Prompt engineering was supposed to be the essential new skill. It was not. Where prompts still matter is work you repeat and need right every time.

Jed Mahrle sat down with Kyle Vamvouris, founder and CEO of SalesThread, for part five of the AI Foundations series. Kyle's read on why the hype cooled: the models got very good, and the harness built around them got good enough to absorb sloppy prompting.

So when is a real prompt worth writing?

Only for repeatable work. Kyle's test is to ask what task needs the output right every single time. For one-offs, a sentence or two is fine. He also notes a prompt and a skill are close to the same thing, a skill usually having a prompt at its base.

Stop telling it not to hallucinate

Kyle's most useful reframe. The entire output is a hallucination. There is no mechanical difference between a hallucination and a correct answer, because the model has no concept of either, so "do not hallucinate" does nothing.

His model for why prompt quality still matters: every possible output already exists in the training data, and your prompt selects a region of that space. A vague prompt gets you roughly eighty percent of what you wanted. A refined prompt makes your ideal output one of the candidates the model can pick.

The three parts of a strong prompt

Objective. Give step-by-step instructions, and Kyle's bar is whether a junior colleague could follow them and do a decent job. Focus on the sequence of actions required, and be specific.

Data. Only give it the context it actually needs. Kyle is blunt that more context being better is backwards, which is why the field moved to context engineering. Specify the output structure or you get thirty pages. And include examples of the input, not just the ideal output, so it can pattern match.

Design. He cites the Lost in the Middle research on models prioritising the top and bottom of the context window while instructions in the middle get missed. So put the important things first and last. His structure: objective, rules, steps, examples, output format.

Prompts should be written by leaders and handed to the team

Kyle's strongest opinion. Companies should own the prompts and workflows their teams run on rather than leaving every rep to invent their own.

The workflow behind a 3.2 million view video

Kyle's Instagram case study, and the clearest self-learning loop anyone has shown on this show. Performance data lands in a spreadsheet from the Meta API, no AI involved. One agent analyses it the way he used to by hand. A content strategy agent reads both and Slack messages him each morning with what to film, what to re-record, and what held a video back. A third agent pulls best practices from high performers into a document that feeds back into the strategy agent's context.

His floor moved from a few hundred views to six thousand, then fifty-eight thousand, then one at 3.2 million.

The caveat matters as much as the workflow. He keeps a human on the best-practices step, because AI judges best practices badly and bad ones poison the loop.

Resources mentioned

• The Sell Better AI Prompt Library
• Free ZoomInfo MCP Access for ChatGPT

The Speakers:

• Jed Mahrle, Host, Sell Better
• Kyle Vamvouris, Founder, SalesThread

Catch The Daily Sales Show live

Follow Sell Better

Explore our YouTube Channel

Watch the replay

Thank you to our sponsor: ZoomInfo

Today, we're talking about how to build effective prompts. This is, like I said, I think part 5 or 6 of the AI Foundation series. We'll drop a link in the chat, or if you head over to the website, you can see the previous AI Foundation series that we've done, we've been covering everything related to AI. Today's gonna be a lot more… we've got a ton of good stuff to share, but we also want to make this as valuable as possible for you guys, so we're gonna be throwing up some polls. I want to encourage you guys to ask, you know, any questions you have so we can make this as relevant as possible for you guys. At this point, I feel like most of you, if you've been here before, you know Kyle Vamvouris. Follow him on Instagram, he's doing some cool stuff over there, some of his posts are blowing out, but he posts just really tactical.

