Loading summary
Transcript101 lines
- [00:01]
Michael Stelzner
Hey, before we start today's show, if you want to accelerate your AI learning, I have a solution for you. Become a member of our AI Business Society. You'll join me as we go deep with live AI training each and every month. Imagine crafting more persuasive content, creating stunning images and automating those time consuming tasks. It's all possible when you join the AI Business Society. Go to socialmediaexaminer.com AI and join today.
- [00:41]
Chris Penn
Welcome to the AI Explored podcast, helping you put AI to work.
- [00:48]
Michael Stelzner
And now, here's your host, Michael Stelzner. Hello, hello, hello. Thank you so much for joining me for the AI Explored podcast brought to you by Social Media Examiner. I'm your host, Michael Stelzer, and this is the podcast for marketers, creators and business owners who want to know how to use AI. Today I'm going to be joined by Chris Penn and we're going to explore the solution to a problem that I think many of you are facing. If you have ever used AI and you've been really frustrated with the output that you're getting out of, it just seems completely out of left field. It's not what you're looking for. We're going to talk about a concept called priming. And by the time you're done listening to today's interview, you're going to understand how to increase the accuracy, confidence and results that you get from AI whenever you prompted to do anything. I think you're going to find it really, really interesting. We go down some rabbit trails, we talk a little bit about the technical side of how it works, but. But by the end, it's going to be very clear to you how to improve your AI output by giving it the proper set of instructions. Let's transition over to this week's interview with Chris Penn, helping you simplify your AI journey. Here is this week's expert guide. Today, I'm very excited to be joined by Christopher Penn. If you don't know who Chris is, he's a data scientist, author of the Intelligence Revolution and chief data scientist for Trust Insights. His company Cons, performs AI workshops and builds custom AI solutions. His course is Generative AI for Marketers. And Chris will also be presenting at Social media Marketing World 2025. Chris, welcome back to the show. How you doing today?
- [02:35]
Chris Penn
Thank you for having me. I'm doing great. Enjoying some of my homemade ginger ale and ready to talk some AI.
- [02:39]
Michael Stelzner
Well, I'm super excited that you're here. Today, Chris and I are going to explore a concept known as AI Priming which will probably be unfamiliar to many people here today. And we're going to talk about how, by following this methodology, you can increase the likelihood that you get the results that you're hoping to get out of generative AI tools. Before we get into what it is, why is it so important to prompt in general? Like, just explain why prompting, when it comes to using generative AI tools is, like, so critical.
- [03:07]
Chris Penn
Here's the way I'd love to talk about generative AI. It is the world's smartest, most forgetful intern, right? This intern has 255 PhDs. They got a PhD and everything, but they can't remember anything. That is an apt description of how these tools work. Right. They are stateless, which means they have no retaining memory whatsoever. And when you prompt them, you are giving them instructions. You would never, ever say to the intern, hey, intern, go do me a SWOT analysis of this brand. Right? You would never give that direction to an intern. You would say, here's what a SWOT analysis is. Here's the brand we're talking about. Here's all the information I have. Now, intern with all this information, go and do the SWOT analysis. And they'll say, got it. I know how to do that now. You've given me the information if you don't do that. These models are all triangulated on three core values. Helpful, harmless, truthful. The first value is harmless because the companies don't want to get sued. The second is helpful. They want. They try to obey instructions. The third is truthful, and truthful is a distant third. So if you give a bad instruction, it will try to be helpful, and it will do what's called hallucination, AKA making things up to try to be helpful. So the more information you give to the AI intern, the better it's going to perform. The more truthful it will be, the more factually grounded it will be.
- [04:26]
Michael Stelzner
Yeah, I love this. Ethan Malik used a very similar analogy, and I can't remember exactly what he said. But, you know, the idea that it forgets everything every time you talk to it is kind of important. Now, there are some ways that you can increase the likelihood it doesn't forget by using, like, custom GPT or Claude projects and this and the other thing. But yeah, so, like, it's only as good as the direction that you give it. It's kind of like, you know, Chris, you and I have known each other for years, but I would have to act as if I'd met you for the first time and I'D have to explain who this audience is that you're talking to before you went and record this podcast today in order to be able to prompt you to kind of output the kind of stuff that we're talking about. And that's kind of the analogy, right?
- [05:05]
Chris Penn
Exactly right. And maybe I have some memory issues. Maybe you told me that last week and then you have the call this week. You're like, I don't remember who talking to. And you're like, remember, this is a I explored podcast, is for business folks and entrepreneurs and small business own owners and marketers. Like, okay, I remember now.
- [05:21]
Michael Stelzner
Yeah. And it's kind of fascinating to me. I have found that if you get good at prompting AI, you can also get good at prompting humans. Have you found this to be true?
- [05:29]
Chris Penn
Yes. In fact, I've gotten better at working with human beings because I just treat them like another form of AI.
- [05:34]
Michael Stelzner
But it's true, right? Because it makes sense logically when you process it in your brain, when you give people better set of instructions and context, you're going to get a better response from a human. Therefore, why would you not also do that potentially every day in your human activities? And I'm finding I'm becoming a better communicator as a result of doing this stuff with AI. But AI, unlike a human, at least today, has a memory issue. Right. So the way that you prompted is absolutely important. So let's now talk about priming. Let's describe what the heck it is and what its benefits are.
