CHRIS HUNTER
0:36
I am joined here with a bunch of our friends. We meet here weekly to talk about AI and how it pertains to marketing and what’s really going on in the whole AI world when it comes to all that stuff. Today’s topic is going to be interesting because we’re going to take it back a few steps, right? We’ve been geeking out and getting really deep into the weeds on all sorts of things. And while our world is getting more and more complicated, it seems like, right? With all of the AI tools that are coming out, and it’s really hard to keep up with everything because every single one of the major platforms are constantly trying to one-up themselves, right? Not only themselves, but everybody else. You know, as a great example, we’ve got— we’re on an agent’s race right now with every single one of the major platforms. They’re all trying to add this agent functionality to essentially take over any, any task that is dealing with a computer these days, right? So we’re going to take it a few steps back though, okay? And, and we’re going to talk about some of the fundamental type things because we’ve heard a little bit from our audience that, hey, while knowing how to do cloud code is cool, what about ChatGPT? How do I even use it, right? So we’re going to talk about how to get good outputs today. We’re going to talk about how to not get hallucinations. We’re going to talk a also dig into some privacy and safety fundamentals that you need to be aware of, right, with pretty much all the platforms out there.
CHRIS HUNTER
2:08
All right, so let’s go ahead and start off here. Um, you know, Brad, uh, you were talking earlier today, or before we jumped on here, uh, why don’t you tell us a little bit about, um, how do we get a good output from ChatGPT. ChatGPT is the biggest one out there. So let’s, let’s focus on that one for now.
BRAD KILGORE
2:30
Sure, sure. Nice to see you all. Or as Chris would say, y’all.
JOHN CLENDENNING
2:35
Y’all.
BRAD KILGORE
2:37
I lived in Texas back 40 years ago, and it was like a different language when I first moved there because I was a California kid. But that’s a whole nother story. So, you know, I’m seeing some shifts when it comes to getting good outputs. Essentially because even in recent times, it’s like, okay, you really need to have a good prompt. You need to be a prompt engineer, so to speak. And I think that that might be shifting some because I think that the AI models, the LLMs, are getting a little bit smarter, meaning you maybe don’t have to have the most amazing prompt in the world to get good output. However, I think it still helps. I think I’ve shared before on here, we use a prompt tool that’s very inexpensive called PrettyPrompt.
BRAD KILGORE
3:28
And so I can put something in there very simple, create an image of such and such and make it look blue and green, you know, very simple. But then when I use this prompt tool, it will then really refine that to let the AI model know at a much higher level what it should be creating. Because what can happen if you’re just using a basic prompt and you don’t really, you’re not really as familiar with some of the tools, then you can get some not so great results. You may be spinning and wasting time and trying over and over. So that’s one of the first things that comes to mind is at least get a decent prompt in there when it comes to your output. I don’t know that I have, that I’ll jump right into the whole security side of things, but also be aware that if you’re, let’s say you’re creating some written content or something more in-depth, you gotta proofread what it produces because there is a term used in the AI world called hallucinating and the robots, so to speak, will do that periodically, meaning you may just take a cursory glance at what it produced and say, oh man, that looks good. And the next thing you know, you’ve published that or whatever and you have somebody else look at it and go, oh wow, that’s really goofed up. You missed that part.
BRAD KILGORE
4:54
So definitely proofread what it produces, especially if it’s longer content. I know that a lot of times we’ll put together 2,500-word blog posts and it just makes sense to review it, you know, because whether you had a human writer or you had an AI writer, your common sense would say you need to proof that, review it, make sure it makes sense, make sure that it’s legitimately talking about what you’re trying to say. Sometimes in prompts, what we’ll do is say, don’t include case studies. Don’t include prices, you know, give it some, give it some guardrails, give it some, some feedback of what you don’t want it to do. That sometimes can save time.