AI sales-related content. I think he blends the two really well and makes it very easy to digest. Some people are just a little bit over my head, but every time I do a show with Kyle, even as much as I am ingrained in AI, I learn something new from him. So check him out, check out SalesThread as well. He's got some cool resources on the website. My name is Jed Mahrle, I'm the founder of Practical Prospecting, frequent host here at Sell Better. And if you guys don't know Sell Better, head over to the website, check out the YouTube channel, there's tons of free content. I always tell people this, but I think there's, like, hundreds of pre-recorded shows we've done in the past, so if you're ever curious about a topic, if you want to type in AI, you'll see all the previous AI shows we've done, or cold calling, cold email, whatever it might be, you'll find tons of relevant content there. I want to say thank you to our partner, ZoomInfo, for making these shows possible. I was super excited about this. I actually want to share a different part of my screen to show you guys this. If you go, we're gonna drop a link in the chat, but this is a really cool.

opportunity, if I can get out of my screen here. They are giving free access to the ChatGPT MCP. For those of you who don't know what an MCP is, I'm gonna ask Kyle in a second. He can explain it better than I can. Essentially, it's like an integration into ChatGPT, but if you click in and you have ChatGPT, you can access it immediately, connect it, and start using it. This is a really cool opportunity, so I wanted to share this with you guys. Kyle, for the people in the audience who don't know what an MCP is, What's your, simplest explanation?

Yeah, I mean, I'm gonna way oversimplify this, but what you can think of is an MCP server is a connection between the tool, so in this case ZoomInfo, and ChatGPT, Claude, whatever tool that you use for your AI stuff. What makes it interesting is what ZoomInfo does is they sit down, they go, okay, we're gonna create an MCP server that's gonna help the AI do things in ZoomInfo, and they provide the tool that you use with all of the different things it can do, a description of what that thing is, so when you ask a question in the chat GPT, it can go and look and say, like, alright, well, the ZoomInfo MCP, I have All of this kind of stuff, and then there's prompts behind each one of those, which is kind of related to what we're talking about today, so it's perfect. to help your AI do things correctly in their tools. And they kind of, like, pre-build this package of tools, is what you would call it, that are specific to ZoomInfo and, using ZoomInfo so your AI tool can actually use it. It's pretty sweet.

Love it. Yeah, good explanation. I appreciate it, man. We've got, like, 90 of you here with us already. I'm gonna throw up a poll here, I wanna see who is in the room. Out of the 90 of you here, USDRs, leaders. account executives, go ahead and fill out that poll so we can make this more relevant to you. Last slide today, and then we're going to be screen sharing and keeping it conversational, but for today's agenda, we're going to talk about the anatomy of a high-performing prompt, what separates good outputs from bad outputs, how to feed the AI the right inputs to get the specific results you want, and how to turn prompts into workflows. We're going to be throwing up a poll later, to sort of make this a choose-your-own-adventure. We're gonna take this in the direction that you guys want the most, because there's a lot of things that Kyle and I can talk about. But with that said, I'm going to go ahead and share the poll results real quick.

We've got 20, you know, 24% of you SDRs, BDRs, shout out to you, 31% AEs, a lot of senior leadership here, we've got almost 30%, and for the 14% of you who are other, let us know in the chat. Are you a founder, marketing, something else? I'm always curious to see who else is here. But Kyle, with that said, before we get into all the extra stuff, let's start off, because we talked a lot about prompts. I guess, like, a good starting point, how do you… you know, how has prompting changed over the last few years, and how do you kind of separate what a good prompt looks like from a bad prompt? Let's kind of start there, and then we can kind of dig into the details.

Yeah, so, I would say, like, kind of funny to look back, because I remember about maybe 2 years ago, maybe even a year and a half ago, prompt engineering was like, this is the new skill you have to learn, this is the most important skill of all time, you're gonna fall… you're gonna be a dinosaur if you don't learn this thing. And, like, that is… that turned out not to really be true. What's happened is these… one, the LLMs have gotten really good, so the models themselves are really good. But then, we were talking about this before the show, Jed, the harness, so the application built around the LLM has gotten so good, so it reduces the pressure on us to write really good prompts. Everyone here probably has decent or great experiences with AI. You know, it's a scale, sometimes it's really bad, sometimes it's incredible, but I'm sure most people here don't write very sophisticated prompts, and they get pretty good outputs from AI, and they're happy with that.