- [06:06]
Chris Penn
Sure. So priming is derived from a technical term. Priming essentially is giving a model a lot of data or having the model generate its own data so that you can prompt it more successfully if you were to compare it to cooking. Priming is the mise en place. Right. You chop the vegetables in advance, you prepare all the ingredients, you get everything measured and laid out, and then you're ready to cook. If you don't do that, if you just wing it. Well, I mean, we've all kind of winged it in the kitchen. If you've. You've been alive long enough and you know that that doesn't always turn out so well.
- [06:37]
Michael Stelzner
Yeah. So what's the upside then? I mean, like, I know when we were prepping, we talked about three things, accuracy, confidence, and focus. Talk to us a little bit about that.
- [06:44]
Chris Penn
So priming, if you are doing any kind of queries that were the importance of the result being factually true, is relevant to you. You absolutely want to do some priming because you will get better Information. Here's the thing about generative AI models. They are trained on a certain amount of information, a lot of it. 15 trillion tokens and upward, which is like a bookshelf of books that goes around the equator twice. So it's a lot of information. So they know a lot about a lot, but they don't know a what's quality and what isn't. Right? So some drunk dudes posts on Reddit are given the same level of importance as a clinical paper from the Mayo Clinic. Right? These are, from a word perspective, the same thing. The model doesn't know what is and is not quality. The only way for a model to know quality is either to prompt it or don't include bad data. And to begin with, and of course, every model is vacuuming up as much text as possible, so they've thrown quality out the window. So prompting is how you get to that quality. So priming is about providing the data in advance. Because if you do that, almost all AI tools will say, I'm going to use this data source as the preferred source. And some tools say I will use this data source as the only source. So Notebook LM the free tool from Google. You have to provide the data. If you provide stuff and you ask a question that isn't in the data set, it will say, you didn't give me that answer in the data, so I can't answer that question. So accuracy is really important. So is focus. Right. Again, if I load all my social media marketing world transcripts from the sessions that I've given over the years into a model now as part of the priming process, it knows what the topic is implicitly based on all the words that I've said over the last 10 years. That focus and that accuracy creates confidence. You can be more confident in the output of a model because you know it's pulling from good sources. I'm in the midst of the fourth edition of my book that's called the Intelligence Revolution. It used to be called AI for Marketers. This version of the book is made from a year of YouTube transcripts, a year of my newsletters, a year of my LinkedIn posts. And I have told these things, hey, here's all the data plagiarize from me. Copy paste my words. You know, we wrote the outline in Gemini and I'm just having it yank stuff I've already written and assemble it. I have full confidence that in what the model is spitting out because it's my stuff.
- [09:08]
Michael Stelzner
Got it. Okay. So really the big advantage here to priming is that you are going to first of all get more accurate results. And I don't know anybody who's using AI that doesn't want to improve their accuracy. You're going to increase the likelihood that it's not hallucinating. Right. That it's accurate. And you're going to allow it to focus its, for lack of better words, intelligence right into a specific area. Those are like the advantages to priming. So let's talk about some applications of where we can actually put this to work. Just some examples so people can wrap their head around this.
- [09:43]
Chris Penn
Sure, there's a bunch of different examples. So to start, let's say you have access to your social favorite social media marketing tool. Maybe it's hootsuite or agorapulse or Sprout Social, whoever, it doesn't matter. And you can extract out your data and you can extract out some competitor data because a lot of tools will support competitive analysis. You could say, hey, I want to take all this data, lump it together. And you would do this in data pre processing, not with generative AI, but you would sort it maybe by engagement. Say, this is the top 20 social media posts performing in our category. This is the bottom 20. You hand that to a generative AI tool and say, explain to me the difference between the top and bottom performing posts. It would give you a nice analysis. That analysis then becomes part of a prompt. You would say, hey, these are the best practices for Instagram posts based on our analysis of real data. As we write Instagram posts today, you're going to use this set of best practices we derive from real data as, as your, your guidelines. So that priming, you bringing in that knowledge that you derive from a credible source and putting it right in. Maybe you are trying to do a SWOT analysis right of you versus a competitor. Strengths, weaknesses, opportunities and threats. Good old business school stuff. You would say, hey, here is a corpus of text. Or maybe it's the actual web pages of my competitor. Here's my website. The priming process there is providing both sets of data and saying, now let's do a SWOT analysis of this to see what are they saying? That is good messaging. What are we saying? That's good messaging. How do we improve, improve our good our messaging to mitigate our weaknesses and take advantage of the threats they've exposed to in their poor messaging. So anytime you have data or you want, you generate data with a model you are using priming and this applies across every discipline. I'll give you a real life Example of my personal life. Last August, I went to the er. Turns out my gallbladder had died. On the way there, I opened up an AI app and I said, you're going to act as a world class physician specializing in children medicine. We're not going to do a diagnosis. I want you to ask me questions about what's going on. And during my drive, you know, I was at that point in a lot of pain, so I was repeating myself. I got to the ER and I said, summarize everything that I've said as an intake form for an ER physician. It consolidated all my medical history, everything. And I got to the er, just handed the doctor my phone and says, I can't talk anymore. Here, read this. The doctor's like reading the cell going, huh, this is really good. Are you a doctor? Like, no, I'm an AI expert. But that priming process was a asking the model, what do you know about this? What do you know about medical intake? And then me giving it all the information to consolidate and summarize.