CHRIS HUNTER
5:41
100%. I think that, you know, and that’s, that brings up a good point of always checking the work, right? You know, just, and what I talk about a lot with, with anyone that I talk with, you know, AI is treat it like a new employee into your organization. They don’t know anything about your organization. They don’t know how to do the actual job., right? And yes, AI has gotten a heck of a lot smarter, but you still need to check the outputs on everything, right? And it’s not just for content, we’re talking everything across the board, right? Um, what do you think, Jen?
JENNIFER CREGO
6:16
Yeah, specifically around this issue, we always put content through another fact-checker prompt. Um, so yes, we also have a human in the loop, but In addition to that, we want to make sure that not only, not only, uh, are we using AI as a fact checker, you know, to make sure that, that there isn’t hallucinating, but to make sure that if we are including some sort of facts that we’re saying, uh, do we have a link to where this was? Like, says, you know, a, a study shows— okay, give us a link to that study. And we also have to check to make sure that that link is actually valid, because sometimes at least it used to do this. We, you know, have the guardrails in now, but it would include links to studies that the— it was a broken link. And so that does us no good. So using AI to, to check it in, in multiple ways like that, I think is really helpful. And then also, in addition to fact-checking in general, if, if we’re writing content for a client, then we also want to tell the AI, fact-check it based on what is already available on the client’s website. Um, that way, like, if it’s saying something that is, it’s just out of alignment, like, yes, we have like the, the custom GPTs and the brand and all of that, but we also want to make sure that it doesn’t contradict anything that’s already on the website.
JENNIFER CREGO
7:57
And so if it does, it will, you know, flag it for, again, to bring the human into the loop to check that.
CHRIS HUNTER
8:05
100%. And so that brings up a good point, you know, on hallucinations. And that’s kind of on our topic here is how do we ensure that the outputs that we’re getting isn’t hallucinations, you know? John, why don’t you weigh in on that one?
JOHN CLENDENNING
8:23
Yeah, so I think what we’re wrapping this all around is, and Brad mentioned it well, and then Jen dove right into it, it’s so, is prompt engineering isn’t as important anymore? You and I, Chris, both know that when you start working with an agent, you don’t do any prompt engineering. You just have a conversation. It figures out the prompt., and then it goes and does a task because it has a skill. So we’ve just added a whole bunch of little things right here and, and there. But skills are basically, like custom GPTs back in the day, skills are transferable. They don’t just have to be on Claude. You can take a skill to any one of the LLMs. A skill is a set of instructions that tell you how to do something.
JOHN CLENDENNING
9:03
So what I kind of wanna put that into bubble is if you’ve got a, what I call a chat in the wild, this is how I explain it to my team. If you just go into ChatGPT, whether you’re on the, you know, and you’re on the paid or you’re not on the paid plan, then you can’t have projects. So you you’re just walking in and starting a chat and you just start the chat up, it is rogue. It is like, I know everything, I know too much, you didn’t give me anything. If you start with, as Jen said, say you’ve got the entire, an entire Google Doc on the client’s voice, the client’s this, the client’s that, you drop that in, right? As a reference context. Then you drop in how you want blog posts to be written with, you know, a summary at the top and, you know, key features and then the rest of the post and you want all the, you know, instructions like that. That, then you say, hey, based on my two control documents here, here’s some new information. You know, can you write me a blog post? Ask many questions before, you know, getting started, blah, blah, blah.
JOHN CLENDENNING
10:00
Right now you’ve got this, um, you’ve given it context, you’ve given it control, and then your output is going to be a lot more based on that. Now that gets into the world if you’ve got a project. Projects are just an isolated bucket. You can put those instructions at the top of a project and have conversations forever, have individual chats that are only inside that project. So the project has a knowledge base that it always draws back against. And now you’ve made that a little bit easier. So it’s worth $20 a month just for your ChatGPT or whatever. So I think those are a couple things.