Where you want prompts, and kind of the difference between a good prompt and a bad prompt, but where you want prompts is for repeatable work. Another term for a prompt is a skill. Skills can be a little bit… there's a little bit more to a skill sometimes than just a simple prompt, but usually the base of a skill is a prompt. And there's things that you do repeatedly, and you want, when you use that prompt or that skill, to get the best possible outcome you could get. So what I want everybody thinking about here is. what is a task or something I use AI to do where I want the output of that to be perfect every single time? So, it might be something very repeatable, like I need to create a report from all of this data, and a good version of a prompt would be, like, a very detailed step-by-step set of instructions, and I'll go over, like, specifically how to write good prompts in a minute, but step-by-step instructions, examples of what good looks like, and a bad prompt would be like, hey, can you analyze this data? share with me what do you think the three biggest levers are for us to improve our team's performance?

That might be okay, and in some ways, you'll be impressed by how good that output is, considering how quickly it was able to do it, and how easy it was for you just to ask it to do it. But the better your prompt is, not only the stronger your output will be, and more accurate it'll be, but it'll also be more aligned with how you would do it as a human. And that's ultimately what we're trying to do with prompts, is we're trying to get the LLM to behave Like us as a human, with our expertise, with our lens that we use in this example to analyze that data, to get as close as we can to the work that we would have done ourselves.

Yeah. Yeah, I think you make a good point, because I've noticed that too, where it's like, a lot of times you don't need to go crazy, and you're still getting decent outputs, but I've found…there are so many AI-related projects or workflows that I've just kind of given up on, because I didn't take it that last 20%, that last 10%, if you will, because I think it's really easy to create an AI workflow or prompt that does, like, a good enough job that you're kind of happy with. And I'll give, like, cold email messaging as an example. anyone can kind of say, like, okay, keep it under this amount of words, you know, don't use M-dashes, blah blah blah, you know, whatever. But, like, to truly get a prompt or workflow to the point where it's something that you would be comfortable with with shipping if you wrote it manually, I do feel like it's that last mile that matters the most, because, like I said, I'm sure other people can relate, maybe you can't too, Kyle, but just, you know, the first step is always easy, like, getting the workflow how you want it, but then getting it to a

point where, like, you trust your team to use it, or your clients to use it, or you actually trust it without additional editing, takes, you know, takes that last, extra bit of effort.

Yeah, and let's… why don't we talk about that for a minute? I can show a visual here. While I do this. Also, I just realized we're matching. Look at our shirts.

Oh, yeah.

Look at us! You got my text this morning. Okay, if you can type into the chat here, what's your biggest frustration with prompts, like, the prompts that you use? Is it outputs are inconsistent? Is it you are trying to get to that last 20% and haven't yet? I'd like to get a little bit of sense of that, and while you're doing that, I want to share with you a really important thing to understand when it comes to using AI, and this is important because it helps it helps ground us in the reality of what LLMs are, so then that way we can be thoughtful when we're prompting, because we know exactly what's happening under the surface. And I'm gonna heavily oversimplify some stuff here, but what you're looking at is, like, in a…Claude or ChatGPT Chat.

when you type in something, a prompt, right, that's what we call a prompt. When we type something into our little chatbot and press enter, that gets added to what's called a system prompt. That's behind the scenes, you don't see it. It gets added into that system prompt, and that's what gets sent to the LLM, okay? It's important to know that that exists, because you don't write this. Anthropic wrote this for you. OpenAI wrote this for you. And your prompt is here. And all of this gets turned into tokens. Can I get a yes or a no in the chat if you're familiar with what a token is? I'm gonna read some of these responses, Okay, perfect. Okay, great. Awesome. What a token is, is, every word that you typed in, including the system prompt, gets broken down into little chunks and turned into numbers, okay? So what you're looking at here is an example of how that happens. You have a…

