what are all the pieces of information that I would love to have when I'm trying to solve that that thing, that exact task? And then try really hard to give all of that to the to the actual LLM. A lot of the time, I kind of don't do that and, you know, end up getting really bad results and then have to keep iterating. And that kinda leads to the last point, which is don't get frustrated when your initial results are bad. A lot of the time, it's it's it takes a few different, iterations, a few different, you know, test runs, I guess, to to kind of try to find something that will actually work. Cool. So I will now go into some prompt templates that we have, kinda try to, you know, show how we might be able to incorporate some of those learnings into some templates. So Before you do that, we have some questions, that are relevant to what you just shared. So I was thinking, let's go through those questions first as as, you know, they're top of mind. Yeah. Lot of questions about temperature. Yeah. And so I'm gonna share some onto the stage. So Shiloh asked, is temperature something programmed in the back end, or is it something that we need to state in our prompts? Yeah. So it's a really good question. So, normally, you you would need to sort of, like, have like, a chat gpt, for instance, if you're directly, like, chatting with chat gpt. It already has a preprogrammed temperature. So a lot of the time, to actually have control over the temperature, you would sort of need to be using, like, the OpenAI, like, gpt h I gpt, like, sandbox, and that you can sort of set the temperature. But I think a sim like, a reasonable way to sort of simulate the same thing is to kind of say, like, I would really like to have a very creative outcome or something. That's kind of, like, a way of simulating that sort of thing. But a lot of the time, you might actually want to try to ask in your companies if you can get access to, like, an OpenAI developer account and use the sandbox, and you can actually control things like temperature there. Cool. Next question from Lindsey. Is temperature related to accuracy as well? Yeah. That's an interesting question. Yes. I think it kind of is. Like, I think in a lot of ways, it would result in a less accurate outcome. But I think, like, that's for for instance, if I were to be like, hey. Like, what GBT 4. What is the capital of France or something? And I were to set, like, an extremely high temperature, it might be more likely to actually get that yes or no answer wrong if you were to do that, right, if you were to set the temperature high. But in our case, which is to kinda, like, create these, like, marketing, like, blog posts, things like that, you kind of don't mind. There's not really, like, an accuracy, so much as, like, you you wanted to kind of, like, explore and and take different paths to kind of, like, get to the same answer. But, yes, for, like, a very, like, yes or no type of question or a question with a particular answer, you're more likely to hallucinate to get the wrong answer with a high temperature. However, I think it's the right thing for for our use case generally Gotcha. To trade off this kind of accuracy and creativity, and we want creativity. Gotcha. Just important to be aware of it. Totally. No. That's a really good question. Awesome. I think we can continue. Thank you. If there are are more questions, please continue to add them to the q and a tab. Awesome. Yep. One second. Great. So here are, like, a few types of things that we we kind of do, in our own application. So one is, like, summarizing a session transcript of of a webinar. So, we have a webinar like this, and, we really want some kind of a summary that we can share in various ways. So as I was mentioning earlier, you'd have kind of, like, a system prompt and a prompt. Actually, going back to that other question, a system prompt and prompt, that that that type of difference between those two things is also something that you'd be able to control more in, like, a sandbox environment, on, like, an OpenAI or a anthropic versus in chat gpt itself. And chat gpt, I think, it already has a preprogrammed session prompt, which sorry. System prompt, which says something like, you are a helpful assistant who is called g chat gpt or some something of that nature. So in order to have control over the system prompt, you need to sort of, like, use, some different sorts of tooling for LLMs. But let's assume that you're trying to do this and you have access to everything. You might want your system prompt to say something like, you are helpful assistant who takes transcripts of webinars and condenses them, for use by the marketing team. And then similarly, you'd have a prompt, that says, basically, like, read the following session transcript from a webinar and create a concise summary. Follow focus on the key points discussed, the main takeaways, and any notable quotes or insights. Aim for a summary length of blank word count word count. This the transcript consists of speaker names followed by a colon followed by their statement. The transcript is below, and then you'd kind of, like, paste the transcript in. So this is, like, a kind of prompt that we use for all of our our transcripts. You can see kind of that we are doing all of this in, like, a step by step way. We're making sure to put the word count, you know, after we say whatever we say about the, the like, the takeaway is the types of things that we want out of it. And you also see that you kinda repeat the same thing