- [12:33]
Michael Stelzner
When we were prepping, you said that for high risk industries like finance, law and tax and all that kind of stuff, there's some applications here as well. Maybe you could talk about that as well.
- [12:42]
Chris Penn
For entrepreneurs and business folks, there are things that AI is. It requires a lot of work to get right. So for example, creative writing, you need to do a lot of work. There are some things where you do not want creativity, like law. It turns out that AI tools are very good at extracting high probability data, like legal templates. So real world example. My friend Julia is starting up her own video editing company. I said, hey, do you have a scope of work? Do you have an msa? Do you have, you know, all these different things to run a business properly? And she's like, not really. And I can't afford a lawyer. I said, let's go into Generative AI. Let's download. She lives in a US state. Let's download your state small business laws and let's generate, based on those laws and its known best practices, what a scope work template looks like for this kind of agency, what a MSA looks like for this kind of agency.
- [13:31]
Michael Stelzner
What's an msa?
- [13:32]
Chris Penn
MSA is Master Services Agreement.
- [13:34]
Michael Stelzner
Okay, cool, keep going.
- [13:35]
Chris Penn
Agencies use those all the time. What is a billing statement look like? What is a creative brief look like? And within 48 minutes, I built her this entire set of legal documents. And now I did say, at some point, once you earn some money, you must get these checked out by a human Lawyer. I'm pretty confident in them because this is stock stuff. But you still should always get AI output checked. Here's the rule of thumb. If you do it with a template, AI can do it probably better. So if you're a business person, you've got a template for an expense report, you've got a template for this, that the other thing, you got a template for a social media post, you got a template for a creative brief that is the most fertile ground for AI to be used in your business and to save you time, save you money and make you money.
- [14:19]
Michael Stelzner
Awesome. Okay, so let's start with this concept of priming. Like let's say, okay, we'll acknowledge that most of us probably aren't doing priming. And you've defined priming as providing a set of data to the AI to increase the likelihood that you get what you want out of it. Right. So where do we begin when it comes to priming?
- [14:39]
Chris Penn
We call it the repel process, the trust insights. It's a role action, prime prompt, evaluate, learn. And so here's the process and I'm going to give you the most simple form where you don't have to go download a bunch of data, where you have the model generate its own data. You would say first you give it a role. Who is you want the model to be using jargon, using industry terms, because again, that helps focus the model very quickly on the domain that you're working in. You are an award winning social media marketing speaker who's, who knows Pinterest in and out like Elisa Meredith. Right. That'd be a good example of a role action. Today we're going to devise a Pinterest strategy. Right now we move on to prime. We're going to ask three to five questions, three questions if you're in a hurry, five questions if you got some time. Number one, what do you know about best practices for this topic? Right. And it could be, you know, whatever it is. Two, what are common mistakes made by less experienced folks in this topic? And then three, what expert tips and tricks do you know about this topic that we have not talked about yet? What this is doing behind the scenes? It is invoking several different AI strategies. One is called contrastive prompting where you say essentially what's the wrong way to do this? And that tends to invoke more specific knowledge. The second is a structured chain of thought where you're having it dig deeper and deeper into your saying, hey, what you've already talked about, don't talk about it again. And the third is, of course, generated knowledge itself. It's every time you say one of these questions, it's going to give you five to 800 words of response. By the time you're done, you may have 1500-2500 words of context on screen that you've effectively primed the model as though you had uploaded the data yourself. So that's the first three steps of the Repel framework, and that gets you ready to prompt.
- [16:31]
Michael Stelzner
Okay, perfect. So I want to back up a little bit and dig a little bit deeper on priming. The three questions we just talked about are what do you know about the best practices for this topic? What are common mistakes made by less experienced people? And then what expert tips and tricks do you have that we've not yet talked about? Now, that's going to give you data which you can analyze if this is a topic that you have knowledge in and say whether this is good or bad. But you also hinted about bringing data to the table a little bit here. We haven't really explored that side of it. I would love to talk about other than you said, if you have a template. Right. But I want to dig a little bit deeper. What kind of data can we bring to the table? Let's explore that a little bit.
- [17:13]
Chris Penn
You can bring anything, anything at all these days. So PDFs, spreadsheets, documents, web pages, video, audio, you name it. Today's multimodal models are capable of analyzing just about anything you give them. Now, I will give you this one caution. They're bad at anything that's not language. So don't have them do math. Right? Don't have them try to do, you know, Bollinger bands or standard deviation analysis. They will not do that. Well, they can write the code to do that, and some of the tools do that as a workaround. That's not what they're really good at. But any form of data that you can download from somewhere else, you can load into any of these tools and have that be part of the priming process. So, for example, maybe you're about to do a YouTube strategy and you found this great paper on archive.org and you say, you know what? I'm going to download this PDF and I'm going to load this in as research and then say to the model, I want you to use the conclusions that you can draw from this paper to help me craft my YouTube strategy about how people retain information. Anything that you've got, maybe there's an article on the Social Media examiner blog about how to do how to, you know a post by John Loomer about how to write great Facebook ads, you grab that download and say, we're going to use this as our guidelines for. And here's the Facebook ads I want to write. So any data you have, the advantage of bringing your own data is hopefully, you know it's good, right? You know that this is a credible source. John Loomer is a credible source on all things Facebook advertising. Brook Selles is a credible source on all things customer care, as opposed to that drunk person on Reddit who you have no idea if they know what they're talking about or not.