The word large is one token, this space and the word language is another token, models is a token, this space and the part of the parentheses is a token, LL is a token, MS is a token, you get the idea. So everything gets broken down, sometimes it's one full word. In the example of tokenization, that's broken down into two tokens, and all these tokens are numbers. So what we do is we convert… All of the text. into tokens, into numbers. And it goes into… this is very important, I hope everyone takes notes on this part. The box of math. A bunch of math happens in this box. I'm not going to explain all of the math that happens, but there's transformers, there's a whole bunch of crazy stuff that happens. I'll show you a little bit of a visual of some of the things that happen. A bunch of math takes place, and this is what makes it so sophisticated. The reason why AI LLMs work the way they work is because people have figured out this incredible

bunch of math that results in us getting what we want. But how do we get what we want? The math happens. And then the output is a bunch of numbers, which are tokens, which get converted back into text, and that's how we get the output. That we get. So, this is…how LLMs work fundamentally. I'm going to show you one more part of this. Now you… now a lot of you, I can see, didn't know what tokens are, so now you should have a good understanding of, of what a token is. And then I want to call out something else that Sean brought up, which is hallucin… like, one of his frustrations is getting hallucinations when asking for no hallucinations. And LLM, by the way, the entire output is a hallucination. There's no difference between a hallucination in the LLM and the rest of the output. The whole thing's a hallucination. It doesn't know it's hallucinating, so you can remove Don't hallucinate in your prompt. It has no concept of what hallucinating is. Remember, it's just a bunch of math.

It does all of this math, and based on the tokens it received, which is from the words you sent. it spits out whatever the output is. If that's a hallucination, guess what? The words we did caused that. Now, I say this because we have control of how good our outputs is by the prompts that we write. And I want to go down here, I have a good, I have a whole bunch of…stuff here.

Oh, wow.

what I want to… yeah, I know. A lot of good stuff. Yeah, I get a lot of questions around. A big part of what I do is I'll do, like, trainings and stuff around AI, like this one here, and a bunch of stuff comes up, so I have a bunch of drawings. So, anyway, keep asking questions, you might get more drawings, very secret drawings. Okay, but we write our prompt. Remember, I already explained to you what happens here. We send our prompt in, and we do the best we can. We write as good of a prompt as we can, and the goal is…Let's get that perfect output that we're looking for. I want this to… I want the perfect cold email, just exactly how Jed would write it every single time. Well, what you're looking at here, in this sea of blue dots, is every possible output you could get. That's all predetermined in the training data.

Every single possible output is predetermined in the training data. It's another important thing. This is why, like, you know, some people debate… I do a lot of content around AI on my Instagram, and I get in debates sometimes with folks who are like, oh, the LLMs follow, like, a similar process as a human brain, and, like, the human brain's just predicting, so is LLMs, it's just predicting the next word. We function very differently. We don't have all these predetermined outcomes. LLMs do. And what we want is we want to get as close as possible to our perfect output, which would be this little green dot, let's say. And when you type in your prompt, what happens is, based on those words, a section of the… all the possible outputs you could get is selected.

If your prompt causes this group to be selected, you'll be somewhat close to your perfect response, but you're probably going to be 20% not there. So, like, 80% of the way there. You'll get one of these blue dots. Again, these blue dots will be somewhat similar, because they're pretty close to your perfect output. But they're not… it's not the perfect output. The only way to get the perfect output is writing a prompt so good that the perfect output becomes one of the possible answers here, and then the LLM selects from one of these possible answers. You're actually… it's possible that you get the green one, and then…Boom, that's your final output. This is why if you have a really well-refined prompt. And Maria, maybe we do the, Choose Your Own Adventure, poll right now, which, we'll explain in a minute. But, That's why…when you refine a prompt over and over again, you can get it to a place where it's, like, 99% of the way there, whatever the output you're looking for is. Jed, you wanna walk them through this poll that we're gonna ask them here?