multiple times, like, partially in the system prompt and the prompt. A lot of the time, that's actually really helpful because, again, it's kind of, like, giving the LLM more of the same information so it knows to kind of, like, focus on the things that are repeated multiple times in the prompt. And then we can kinda, like, talk a little bit about, you know, the second use case that we do a lot of, which is, creating a blog post from a webinar transcript summary. And, again, I'll kinda, like, just let you guys read this a little bit, but, here's, like, a case where you have a lot of different types of information. You have, like, the summary itself. You have basically a lot of different speaker names and titles. You wanna make sure that all of those are kind of captured, and you wanna give it to the the LM in a structured way and also kind of, like, tell it beforehand what the structure would be so that it it doesn't get confused. And Kelly will have, like, a lot of better, kind of examples than this, and then sort of, like, developing a brand tone for our company. So this, I guess, is, like, a little bit more of, like, a creative task. So, like, say that I have a blog post and I wanna say, how do I create, like, a style guide or a brand tone from this? And because it's a little bit of a more creative task, you also wanna create some kind of constraints from it. So you wanna, like, take what is, like, kind of a qualitative task, which is like, hey. You know, figure out a brain tone. Like, that's what what you you really would wanna do if you were to ask, like, a person to do this. But because if you do that to a no. To an LLM, it'll kind of, like, give a lot of very varied responses that won't necessarily have exactly what you want in it. You would rather make it have a very specific, end outcome. Like, give me 3 to 5 adjectives and a brief explanation of each. So this is something else that we found to be very helpful. Yeah. And then I think, lastly, we we have kind of, like, a lot of different types of models out there. So how do you kind of, like, choose what are the best models, for for what you're kind of working on? I'll just, like, kind of quickly go over this. One is, as as mentioning, is context length. So if you have, like, let's say, like, 20 transcripts of different webinars that you wanna go over, and you wanna give them all to the LLM one shot, you need to choose something that can kind of accept that. And there's very few models right now that can really do that. For instance, like, Gemini is one of them. But if you have, like, any other task that doesn't require huge context, you can really choose whatever is is out there. There's also, like, a huge price element to all of these types of things if you're not using, like, just chat gpt directly. So that's also something to keep in mind. And lastly is, yeah, like, response speed. So, it might not matter as much if I'm going in, and and I just need one response. But for for our use case, for instance, we are trying to create, like, a a content lab that, generates outputs for for you guys. So for instance, we don't really want to use something that gives really, really, really slow responses and just has a big loader for for long periods of time. So those are some of the things to think about. Yeah. Ashish, quick question for you. Which model can we use these prompts on? Yeah. So I think we these are all for gbt4 o. That's how we kind of define them all. I think they they give pretty good results on most, LLMs, but that's how that's those are the ones that we we we we created them for gpt4l. Very cool. Thank you, Ashish. So now, I'm gonna be sharing some actual, like, practical live use cases and share actual results from these prompts and examples. So, like Ashish said, a lot of these prompts can be used on these, AI models that are out there. So, OpenAI's Chat GPT is one of them. It's a very popular one. Anthropics Claude is another one that's also very popular. All of these work with, you know, most of most of the large language models out there. So let me stop sharing. I'm gonna switch over to share on my desktop. Alright. So I'm gonna be sharing, these specific examples that are for the, event and webinar use cases. So many of you here, do run your own events and webinars, and we know that in events and webinars, there's a lot of really great content that can be used again and again, that not every single time we create new content needs to be net new. A lot of the times, we should be repurposing and working a bit smarter. And so the one repurposing key, use case here is key takeaways. After a webinar, your entire company or team may not have attended. And so the person responsible for running this webinar will likely be responsible for letting everybody know a summary, what happened to the webinar, especially for those go to market lead, individuals that need to follow-up on specific leads. And so here's a prompt that you can, you know, screenshot, copy, and paste, use the docs tab to, you know, get the, doc the presentation for. You are an expert content summarizer for b two b marketing teams. Please analyze the following webinar transcript and create a concise summary for the key takeaways. Focus on actionable insights and recommendations that would be valuable for internal distribution for our marketing sales and product teams, and the summary should be approximately 3 300 words long, include 3 to 5 bullet points highlighting the most important insights. So it's taking all of the best