- [18:49]
Michael Stelzner
So talk to me a little bit about internal documents versus external documents. I know that a lot of people listening might feel a little squeamish about providing the AI something that might be internal, confidential, and they also might feel a little weird about also taking information from someone else and putting it into the AI model. So let's talk about both those a little bit. Let's start with the internal side.
- [19:10]
Chris Penn
So internal, the level of concern is based on the risk, right? So our general rule of thumb is always this. If a tool is free, you're paying with your data, right? Your data is going to be used to train a model. So any tool that's free, probably you're paying with your data and you can check the terms and conditions, you're probably saying, yes, you can use my data to train on some tools like Anthropic Claude. If you're using the paid version in any version it is secured, which means that they do not use your data to train on. And unless you trigger like the abuse warning system, no human will review it in Google. Google Workspace. The Google Workspace version of Gemini is protected anything that you're doing in that.
- [19:47]
Michael Stelzner
Version, which is a paid version, right?
- [19:49]
Chris Penn
It's not only a paid version, it's a paid version within Google Workspace because the individual paid version still does train on your data.
- [19:57]
Michael Stelzner
Ah, okay. What about ChatGPT?
- [20:00]
Chris Penn
So ChatGPT by default does train on your data for individual and plus for pro teams and enterprise. It does not. However, in ChatGPT you can go into the settings and turn off the training data in any version of it.
- [20:13]
Michael Stelzner
Got it. Okay, so what I'm hearing you say is, and this goes without saying, but I'm going to say it anyways, remove any customer personally identifiable information from your data sets. Right? We don't want to be putting the email addresses, phone numbers, IP addresses into the even, no matter what. I mean, it's probably going to violate your privacy policy anyways. What about public data that's out there on the web?
- [20:35]
Chris Penn
Public data that's out there on the web is going to depend on how you're going to use it and whether it would fall under fair use. So for example, if I, if I copied a post off a social media examiner and I wanted to use the ideas that John Loomer had shared in the same way that I could do the same thing as a human, I'm not making a derivative work of that piece of content. If I said, here's this post by John Loomer, rewrite it in my tone of voice, that's a derivative work. You're violating someone's IP rights there.
- [21:02]
Michael Stelzner
So let's talk about tips on finding research out there on the web because I know that you use Perplexity and I think you may have also used Deep research by Google. Talk to us a little bit about your thoughts on those tools.
- [21:14]
Chris Penn
So when it comes to finding credible data sources, number one, you should have an idea of what is credible. So here's some things that I personally use. Number one, I like to look at academic journals, I like to look at peer reviewed sources, I like to look at any kind of article where I know the organization itself is credible. So for example, if I'm looking for medical stuff, I'm going to look at the Journal of the American Medical Association, I'm going to look at nih, I'm going to look at Stat News. Those are very credible resources. I'm probably not going to look at, you know, Madame Peony's Healing Crystals Incorporated, because from my point of view, that's not what I consider credible. I always look for document object indicators, these called DOI numbers. And so any paper that's been peer reviewed has a DOI number and those DOI numbers indicate that this is a registered publication of some kind of at least a minimum level of standard. So in a tool like Perplexity, you might say, today, you know, you're going to be a world class internist to help me to understand the effects of having your gallbladder removed. Today you're going to find peer reviewed articles and research and papers about the effects of gallbladder removal. Restrict your sources, peer reviewed academic journals only. And restrict your sources to anything that has a DOI number. Disregard anything that does not have a DOI number. Disregard all mainstream media. Disregard all social media. And then what you get out of Perplexity will be a sources list along with its summary. Ignore its summary, its summaries are not Helpful. What you really want is that side rail that has all the sources that say, here's where I got this information. Google's Deep research in Gemini 1.5 for paid individual users of Gemini does the exact same thing. You would use that exact, exact same prompt and you would say, hey, I only want things with DOI numbers from these journals. Maybe you even know specific journals. And it will present its list of sources. Then you go and download the PDFs or the files of the stuff you care about and then load it into. Depending on the level of risk. You could load it into ChatGPT if you just want to chat about it. Generally, if you want to make sure you're not hallucinating, you load it into NotebookLM.
- [23:18]
Michael Stelzner
Yeah, and just so folks know, Google Deep Research, as Chris said, is only available as of this recording for personal Gmail accounts, not for workspace accounts. And it is a paid service. And perplexity. Is everything you just talked about available with their free service?
- [23:34]
Chris Penn
It's free and paid. And so free users get like three pro searches a day. And just be aware, you know, check the terms of service, because the free service, your, your props are getting stored there too.