Yeah, by the way, I love the visuals. Like I said, makes it really simple to me. I also want to reiterate something you've said before on previous shows, which is that, like, and you kind of already said it, but it was a good way for me to understand AIs, is that it's just… it's like a really smart, what do you call it? Autocomplete. Autocomplete, yep, exactly. And so I think it's a really good perspective to look at it from. But yeah, so we're throwing up a poll here, guys. We want to know what you care about most. I'm already seeing the trend. You know, how to write a really good prompt, we can dive in deeper to that. How do you refine a prompt that you have that is already good and get it to that next level? Or how can I take prompts and use them in a workflow? So, so far, I'm not sure if you're seeing this, Kyle, but from what I'm seeing, 55% so far are already leaning towards how to write a really good prompt. Number two is how to take and use it in a workflow. So maybe give it a couple more seconds here and see if there's a comeback, otherwise it looks like, how do you write a really.

And I can mix those two together, because I can show a workflow, and then we'll go into a prompt. And if somebody, if you guys could type in, if there's a prompt that you've been wanting to write, or one that you have that you're like, hey, I wish, like, I had a better prompt to do X. If you could type in… I wish I had a better prompt to do X, and then whatever, you know, obviously type in what X is. We'll pick one, and I can work on a prompt live and show you how I go through writing a really good prompt. I… let's do, let's do this then, Jed. Let me just talk about some of the best practices of writing a prompt while we're getting that, and then I can do the workflow, and then we can, build a prompt together. Yeah. Okay, I'm gonna go quick through this, because I have a prompt writing prompt that, writes prompts for you. So I just want you to understand the structure, and I think you'll learn a lot more when I go through examples, so I'm gonna go real quick through this, but it's in the recording if you want to watch it again. The first…

part of having a really strong prompt is a really clear objective, and there's three things you need to know about, having an objective, or a set of instructions. There's two parts of a prompt when it comes to objective. One is the actual objective itself, what goal you're giving it, and then the other one are the steps you're asking it to follow to, complete that goal. And the first thing I want to highlight here, and this is critical, especially, I think, what was it, like, 25% of people here are leadership. This is a really, really, really important thing, because I expect leaders to do a lot of the architecture around how the team uses AI. I think prompts should be written by leaders and given to the team. It can be done in collaboration and stuff, but I think companies should take more of an effort in owning the prompts and the workflows that use AI. Anyway, the first thing here is providing step-by-step instructions, and this is where I want to make sure everyone is clear. Like, you're delegating to a junior colleague.

AI is not, like, all-knowing and special. It's not as good as humans, yet. And because of that, I like to think, okay, let's pretend this person's a junior colleague. If I couldn't give my prompt to a junior colleague and they'd be able to do a decent job, it's probably not good enough. The second thing here is focus your instructions on the sequence of actions required to complete the task. The sequence of actions. And this might be your sequence of actions, right? So, like, one of the things I like to do, like, let's break down a workflow. What's the hu… what are the human steps in that workflow? And what's the steps for the AI? They map, but sometimes an AI needs two steps, right? But by giving it the human step-by-step breakdown, it's more likely to be accurate. And then number three here is be specific about you about what you want the AI to do. Right, be very, very specific. So that's the first set of things to understand. Again, we'll go over an example, and I'll go into more detail there.

Data. This is the… this is really the context. What does the LLM need to know in order to do this job well? This helps us remember our green circle? Whoop! A big part of the green circle hitting the right spot is by being very thoughtful about what data we give it, what context we give it. The first thing is, you only give the AI background information it actually needs. A lot of people think more context is better. It's the opposite. It's the opposite. In fact, you may have heard the term context engineering, that's sort of the evolution of prompt engineering. Now they're starting to talk about harness engineering, which is kind of a rebrand of context engineering in a lot of ways. It's because you're trying to control what you put in the context window of the LLM so you can get better outputs. So you want to think with your prompt, what do I need to give the AI in order to do a good enough job? The next thing here is include details about how to structure the output you want structured, like, hey, this is how I want you to present the data. If not, we're going to get, like, a 30-page document.

And then the third thing here is have examples of the input itself that you… that it receives. So let's say you have a prompt to write, LinkedIn messages. The input might be who you're writing the, the email to, and you know, some details about your offering and maybe them, okay? That might be the input. And if you give, within your prompt, examples of inputs, and then the perfect LinkedIn message output. It can start pattern matching a bit there, too. That's the data part. Again, we'll go into an example here. And then the third part of it is design. How do you structure your prompts? Because that does make a difference. There's… if you're nerdy about AI, you can go read the, it came out at the end of last year, if I remember right, a research paper called Lost in the Middle.