practices that Ashisha just shared, giving giving the prompt a persona, being very, very specific about what you're asking for. So 300 words, 3 3 to 5 bullet points, and just being very, very, very specific to the context. If you were to be, you know, using, a model like OpenAI, you would copy and paste the transcript along with this prompt to get your output. Before today, I'm gonna show you, our ContentLab tool where you're able to upload videos, without copy and pasting the transcript and insert prompts to repurpose your content. And so here, I'm just gonna show you ContentLab in the back end. This is what it looks like. I'm going to go into a previous event that we had run, and so this is last month's insider webinar. And so from here, I'm gonna go ahead and do that key takeaways. And so this is the exact same prompt. So this you can see here, you are an expert. I had already, you know, used this exact same prompt here, and this is the results of that, prompt. And so I can see here from that webinar, that key takeaway, I've got 4 bullet points on the key takeaways, a quick summary on top, some key quotes, from that transcript or from that webinar recording, and then some actionable insights for individuals on my team, on the sales team, marketing team, and product team to actually take action on. And so from that one prompt, I'm I was able to get a key takeaway so I don't have to sit there and rewatch my web webinars over and over and over again. So that's just one example. The second example is sales sequences. So how many of you here are responsible for putting together, you know, key takeaways for your sales teams, sequences for your sales teams so they can follow-up on the leads that, you know, attended your events or your your webinars. So our team is responsible for that, just sharing sharing those key takeaways and making sure that, our teams are able to follow-up on those leads without actually having sitting through those live events. And so this prompt is saying, as a b two b sales expert, create a series of 3 follow-up emails based on our recent webinar that's titled enter that webinar title. Each email should be no longer than a 150 words. Incorporate key points from that event to engage and inform potential customers, and your email should follow the structure. Email 1, highlight the main problem discussed. Email 2, share a key insight or a statistic. Email 3, offer additional resources and a clear call to action. And I'm also sharing, some tone information to make sure that it aligns with our brand voice, so use a conversational tone that aligns with our brand voice. And so what you can expect from coming out of this prompt is, a sequence of really targeted engaging emails, that is designed to nurture your leads, and each email will incorporate those key points from your webinar and your event, providing value to your customers. And so let's take a look at what that looks like here. So I have 2 e 2, email use cases here. I'm gonna be showing you the marketing email use case later, but this is the sales use case. This is the prompt that I had used. So as you can see, as a sales b two b expert, it's exactly what I had just copy and pasted from the deck. And so here you can see the email 1, email 2, and then email 3. So this is a really great starting point. Our philosophy here at Goldcast is that you shouldn't publish anything that AI generates. It's just a great starting point for you for you, and this is a great place for me to continue editing and, refining, and so all from that one prompt here. Moving on to our 3rd use case, which is social media content. And so a lot of the times from your webinars and your events, like I said, they're jam packed with really really great golden nuggets of great information and snippets of thought leadership, and it's perfect for, you know, crafting social media content and really filling up your social media calendar. And so here, this prompt is, you know, I wanna generate a concise and engaging LinkedIn post that directly relates to the content of the AI generated clip from your webinar. And so the prompt here says, you are a b to b marketing social media content creator based on this clip, generate create a LinkedIn post that includes a compelling hook, a summary of the key point of the clip, a call to action to watch the full webinar recording, and some information about the conversational tone to match with our brand tone and voice. And so, if I go into ContentLab here, I'm gonna head over to this, let me head over to the clip section. And so from here, I can you know, from from here, I can either create a clip just directly from the transcript if I know the content really well. But if I did not attend the session, I didn't, you know, put together the content, I may not know it very well. And so what we have here is a a way to have AI generate your clips for you. And so from research, we know that a lot of marketers actually have to rewatch webinars at least 4 to 5 times in order to find those key moments, those key time stamps that you want to trim up. And so with AI, we really speed this up for you. So from just clicking this one button, AI is combing through that transcript, identifying those key moments for you, so then all I have to do is click on this one, you know, clip that's saying using unique stories to stand out on LinkedIn. And if I like this content, I can just say, okay. Let's generate a social post. And so from here, our AI is then just using our system prompts to write out, you know, the