- [23:47]
Michael Stelzner
Here's the thing about AI. It's not just hype. It's transforming everything we do as marketers. If you're ready to master AI for your business, I've got something special for you. At Social media Marketing World 2025, our dedicated AI track will give you the exact playbook you need to thrive as an AI enhanced marketer. Learn from experts like Matt Wolf, Chris Penn and Nicole Leffer and many more. Don't get left behind. Join us in San Diego this March. Visit social media marketing world.info and get your tickets now while you still can. Save big. Got it. Okay, so just to summarize some of what we're talking about. In an ideal world, you are bringing data to your inquiry, your prompt, for lack of better words. And depending on how complex that data is, it could be something as simple as some internal documents that you've created about whatever it is that you do for your job. And it could be something as complex as literally research papers. Right, and is there a limit to how much of this data we should provide when we're priming?
- [25:01]
Chris Penn
The limit is whatever tools working memory is. So for example, ChatGPT has a working memory of 128,000 tokens, which take tokens, take 75% of whatever the token window is and that's the rough number of words. So ChatGPT can remember about 90,000 words before it runs out of memory in any given single. Single conversation. So if you've got a year's worth of sales calls, you know that's 800,000 words. You're not using chat GPT on it. Anthropic's Claude can hold about 200,000 tokens or about 150,000 words. Google Gemini Pro can hold 2 million tokens, which is two of these.
- [25:38]
Michael Stelzner
Yeah, say that for the audio audience so that they can see what that is. William Shakespeare.
- [25:41]
Chris Penn
This is the complete works of William Shakespeare, all 38 plays, 800,000 words. Gemini can hold two of these in memory before it runs out of memory. Generally speaking, One of the reasons why I use A tool like NotebookLM for example, is NotebookLM can hold. I wanna say it's about the same amount, but it's with up to 50 different sources. So you can just like drop PDFs in there left and right. I did that recently. I was working on a court case and we had 3,000 pages of court documents and just dumped it all into NotebookLM and said, okay, let's summarize specific parts of this so that we can get at the information we actually care about. Cool.
- [26:18]
Michael Stelzner
We're going to get back to NotebookLM in just a little bit. Okay. So where we're at right now is we've got all this great amount of data. With all this data that we're providing. I'm going to delineate this in two different ways, because on the one hand, you started with ask it three questions. What do you know about the best practices? What are common mistakes and what are expert tips? Do we need to do that if we're already providing the data set?
- [26:40]
Chris Penn
No, if you're providing the data set, you can just use the data that's in this. You'd say, you are this. We're going to do this. Here's the data. Summarize it at a high level. You always do that with data upload to make sure that it's reading it, that it's actually reading the correct information.
- [26:54]
Michael Stelzner
Okay. And then it's going to output something that you're going to use over and over again in your prompting. Is that the idea? If you're talking about this topic, you.
- [27:01]
Chris Penn
Can do that, or in this sequence, the prompt is the next step. So you've done the role, you've done the action, you've primed it by either asking it those questions or loading the data yourself. And now you're ready to prompt it, to ask it to do the task.
- [27:13]
Michael Stelzner
Okay, now I want to talk a little bit more about NotebookLM. So once we've got all this data and whether the data was found on a search using Perplexity or some other deep research tool, or whether it was provided by you in an upload or it was prompted out of the model through a series of questions, you had recommended having AI digest the data. And I know you've talked about NotebookLM a few times, so let's talk about this digesting the data side of things a little bit so people can wrap their heads around that.
- [27:48]
Chris Penn
So one of the things that if you're using a tool like chat GPT again, it only has a memory of about 90,000 words, that is a budget you go through really fast. One of the hacks to get around that is to take, you know, if you have 3,000 pages of court documents, a tool like Notebook LM can summarize it. So you would say, I'm doing a specific line of inquiry I want to know about. These are the earnings call transcripts for the last year from my competitor. From their earnings calls, you put all that Notebook alum and say summarize for me the key points from all these transcripts that deal with social media marketing. And what it will produce is short 5 to 800 word extracts. Those you can then copy and paste and put into a Chat GPT and instead of trying to cram all however many thousands of words from the earnings call transcripts, you're now taking the summaries. More practical application for a sales and marketing person would be maybe you've got a tool like Gong, or maybe you've got your own call system and you've got all the sales calls that your team has made over the last year. You've got thousands of these things, you consolidate them into one big file, load that into Notebook LM and say identify the top 10 themes in these calls for, you know, ideally you have the calls listed like these are calls that resulted in a deal, these are calls that did not. You could have extracts, say, summarize the top 20 themes and the calls for deals we want. Summarize the top 20 themes for the calls we did not win. Then you take those extracts which are going to be 2000 words each, that goes into a chat GPT or Claude or Gemini, and it takes only the best stuff out of all that stuff because you know, in a, you know, even in a single 30 minute sales call, that's gonna be a lot of fluff. Oh, how was the weather? How's this, right? Yeah, you don't need that. You want the distilled down digested best content. And a tool like NotebookLM will get that.
- [29:36]
Michael Stelzner
Yeah. And we should state that NotebookLM has a free and paid version. Recently they came out with a paid version within the last couple of weeks. As of this recording, the free version should be able to do most of what we want to do. Is that correct?