It talks about how LLMs typically prioritize what's at the top and bottom of the context window, and things in the middle get lost sometimes, or instruction following doesn't happen as well in the middle. So I recommend people have a logical structure to your prompt that a human could follow easily, define all the rules about what to do and what not to do, and then organize your prompts in a way to where the most important details are at the beginning and the end. And I have this little visual if you wanna… there's… I have it for image prompts, too, if you wanna check it out. But, I'll let people screenshot this if they want. here's our objective. Our objective is going to reference that there's steps that need to be followed. There's rules and reminders, there's the steps themselves, there's the examples, and then there's the output format that you want. And it's a similar structure with image prompts. Again, you're only gonna write these prompts for work that you do repeatedly. If you're just doing random stuff, you're fine just saying, like, oh, do this, do this, and writing a couple of sentences. Like, you're gonna get decent enough responses, and then you're gonna give more feedback.

feedback, and it'll be refined over time. But if you want something that's part of a workflow, this is a job to be done, and I want this job done for me, and I don't want to spend time on this thing, I want the output itself, and I'll give you an example of this as we talk about workflows, then that's when you're going to want to write a prompt. Alright, let me pause. Jed, anything you want to add here?

No, I think, like, first off, I think it was important that you called out, like. identify the things that you're doing repeatedly, because there's no reason to, like, go through all these crazy steps if it's just a one-off thing, right? Like, this is the sort of, like, the stuff that you do repeatedly that's important. Matt had asked a question around, like, hey, let me pull up his question here, it was around email messaging. But he said he wants… he… hold on…I wish I had a better prompt to… for cold email outreach sequences. I'd put in the problem we solved, the person's title, and their company, and I'd write a short, strong email that isn't cringey. So, Kyle, you may have prompts for this, I know we have our AI prompt library, which we'll share in the chat. I will just give my feedback on that, and just say that, like, two of the most… two of the things that have really helped me with prompts is one, having, like, Super Whisper, Wispr Flow, or any of these tools where you can just talk to the LM and just talk… dump out everything logically. Because cold emailing, specifically for me, there's so many variables.

And this was something I struggled with, is we needed a cloud project or workflow that would help create the initial messaging for our clients. And our clients are super broad. They sell to different ICPs, they all have different offers. And so, I needed to talk through everything, you know, depending on the type of ICP, this is when you should do this, and that just… that was the easiest way for me to talk through it. And then, at least in my experience, and I'm curious to hear your opinion too, colleagues, you just have to do it, and then… and then find the edge cases. And then update, and keep going through it. And, you know, one thing I was struggling with was, like. you know, my email messaging kept, it kept outputting, like, these super long CTAs or CTAs that I didn't like. And so I would just ask Claude, like, hey. the prompt I have is I'm trying for it to achieve X, but I keep running into this issue. Here's my prompt, can you help me fix it? Like, these… you have to go through an iterative flow, and I think a lot of people don't want to spend that time, but a lot of times that's necessary and worth it if it's something you're going to keep using over and over again.

Oh, totally. You need to refine your prompts. Very, very important. And you can, like, even with what Josh is saying here, like, as I get an output, if the output doesn't look great, I want to understand why it doesn't look great, and then make a tweak to the prompt, a tweak to the context that I'm giving the prompt. You're going to have to do that refining. and then you'll get a prompt to a great place. And because, it seemed like the more… the direction people wanted to take this was, how do you write a really good prompt, as well as the workflow itself? Maybe, Jed, I'll jump into a workflow that I actively use. I liked, in terms of a prompt, Margaret here, and if you can keep track of this, make sure we don't lose it, I wish I had a better prompt to find out if my timing is right to reach out to certain accounts. I like this one a lot, so let's work on this prompt live. Margaret, if you can type out, just give me, like, a couple of bullets of how you as a human know if it's the right time to reach out to certain accounts.