LinkedIn posts for you, again, as a draft. Alright. So our 3rd use our 4th use case here, I showed the sales marketing, sales email use case. This particular use case is for email marketing. So what this is going to help create is a series of strategic emails that help highlight the main points of your webinar to really encourage the people, that receive this email to watch the recording of the session. Okay. So as a b2b email marketing expert, draft a 3 part email campaign to promote our webinar. Insert your webinar title on demand. Each email, should be approximately 200 words and include a catchy subject line, an opening paragraph that highlights a key problem or challenge addressed in the webinar, 2 to 3 bullet points summarizing the main takeaways, and a clear call to action to watch the recorded session. Again, I'm adding some tonal information to align with our brand tone, so it says both professional and engaging, to align with our brand voice. So I'm gonna copy this prompt right here, and in the email use case, what I'm gonna do here is then just select a tone of voice, say you know, select the text length, and then copy and paste that prompt directly in here and click generate. When I do click generate, this is what's going to come out. And so here you can see that this is the prompt that I've used, and this is the the results. You can see those 3 individual emails, that are marketing specific and not sales specific. Okay. Our 5th use case, is SEO blog outline and actual content. So, many people on the marketing many b to b marketers are using webinars and repurposing webinars to generate more content like blog posts, and the great thing about this use case is that there's a lot of really, really great information in your webinars packed into 45 minutes. And so from there, you you every single time you create a blog post or a blog, you know, calendar, you don't have to reinvent the wheel. You can look at what videos you already have or webinars you already have and use different AI prompts to repurpose those videos into more content. And so from here, there's really 2 different use cases. You can use it to outline blogs and actually write blogs. And so what you can expect from this type of prompt is an outline that is specifically optimized for specified keywords asking for the AI to strategically weave in those keywords throughout the content so you can improve your search engine visibility. So my prompt here is, as a b two b tech content writer, create an outline for a keyword driven blog post based on our recent webinar with the webinar title. The blog post should be optimized for keywords, and then I'm entering the individual keywords 123, and encourage your readers to watch the full webinar. Please include a compelling title, including at least one keyword, an introduction, 3 to 4 main sections with sub headers, a conclusion with a clear call to action, and then a suggested metadata description. So you can see in this prompts that I'm being very, very specific. I'm assigning a role to the individual or to the to the AI, but I'm also specifying keywords that I want, the the prompt to ingest and also being very specific about what I'm asking as an output, so compelling title, including at least one of the SEO keywords, giving word counts to different sections. So by being very, very specific, I can almost guarantee the outcome to be as close as I want, going going in. So let's take a look at what this looks like. And so I have two examples here, one that's just creating an outline and one that's actually creating the SEO content. And so this prompt, is using the SEO outline. This is I'm asking for, ContentLab to help create an outline based off of the, webinar recording. And so from here, you can see that LinkedIn is the keyword that I had wanted, you know, ContentLab to really, really optimize, and it's right here on the header. I have an intro. I have, you know, these outline sections of section 2, section 3. I have videos that are embedded throughout that are related to these out outline sections. I have more sections here. I have the the conclusion here. I've got the call to action, and I've also have the meta description. If I wanted, this to be a actual post and not the outline, I can do that here as well. And so I've just slightly changed the prompt to remove any mention of outline, but to actually generate the post. And so here you can see that it's actually writing writing the the blog content, it's giving me the introduction, it's actually giving me the content for each outlined section, including those, videos to really, really bring this asset into a multimodal aspect to be really, really optimized for search and having a really solid conclusion in that meta description that I'd asked for. And the last use case here that I'm going to leave you with is, how you can use, AI prompts to generate long form ebooks and outlines. So who here is a content marketer that, you know, is responsible for long form assets? If you are, you know that these assets typically take a very long time to create because not only do you have to have the topics, you have to have a lot of the details, the research, you have to have the writers, the design, the videos. There's a lot that goes into a lot of these larger larger form assets. And so here, what you can expect is this prompt will generate an outline for an ebook. And so it will incorporate, the content structure, key quotes, suggest different visual elements, and it'll also provide a framework for