- [29:48]
Chris Penn
That's correct. The paid version, there are several versions. So there's the paid individual version. And then Google Workspace users through Google Cloud can integrate it into Google Cloud. So if your company is a Google Shop Notebook LLM, you can have a secured instance just for your company.
- [30:02]
Michael Stelzner
So the prompt that we're going to be asking it for is to summarize key points around topic X. Is that really what I'm hearing you say?
- [30:09]
Chris Penn
That's correct.
- [30:10]
Michael Stelzner
And that data will be crafted generally in a very easy to understand language set. Because if I'm not mistaken, this was designed for academic use in the beginning, right?
- [30:19]
Chris Penn
That's correct.
- [30:19]
Michael Stelzner
So it's designed to kind of take the complex and make it more easy to understand. Right. And that data should be sufficient enough to prime most AI models without providing excess amount of data, right?
- [30:32]
Chris Penn
That's correct.
- [30:32]
Michael Stelzner
Okay, is there any downside to providing all of the data versus the summary of the data or any upside to providing a larger data set in the priming?
- [30:41]
Chris Penn
It depends on the data set. There's a lot of data sets where there's a lot of garbage in them. There's a lot of not helpful stuff. So for example, if you were to go on Reddit and using the Reddit API, go into the social media marketing subreddit and extract the last 90 days worth of posts, there's some good posts in there, there's some good comments in there. There's also a lot of not so good, low quality content. That's a case where, remember that the way an AI model works is it does associations with all the data you load. If you load a data set that's especially dirty, that's filled with, you know, ridiculous memes and things like that don't answer the question. Well, that will throw off the accuracy of the model because it's trying to do predictions with all these stories, swear words and all this, you know, these, these vulgar terms and you're like, that has nothing to do with my Pinterest strategy. When you use tool like NotebookLM, you, you digest that down to say, here's the top 20 themes in this huge data set about Pinterest then you can take that very clean, extract that distillate, put it into your regular AI model, and it is super focused now on just that Pinterest discussion and not on, you know, Bob's your uncle or whatever the things people are saying.
- [31:53]
Michael Stelzner
Okay? So we've been spending a lot of time talking about how to get really good, high quality, clean data so that when we actually start prompting the AI model, we increase the likelihood that it actually works the way we intend it to work. And earlier in the discussion, you introduced your repel model, and we're going to now review all the elements of it. But before we do, help me understand if this is a pie chart, how much of the data in a prompt falls under priming. Do you understand what I'm asking?
- [32:27]
Chris Penn
Yes. Probably like 95%.
- [32:30]
Michael Stelzner
Okay, that's important, right? Because that means that probably most people aren't doing this, right? Like, if 95% of your prompt is priming and most people think to themselves, there's no way 95% of my prompt is priming. Right. They're probably spending 95% of their time talking about the rules and the action and not focused on this. Right. And this probably explains why so many people struggle.
- [32:52]
Chris Penn
It does. I will give you just a very simple rule of thumb that I use for myself. If you're prompt, by the time you get to the prompt stage, if you're not at about 5,000 words, you don't have enough information, right? Whether you generate it while you upload it, whatever. If you're like, I've got the ultimate chat GPT prompt, it's a paragraph long. Like, no. No, it's not. You really want to let these tools talk again? Going back to the whole architecture of the way the system works, they need to talk. They need a lot of text to grab onto. The more you can provide, either by generating it internally or by uploading it, the better they will perform. It's data is the ingredient here, right? How much of an omelette are you going to make with half an egg? How much of an omelet are you going to make with, you know, five crates of eggs? That's the basic analogy.
- [33:38]
Michael Stelzner
Okay. Does it matter how we format the priming side of this? Can it just be a whole bunch of text or do you have recommendations on how to structure that data?
- [33:46]
Chris Penn
It doesn't matter a ton. If you do know the format, you can that. For example, if you're uploading a JSON file or a YAML or a SQL database, that certainly helps the Model understand what it's looking at. But most of the models, most of the time pretty clear they know what to do with a piece of information.
- [34:03]
Michael Stelzner
So it's just really a massive piece of text essentially is what this is in the prime section. Right, Right.
- [34:09]
Chris Penn
Because at the end of the day what the model does is it digests it down into tokens. No matter what format it is, it turns it into tokens, which is fragments of words.
- [34:17]
Michael Stelzner
Okay, so roll is the R part. How big of a description do we spend on the roll? Like we talk in the sentence or what are we talking here?
- [34:26]
Chris Penn
Yeah, sentence or two. I usually do the structure of you are or whatever and you specialize in whatever those enough keywords in that to at least start the steering.
- [34:35]
Michael Stelzner
Now, once you get to this is what you're going to do, do you recommend using a separate paragraph or delineator of some sort to delineate between the role and the action and, and the prime and stuff. And if so, talk to us a little bit about how that works.
- [34:49]
Chris Penn
So there's a particular markup language called Markdown that is a structured text only formatting language. But these tools can understand any structured format. So you could put in, you know, one role, two action and just have it be like a numbered list. Anything that communicates structure models do better with. Again, it's like an intern, you give it a wall, the internal wall of text or if you broke up that wall of text to make it readable. Because models have trained on us and all of our data, we do that intuitively, they understand that better.