JOHN CLENDENNING
10:36
And then if you’re ever having like a, I know Laura was gonna touch on, we’re talking a little bit about, you know, a lot of people don’t know how insecure these, these, these LLMs are. Everything you type in there is public unless you either go into the, the private mode. There’s an option in ChatGPT in the top corner where it’s temporary mode where nothing is shared, that kind of thing. You can set your privacy settings. They’ve added two-factor authentication, um, just like you should be doing on Facebook because people can get into your Facebook account and then, you know, and go rogue and steal all your stuff. Um, people can get into your ChatGPT too, um, if they figure out your username and password on stuff like that. So you should have the two-factor authentication now, which wasn’t there like, what, 2 months ago? But it’s there now on the backend, and they’re recommending everybody set that up. So think of your privacy, think of context when you’re doing stuff, and you’ll get much better output, and you can kind of control your environment a little bit better.
LAURA SUTHERLY
11:30
I think that’s well said, that all of us, no matter what, need to take time to look at our settings and see what’s currently in there, what they have updated, what is included, what’s not included. I know one of the key things that we like to mention is if you’re not going to email it out to the world, don’t put it, don’t copy and paste it in there. So you do want to think about what you’re putting there because it does start to seem very intimate and you’ve got this private conversation going on with your computer or your phone and you don’t think, oh, it’s not going to go anywhere. But that’s the thing is depending on your settings, it could be training other LLMs or the AI is using it for other fashions. So definitely take the time to check those settings and see what you want out there and what you want them to be. So the other thing too is you can sanitize your inputs. So what we mean by that is maybe be a little bit vague. Don’t talk in specifics.
LAURA SUTHERLY
12:33
So if I’ll just use a tractor as an instance, so you don’t have to put everything specific of your serial number and everything about your tractor to find out information. You can go ahead and be vague and just state the year of the tractor, the, you know, some basic things. And once that’s in there, then you can decide what to do with it. So think about being a little bit vague once in a while and not be so specific.
JOHN CLENDENNING
12:58
Don’t tell it your social insurance number.
LAURA SUTHERLY
13:01
No, no, I definitely don’t want to do that.
BRAD KILGORE
13:03
That’s exactly right. Don’t do it. Yes. I know that there are some settings in general that you can adjust in the different LLMs. Do any of you recall exactly some of those settings that you would want to keep where it’s basically you’re saying, don’t use my information to train the model? Do you guys remember any of that?
JOHN CLENDENNING
13:31
Yeah, yeah, go through your settings. It’s in there. You think like there’s like ChatGPT has the whole security section. So you want to go there as well and double check sort of your setup and security and stuff like that. But yeah, just pop through the, ’cause they keep adding more layers to it, but just make sure that, you know, things like memory and training and stuff like that. I know Gemini has, you know, really harsh settings or really good settings on, you know, do you want this to be used anywhere else at all kind of idea and asks you really straight up. The other ones are a little bit harder to find or a little bit, you know, more nuanced. But yeah, they all have an ability to hit a temporary chat that is completely off the beaten path and will not do anything if it’s going to be that private.
BRAD KILGORE
14:20
Yep.
LAURA SUTHERLY
14:20
Always a good thing. Always a good thing. Back to what we’re talking about, the facts and stuff. One other thing that we also like to do is ask it to flag assumptions because again, AI likes to be very pleasing, likes to provide you all this great information. But if you ask it to flag any assumptions, that is helpful. Also asking 1 or 2 qualifying questions after what, you know, like clarifying questions, not so much qualifying, but clarifying questions is also a way to kind of cross reference what it is putting in front of you. And again, to everybody’s point so far today, is human in the loop. Remember, there was only a few years ago we did not have this type of assistance, and we were very capable of writing phenomenal blogs with our team members.
LAURA SUTHERLY
15:10
So keep ourselves in the loop to verify, make sure we are signing off on whatever might be put out. And to Jen’s point, definitely click on the links because it does just like to throw in those pretty blue underlined links, and we take it for granted that those are valuable assets, and you can find a lot of dead links within it.