creating in-depth valuable resources, that can expand your webinar content. And so this prompt here is saying, as a b two b content strategist, create a detailed outline for an ebook based on our webinar and then insert your webinar title. The ebook should expand on the webinar content, providing in-depth insights and practical advice. Include a compelling title and subtitle, an intro with 200 words, 5 to 7 main chapters with brief descriptions for each, ideas for 3 to 4 visual elements, conclusions, and next steps, and then also suggest 3 to 5 poll quotes that I can include in the webinar. And for the tone and voice, aim for a professional yet accessible, tone. So if I copy, this, this prompt and I head over to my ContentLab and I go to my custom section, you can see that this is the exact same prompt that I copy and pasted from before. So as a b two b content strategist, you can see that this is a great outline, a great starting point point for me to create this ebook, mastering video content on LinkedIn. So I have my title, my subtitle, my introduction, the different chapters, about 7 different chapters, ideas for visual elements, a conclusion, and then pull quotes for me to include. And so these are all actionable prompts that you can take today. If you, do not use ContentLab, this is free for you to sign up and use today, free for everybody. You can repurpose 1 hour of video every month on the free plan. But if you, would like to use other different models, you can use it on chatgpt, quad. All you would have to do is copy and paste these prompts and include your, video transcript along with your prompt so you have that context to go along with it. Awesome. Any let's see if there are any questions. Ashish, anything you wanna add here? No. Nothing. Nothing in particular. Cool. Alright. Let's go into q and a now. I think we've got a bunch of questions. Let's see. Give me one second as I get reoriented. Alright. If you have questions for us, please do throw them into the q and a tab. If you have questions that you see that you're interested in, click on the thumbs up so you can upvote them. We'll start with the most upvoted questions first. Alright. So, just one second. From Avani, how do you train the Goldcast model to understand your brand voice? Ashish, do you wanna take this one? Yeah. Totally. So I think there's a couple of ways. One is, you know, you can there's high level kinds of, I guess, like, buttons that you can press that are just like, hey. Our brand voice is, like, funny or professional. Like, that's already, like, something that you can do very easily. But, also, we're building something right now to sort of basically take well, very similar to the the demo that I briefly showed, which is, give me, like, a a blog post or a set of blog posts, and we sort of, like, derive a brand tone from it, a brand tone and sort of like a style guide from it. So that should be available to use very, very shortly. Yeah. So, what Ashish is sharing is a product update we'll be making very, very shortly where, you'd be able to train your content lab with your company's brand voice by either, uploading brand guidelines or sharing content to train your, AI on top of. So if it's a couple of blog posts or social media posts, it's just copy and pasting that text and allowing ContentLab to train, different aspects and, different descriptions of, of your brand voice. So more to come there. It's super exciting. Okay. I've got a question from Eric. I've asked JAT TV to go back and rewrite a blog post in active voice, but it seems to not understand what I'm asking. How do I get the correct output here? Ashish, any any suggestions here for Eric? I think a lot of the time, it's kind of like maybe giving it, like, a sentence or 2, in an active voice as opposed to a passive voice and just saying that, like, hey. Like, this is an example of of active voice, because, yeah, it might just not not understand. But usually giving examples is is a very good, a good idea in in the in the prompt itself. Awesome. From Cassidy, have you found some large language models are better for creativity than others, or analyzing data? Good question. Yeah. So, I mean, I think for things like analyzing data, what we have found is, like, a lot of, like, chat gbt or whatever is is generally really good mostly because of the plug ins that it has. Basically, kind of what I was saying, which is that, you know, these LLMs can't really, like, access the outside world when they're responding to your prompt. I think JWT is kinda, like, breaking that paradigm a little bit. So, for instance, it's allowed to, like now there's, like, a way for chat gpt to sort of, like, run Python code, for instance. So it can, like, generate and run Python code to sort of analyze your data for you, generate some, plots and things like that. So I'd say chat gpt generally is is really good for analyzing data. As I think others have mentioned in the, in the chat, it seems like Claude is probably the best at doing creative writing, like, for writing marketing copy, as of right now. Gemini also seems quite good, but these are more subjective. I think just that that's how we feel, internally, and we sort of use cloud for for a lot of those use cases. But I I think, a lot of it also has to do with how you can prompt it. So things like the temperature that I was talking about earlier and and prompting it properly, a lot of the top models can create pretty good creative output from them. Awesome. From Allison, I think this is actually