- [35:20]
Michael Stelzner
Okay, so on the action side of things, how big of a descriptor are we providing there? You said the rules about a sentence. What about the action?
- [35:27]
Chris Penn
The action is typically maybe a paragraph and or four or five items in a numbered list. Because what you're doing is you're providing a high level overview of what the task is. You'll get to the details of the task in the prompt section, but you want to initially start out by there's a whole technical term for bidirectional encoding. But essentially what you're saying is early on in that action statement, you're telling the model, here's what you're going to start building these associations for. You're going to make a report, you're going to make an analysis, you're going to make whatever. And those words will activate certain parts of the model's knowledge to say, okay, now when I get all this priming data, I know how to start finding associations to fit this General action.
- [36:07]
Michael Stelzner
Okay, so in the action, obviously it could be lots of things. It could be ask me a series of questions to guide me down the path, or it could be analyze a data set that I'm going to upload. It could be almost anything you could possibly imagine.
- [36:18]
Chris Penn
Right, exactly. So, like, when I'm debugging software, I'll say, we're going to be doing some debugging today. First, you're going to summarize the application and explain the error and stuff like that. And this is the general process that I want you to use for this. Now, here's all my data, here's my code, and then here's the steps I want you to take for debugging my code. Tell me where I screwed up, and it will then do that.
- [36:36]
Michael Stelzner
And your code in this case is the prime content. Right? Is that correct?
- [36:39]
Chris Penn
Yes, exactly.
- [36:40]
Michael Stelzner
Okay, so we're now at the point of the prompt. So we've already talked about some of the stuff that goes in the prompt, the role, the action, the prime. What else goes into the prompt that we didn't talk about?
- [36:49]
Chris Penn
So the prompt is where you give it specific, detailed instructions, ideally what you want it to produce in detail, and then you give it instructions for how to think. And this is a critical part. This is a part that is almost nobody does. And they should. There's an old public speaking saw. Tell them what you're going to tell them. Tell them, then tell them what you told them. Right. It's not the best format for a public speaker, but it's better than rambling mindlessly on a stage for 45 minutes. That works really well with generative AI. So let's say, you know, we're going to do Pinterest strategy. We declared the role, we've declared the action. We've loaded some example data, and you're going to say, and you go into the. The prompt portion. Now you say you're going to now produce Pinterest strategy. First, recite back my instructions for me. Explain to me what it is we're doing. Second, identify the key points of a Pinterest strategy. Third, explain why you chose those points for Pinterest strategy. Fourth, give me my Pinterest strategy in full. And so you tell it what you're going to tell it. You tell it. You tell it what you told. And then you do the thing that basically lets the model almost do kind of like a trial run or rough draft of what you want to do is to let it think out loud again, filling up that memory with more information and then it produces the final output. So that process is part of prompting. There are many forms it will take. There's, you know, instructions for debugging. Tell me what functions, dependencies this code has on. You know, tell me what this error means. Tell me why you chose what you did, but you basically just want it to talk. What are you going to do? Why are you going to do it, how are you going to do it and then do it.
- [38:21]
Michael Stelzner
So is this stuff you just talked about part of action or is this part of some other.
- [38:24]
Chris Penn
This is part of prompt.
- [38:25]
Michael Stelzner
Okay, so literally, but when we're formulating a prompt, I understand that you call it prompt, but in the prompt you've still got the role and the action defined, right?
- [38:34]
Chris Penn
Yes. So that's already happened.
- [38:35]
Michael Stelzner
Okay, and then this instruction set, you just call it prompt.
- [38:38]
Chris Penn
Yes, that's the second P repel.
- [38:39]
Michael Stelzner
Okay, so it's useful to delineate between the action and the prompt. So the action it sounds like is just a high level thing it's going to do and the prompt is the very specific thing it's going to do. Is that correct?
- [38:49]
Chris Penn
That's correct.
- [38:50]
Michael Stelzner
Okay, so what if we don't know what kind of, I mean, can AI help us come up with a better prompt, for lack of better words in this section? Do you understand what I'm asking?
- [38:59]
Chris Penn
Yes. In fact, there's an entire subset of prompted called metacognition prompts that do exactly that. Where you say, I need your help making this prompt better, here's what I'm trying to do. Tell me what the intent of my prompt is. Tell me what the likely outcome is going to be, then tell me how you're going to make it better and why, and then you will iterate through that in this sort of metacognition style. If you have used chat GPT's 01 model, which is one of their newest models, that's essentially what it's doing behind the scenes. So it has this sort of structured chain of thought going on and you don't get a choice. Like that's how that model particularly and its infrastructure work. What we're talking about here is doing that in regular Chat GPT or Claude or Gemini or whatever tool you're using. Because in across the board, in every tool, getting it to think through things and maybe even reflect on what it's done always improves it.
- [39:49]
Michael Stelzner
One of the questions that came up inside our AI business society that you were kind enough to answer is tools for saving prompts. Because if these things are as Big as they should be. That's a lot of data and you don't want it just buried in some random thread instead of chatgpt. Right. So what tools do you recommend using to enable everyday people to kind of be able to reuse their prompts and maybe modify them?