So how would you make that decision? So give me, like, 5 bullets, or whatever you would do. Okay, and then I'm going to… I'll show you a workflow. This is a workflow that I built for myself. So I've been creating content on social media for a really long time. It has been abysmal. Other than… I mean, like, LinkedIn's been fine. And, like, I've tried a bunch of YouTube stuff, I've tried Instagram for a long time. I'm one of these people that can, like, talk off the top of my head, I'm comfortable on camera. It always drove me nuts that, like, LinkedIn worked well for me, and I hate writing. It's like, can't I just talk to a camera? It's, like, so much better for me. It makes me happy. So here's what I did. I'll give you an example here. I started focusing a lot on my Instagram.

And I would get, like, a couple hundred views, and then I started doing more AI content versus sales content, and I started getting, like, 600 views here. So if you look down here, you can see, like, oh, we got the 600s here, okay? And that was way better than what I was doing before, so I just started doing more AI. You can see, like, all this stuff is more AI-driven stuff. And then I had one that did, 1700, which was good, and then I had one that did 1,000, and then 700, 500, 600, started getting better. And then I had one that did 6,400, which is, this one here was about, Anthropic, okay? I forget exactly what it was about. This was a big signal for me. And every time I had a post that did well, so like that one there, I would re… look at the data, and then I'd either re-record that video with better structure, or I would do a similar topic, whatever it is. I had all this stuff that I would check. And then right here.

I got 58,000 views. This was by far the most… I mean, 6,000 was the highest viewed video I'd ever gotten, 58,000 was the highest viewed. So then I was like, oh, I'm on to something. So I kept looking at the data, I kept refining. You could see, like, now I had 2,400, 2,100. I'm like, oh, great, like, my floor, this is 5,600. My floor is increasing. So, the next step I did was, I was like, alright, well, I have a good workflow as a human, let me build an AI agent to do this. I'm gonna show you the workflow of that, and then we're gonna dive into a prompt. But, so I built an AI agent to do what I was doing with the, with the data analysis, and it would send me a report every morning. Then I had a video do 17,000, 12,000, blah blah blah blah blah. This one did 3.2 million views. And since then, I've had several that do, like, 50,000, 100,000, 200,000, so I'm getting a lot more traction. It's all because my AI agent tells me what to record every morning.

It just says, hey, Kyle, based on the performance of the recent videos, you should re-record this one, you should shoot a new video on this kind of topic here. For this one that you're re-recording, the hook, I think, really held it back. You should make these changes. That was all based on what I was doing manually earlier when I was scrolling. And what that looks like… is this. I'll do the simple version of it. This is a workflow that I've mapped out. I post something on Instagram. Right here, this is a, this gets updated from the Meta API. This is a spreadsheet, it looks like this. The spreadsheet has all of my performance metrics in it, okay? This is all my performance metrics. That spreadsheet. gets updated from Meta's API. I'm not using AI to update that spreadsheet with that data. So the performance data comes directly from Meta. Then… It gets enriched with

a performance analysis that the AI agent is doing. That performance analysis is based on my own personal analysis. I would look at the Instagram post, I'd look at the data, I'd say, okay, the hook didn't work. This was shared heavily, but it didn't get a lot of views, but the people who watched it shared it. This is probably going to be a good video, because shares ranked really highly. that's the kind of thing that I did. I wrote a prompt that does that, right? This agent analyzes the data, it updates it in the performance sheet, but the output of this here, goes to my Content Strategy AI agent. What that does is it looks at all of the data in the performance data sheet, as well as the analysis from the agent that does performance analysis, and then determines what Kyle should do every morning. And it Slack messages me, Kyle, you should do this. And then I say, oh yes, and I bow down, and I light a candle in my shrine.