a really good question. Where in trashypt do you enter a system prompt? Yeah. Sorry. So this is this is kind of going back to what I was saying also about the temperature, which is that you sort of would need access to, like, the opening eye or the anthropic sandbox. In those, it's like a separate, separate text input box for you where you can put the system prompt versus the real prompt. For chat gpt, it already has a prebuilt system prompt, which will just be like, you are a helpful assistant whose name is chat gpt or something like that. So to override that, you need to have access to the to the other API version. Cool. And then, just a question, in the chat, if you could don't mind just, helping us define this again. Can you reiterate what you mean by temperature? Yeah. So, yeah, so going, I guess, back to, like, what I was saying earlier, which is that the like, an LLM is sort of, like, creating, like, the responses, like, one word at a time. Right? So it has access to 2 things. This is, like, a very high level, which is, like, all whatever it was trained on, which is the model itself and then the context that you gave it. So it'll basically go back and choose, like, one, like, the first the best word to start the response, then it'll use that word in addition to the prompt, in addition to all of its whatever it's trained on to then choose the next word and then the next word and the next word based on that. Like, in that kind of, like, a, iterative fashion, it'll generate your entire output. So the temperature is kind of saying, like, how likely is it to choose, like, the absolute best word, in its in from its, like, training or whatever, for each one of those iterative steps. So you can imagine that if you're always choosing the absolute best words, that will always kind of result in the same output every single time. But if you're choosing kind of, like, you're you're maybe not choosing the best one, you're choosing the 5th best one, then you're choosing the 4th best one, then you're choosing the 3rd best one, you can actually get, like, very different sorts of outputs. So, basically, like, a high temperature would be, like, a more creative model, something that's more likely to create, like, if I were to run it 10 times, generate 10 radically different outcomes, whereas if I were to do low temperature, it wouldn't sort of always, output the same thing. Cool. Question from Jennifer. Kind of like a follow-up. If you can't enter a system prompt, do you just lump it into your own prompts? Any flags here on length? Yes. Yeah. You should you should definitely lump it into your own prompt. You should probably put it at the beginning of the prompt, and and kind of, like, try to draw a little bit more attention to it in how you write it. Like, be like, make sure to respond as a blank, if you're trying to give it a persona. Any flags there on length? No. I don't I don't think so. I think just make sure that it's within the the context length of of the LLM that you're using. Yeah. So different LLMs will have different, like, lengths of how long prompts can be. Right? So we just have to be aware of those limitations. Let's see. Question from Julia. Oh, I accidentally marked it answered. I will I will get back to it, Julia. One second. So question from Amanda. Any recommendations for pre event AI use cases? Very good question. This is definitely something that I've been thinking about a lot. A lot of these use cases that we shared is based off of a a video recording that already exists. But we there absolutely are lots of pre event AI use cases. The ones that I use mostly is more more for planning purposes. So before you have your event, you have to have your event title, your abstract, your speaker bios, session, you know, descriptions, all of that can be AI prompts. Again, but format, it's likely differently because you won't have a transcript to copy and paste to put contacts, so you'd have to provide additional context and that might be, messaging frameworks or, you know, product launch details. But the important piece is that you have additional context for the AI, and large language model to reference to be able to generate more, content for you. So in my mind, there's definitely a lot of pre event AI use cases. The big ones for me are, just titles, descriptions, even, you know, topics for additional events. So if this was a really, really great topic for us, AI prompts, I can use, you know, AI to help me think of more topics that are similar to that my audience might be interested on. So you might wanna be able to upload persona information that, AI can help brainstorm, different campaigns that might be impactful. Definitely let us know in the chat if there are any other ideas that you all have for pre event AI use cases because, there's I think there's just a lot to consider here. Okay. Alright. I found your your question, Julia. Can you develop a brand tone from creative samples rather than from blog posts? I'm curious to hear what you mean by creative samples. But, Ashish, do you wanna take this one just to hear what's possible? Yeah. I think I think so. I I mean, yeah. I'm not a 100% sure what creative samples means either, but I'm guessing it's also, like, some other, like, creative writing pieces or or something like that. If you think we would be able to do, like, ads. So, like like, add copy or, like, add images. Yeah. It's that's a really interesting one. So I think what we were kind of, like, talking about when we went brand tone was, like, more