- [40:12]
Chris Penn
Any note taking software. So you could use Apple Notes, you could use Evernote, Google, keep OneNote, Google Docs, right? Yeah, Google Docs, yeah, anything that, where, where you can organize. And what you want to do is you want to organize in a couple different ways. Number one, every model is different, every model behaves differently. So your prompts should be delineated by which model it's for. Is this for Gemini? Is this for Chat GPT? Is it for Claude? That's important. Second thing you want to specify is what kind of task is this or what kind of content is this? Is it an actual prompt? Is it a scoring rubric that you're using to evaluate things? Is it a knowledge block? So for example, with Trust Insights, I have like a two and a half pages of just about our company as, as a piece of text, I have another one of five pages worth of here's how we do our marketing. So when I'm asking questions, I can just bring in these, these pieces. I have another one which is a 19 page ideal customer profile distilled down from, you know, our CRM system, things like that. So I can drop that in when I'm prompting to say, like, hey, what would our customer think about this? So you want to have those kinds of pieces in your document management system and then you sometimes will want to specify, you know, what the actual task is. This is a generation task, a summarization task. Is it an extraction task? Is it cleaning up a blog post? Is it doing SWOT analysis, whatever the thing is, but that structure, you want to have defined and governed internally, whether it's just you or whether it's ideal across your team so that people know, oh, I need to do some podcast transcript cleanup and I'm doing it in Gemini. I'll go to the Gemini folder, I'll go to the podcast folder, I'll do the transcript cleaner, pull that out, use that, that and then, oh, I need YouTube captions or I need YouTube description and YouTube tags. Okay, there's a prompt for that. So go to Gemini, YouTube and stuff. So you just, you want to be orderly and organized.
- [42:04]
Michael Stelzner
Okay, thank you for that. Rappel is roll action prime prompt. We haven't talked about the E and L, so why don't you give US a quick little insights on that.
- [42:12]
Chris Penn
So E is evaluate. That is you look at the results and say, yes, you did what I wanted or no, you didn't. And then you provide instructions for how to clear me. And you go through that, that until you get the thing you want. Right, that's what that is. That's straightforward. The last one, L stands for learn. And this is again something I see almost nobody do. What you would instruct it at that point is to say, based on everything we've talked about today, I want you to write system instructions for an AI tool to perform the same exact task we did. So if you do, if you're doing Pinterest strategy, as we're talking about, you would say something like, now using everything we've talked about, you're going to write system instructions for an LLM to help formulate Pinterest strategy. The user is going to provide you whatever the data you provided in the conversation and you're going to step through the processes that we walk through today using your knowledge of prompt engineering to make a prompt for this. And then if you're in Chat GPT, you can build a custom GPT from that. If you're in Claude, you can build a project from that and include those system instructions. If you're in Gemini, you can build a GEM from that. What you're doing is in this last step is you are encoding the process that you went through. So the next time you do that process, it's much faster to get up and running because you don't have to go through as many steps.
- [43:23]
Michael Stelzner
Well, and I would imagine you should update your prompt library too, if you have such a thing. Right. Because presumably it will have evolved.
- [43:29]
Chris Penn
Right, Exactly. And depending on the system you're using, if there's been a model change, you need to update your prompts, you need to go back and update your prompts again, because every time a model changes under the hood, there's a bunch of stuff we don't know going on. So Google, about a month ago, as of the day of this recording, released Gemini 2. Gemini 2 behaves very differently than 1.5. So all of my old Gemini prompts, I noticed when I started using them, they broke, they behaved unexpectedly. So I said I went back and had to have Gemini regenerate them for the new model.
- [43:59]
Michael Stelzner
Absolutely fascinating. Chris Penn, thank you for answering all my questions and helping us wrap our heads around the importance of prompting and priming. In particular, if people want to discover more about you, what's your preferred social platform and if they want to work with your company, Trust Insights, where do you want to send them?
- [44:16]
Chris Penn
Trust Insights AI is the place to go. I I don't really do much with public social media anymore because it's such a crazy place. I do have a substack. That's where I keep my newsletters. You can find that almost timely substack.com, but the best place to go is Trust Insights AI. You can also catch me. I'm in the AI Business Society. And of course you're going to be at Social Media Marketing World. I will be speaking there on how to use AI to analyze data and make your business better.
- [44:39]
Michael Stelzner
Chris, thank you so much for sharing your insights with us today.
- [44:42]
Chris Penn
Thank you for having me.
- [44:43]
Michael Stelzner
Hey, if you missed anything, we took all the notes for you. You over at social mediaexaminer.com a40 and would you do me a favor? If you're a longtime listener, would you give us a review on whatever platform you're listening on? And if you're brand new, be sure to follow this show on your favorite podcasting app and also let your friends know about this show that really helps us get the word out. We've also got a couple other shows, the Social Media Marketing Podcast and the Social Media Marketing Talk Show. This brings us to the end of the AI Explored Podcast. I'm your host, Michael Stelzer. I'll be back with you next week. I hope you make the best out of your day and may AI help you become more successful.
- [45:23]
Chris Penn
The AI Explored Podcast is a production of Social Media Examiner.
- [45:31]
Michael Stelzner
If you're serious about learning more about AI and marketing, I'll see you at Social Media Marketing World 2025. Go to socialmediamarketingworld.info and secure your spot today.