JENNIFER CREGO
15:31
Yeah, well, I’ll tell you, we don’t actually check— or I’m sorry, we don’t click on the links at that point. We tell AI, like, we tell AI to check the links, and then we, we do eventually, but at first we just let— make sure we’re sure that we’re prompting the, the LLM to check those links after they’ve added or it has added.
JOHN CLENDENNING
15:53
So good point. One custom instruction that I learned recently. So inside your settings under the personalization in, in chat, I’m trying to remember where it is in Claude right now, but you should set how you want it to, you know, to talk to you. Do you want it to be conversational, talkative? Light, humorous, with mine, call me, hey dude, or hey JC, instead of being all that kind of, instead of being formal. But one of the things that you should set once that I did a long time ago, and I just found it on my phone here, putting in the bottom of those custom instructions, prioritize substance over compliments. Never soften criticism. If an idea has holes, say so directly. This won’t scale because of X.
JOHN CLENDENNING
16:39
Is better than have you considered? Challenge assumptions, point out errors. Useful feedback matters more than comfortable feedback. If you tell it that, and that’s the wording that this one fellow gave, then you’re not gonna constantly get, you know, again, do you want the ego stroke? Okay, you’re great, you’re great, you’re great. I’ll tell you you’re great right now. So now you got it, write it down. You don’t need ChatGPT to always tell you that you’re great if your idea sucks. Actually have it push back a little bit and go, well, maybe that’s not a good idea. Kind of idea.
JOHN CLENDENNING
17:10
So, because it does default to be your most agreeable friend, and sometimes that agreeable friend that agrees with everything gets a little annoying.
CHRIS HUNTER
17:19
Well, I think that we brought up a lot of valid points there, and that just kind of wraps everything up. Does anyone have anything else to kind of add on to everything that we just talked about?
BRAD KILGORE
17:32
Check your work. It’s like what was shared earlier. You know, an AI assistant or an AI tool is just that. It’s just like if somebody worked for you, you still want to look at what they produce because nobody’s perfect and things can go goofy sometimes. So I think that’d be the tip of the day is check your work and don’t just trust that it’s always going to be right.
JENNIFER CREGO
17:57
Yeah. And check the privacy settings of your account. Check it. Make sure that you, that you know where you stand, whether you whether that means that you upgrade to a paid account or you’re fine where you’re at, whatever it is. But check the settings and make sure you know where you stand.
JOHN CLENDENNING
18:15
Yeah, yeah. Especially today with all the agents coming out. OpenClaw, you don’t have to be into that, but like Chris was chatting to me about the Perplexity computer, as they call it, which is their workflow. Claude Cowork, they’re all the, the, the agentic race is on. We’ll have an episode about that soon. But they’re all doing that and that is rogue like crazy if you don’t know what you’re doing. Like, that’s, that’s go in and answer emails and move things around and do all that kind of stuff. You, you better kind of know the basics we talked about here today before you ever just decide to say, yeah, just go do it, I’m going for lunch, and, you know, come back to whatever happened while you were gone.
LAURA SUTHERLY
18:53
So yeah, that comes back to a strong foundation. But like you said, we’ll, we’ll visit on that because if you’ve, if you’ve if you’re unorganized now, it just amplifies that you’re unorganized.
BRAD KILGORE
19:06
Exactly.
JOHN CLENDENNING
19:07
Yeah.
BRAD KILGORE
19:07
Okay.
CHRIS HUNTER
19:07
All right, y’all. Thanks so much for tuning in. This is AI Marketing Experts Podcast. Be sure to hit like and subscribe down below. Make sure you tune in every week because we are bringing it with all sorts of cool stuff that we’re talking about. Yes, we did go back and go some fun, you know, foundational type stuff today, but We’re always talking about the newest, latest, and greatest and all that kind of stuff in the AI and marketing world. So with that being said, thanks for joining us and have a great week, y’all.