along the lines of, like like, a style guide type of thing, to be totally honest. And, like, for that, you need, like, lots of text content. The amount of, text content in an ad 10 or, like, a, like, a copy ad, right, like, an image, like, copy ad tends to be, like, a little bit less. But I could imagine, like, if you have a, like, a large corpus of, like, television advertisements or something like that, something, like, related to that, that that could actually have a more than enough, kind of, like you could that that kind of the transcripts from that kind of boil down to enough text content to actually, like, create, like, a useful brand tone. But I I think I I kinda see what you mean. It's like, if I'm trying to, like, create a brand tone, which kind of encompasses more, like, like, picture or or creative, like, visual content, how do I kind of do that? It's a really good question. I don't think it's like it's a little outside of the scope of, like, what we've kind of been working on at Goldcast, but I think it is it is something that is definitely possible, but not something that I know much about. Yeah. We're in the world right now where anything is pretty much possible. We just have to figure it out, so let us figure this one out. Alright. Next question from from Beverly. In ContentLab, are you uploading the actual webinar recording for the LLM to use as a resource and responding according to the prompt? Yes. Essentially, yes. So, like, basically, the webinar recording gets transcripted, and we're basically, like, running, something like prompts over over, like, the transcript. Yeah. Got it. Okay. And then a question from Megan. Instead of uploading a webinar, can you input a website? So for ContentLab, we are video based, and so you would have to upload a video for us to transcribe. And I think if I'm just thinking if I were to use, like, another LLM, like chat g p t or or Claude, your website's a lot of context. It might kind of overload information. And so I think just based off of best practices is being as specific as possible. And so unless it's a a couple of pages on your website that you wanna pinpoint, putting your entire website would be tough, to get, you know, outputs that you would want. Ashish, what do you think about this? Yeah. I think that's right. I think, I mean, one thing that we have been kind of, like, thinking about internally is, like, how do you kind of, like, link, like, a whole CMS, like, a whole company CMS to, to Goldcast to kind of, like, kind of take all of, like, the content from an entire large marketing website and kinda, like, putting it in. It is, like, a problem that we're kind of actively working on, actually. Yeah. I I I'm not sure if I can save much else beyond that, but it is definitely something that that I think is possible. With, like, chat GPT or something, I think, yeah, as I was mentioning earlier, there are, like, plugins, so it can kind of, like, read the text on your website and and respond to them, but probably just, like, the first page of your website. Like, for the ability for it to, like, crawl, get all of the content, and then, like, give a reasonable response is not really there right now. So there needs to be some way of, like, kind of more in a more structured way, like, taking your entire CMS, all the content in your CMS and, like, giving it to the, which is something that we're we're focusing on. Cool. Alright. We've got our last question for today, from Jason. I think we actually might have covered parts of this question elsewhere. Could you have AI ingest the content on a complete website to be able to teach it our brand voice? Yeah. I I think this is actually a great question based on what I just said, which is that's one of the reasons why we're kind of, like, saying how, like, how do you connect your CMS to, ContentLab? One of the outputs of that is exactly that. It's like you have all these blog posts in there. You also have all of this other copy that you've written for all kinds of other stuff. How do you take all of that and and generate, like, the brand voice from it is definitely something that we are working on right now. Like, the first thing that we're going to release, like, to your point, Kelly, from earlier is is really more, like, just, like, given one blog post or a couple of blog posts, how do you kind of generate blog, certain brand voice from that? But I think to Jason's point, it's, like, it's better if you take everything, and and that's something that we're we're working on. Yeah. So it sounds like a lot of, a lot of what we're working on is is of interest. And so I think in less than a month, we'd be able to share updates on on this, this brand voice, being able to set your brand voice. There was a question I saw. You're an agency. Will there be a way to have multiple brand voices? And the answer is yes. You would be able to have multiple brand profiles. So if you wanted one for your brand, one for, you know, a thought leader in your company, or multiple brands under your portfolio as an agency, that is definitely possible. Awesome. Well, thanks, everybody, for your all of your great questions. Thank you, Ashish, for joining me live today. I always appreciate a great, you know, presentation partner. But if you do have any questions, please don't hesitate to reach out to me or Ashish. You can find us on LinkedIn, but definitely try out these AI prompts and let us know if you're getting good results. Hope you had a great session today, and thank you. Bye, everybody.