The other thing that's going on here is the performance data is also… there's an agent that looks at that reel, the transcript, and the performance data, and says, hey, did anything work exceptionally well here that's worth saving as a best practice? And then it updates this document with best practices. That's also part of the context window of the content strategy agent. That is now a self-learning loop. as we do things, and a lot of my Instagram, it's… my Instagram is either me off-the-cuff just riffing on something, or me structuring it because something I off-the-cuff riffed went… did well. So now I'm doing a more structured version of it, and those always perform better, which is annoying, but that's just how it is. So…that is, like, what the system looks like. I'm gonna transition to the prompts here. Anything you want to add, Jed?

I think the most interesting thing you said there was, like, the self-improving loop. Yeah. I think that's the coolest part about AI, where I'm starting to do that with our cold email messaging, right? We want to run a lot of campaigns, we're having it look at the email data, and then improve the prompts that we're using to write that messaging. I think that's the next level stuff that's, that, you know. just takes it, again, takes it to the next level, because you're not always having to be the one that has to go in and improve the prompt, etc. So yeah, I think, you know, this is just one example that Kyle's using for Instagram, but there's… there's so many ways you can apply these types of workflows to everything you're doing. And it looks like, you know, Margaret gave…some examples on that timing prompt she was talking about. I saw that. So there's some good stuff here.

Perfect, yeah, so what we'll do is I'll show you one of the prompts for this workflow here, and then we'll transition to Margaret, and I'll use the prompt writing prompt that everyone will get access to. Maria will send the link to get that. And then we'll build that prompt real quick for you, Margaret. We'll build the first version of it. Real quick, this self-learning loop is a little bit more advanced. This is… this box and this arrow are doing a lot of heavy lifting. It's pretty hard, it's…Not hard, but…as a human, we're, like, we can catch best practices pretty easy, like, we're smart, you know? AI, sometimes struggles with that, and you can get pretty crappy best practices, so I usually say, like, this is where the human lives a lot, is to, like, make sure our best practices that are getting added are actually good best practices. And there's a couple of ways to manage it. It could literally be a document, which is what I do for this. You can also do it as a database thing, it depends on, like, the agent and what you're trying to do. But building this out isn't, like…

It isn't super easy, so we're not going to talk about that too much. Real quick, I'll go over a prompt here. So, this is the weekly, the weekly strategy report, so every week I also get a report. how the videos have performed over the past 14 days, and a recommendation on what videos to do next. So, here's the objective portion, and we're not gonna, like, read this word for word, but I want, just to show you, like, how I think through writing these prompts. Here's the objective. I'm your content strategy agent for Instagram Reel Creator. You receive one table containing all the data for 14 days. I'm explaining what it does. And I'm explaining the job that I want it to do. Okay? I use… I do this sometimes too, which is, like, a core, like, almost like a, a centering question of, like, hey, what video should the creator film today based on the last 70 days of… or 14 days of performance?

and were any posts that should have been re-recorded but appear to have been missed. Like, this is, at the end of the day, what I want out of my weekly report. If I wanted a ton of stuff, I would not use a central question. I would list out the stuff that I wanted. Okay, I talk a bit about source of truth here. This is more of, like, the rules section, how to interpret the metrics, so Instagram waits, and they tell you how they weight each one of these metrics for the algorithm, so I make sure I'm clear about what that is, and then I give really clear, important rules. I don't know why there's a cross-through on this, but whatever. The input is one table, do not expect separate inputs, like, these are just rules for it to follow, okay? Then I give instructions. Step 1, Step 2, identify the patterns. Step three, build a new topic opportunity list. Kind of two things, like, hey, you need to re-record these, you haven't yet, and then here are some new topics that you should, cover.

You can see, like, best practices context when it refers patterns in the table, topic suggestions from in the diagnostic reports. You can see I'm, like, telling you kind of where to, like, base this on. Then there's step four, detect these missed re-record opportunities, like, hey, this one seemed like it was pretty good and you missed it. Step 5, write the actual performance overview. Step 6, build new video topics to form… to film. Step 7, build missed re-record opportunities. Step 8, build key takeaways. And then, boom, here's the output that I'm looking for, and this is how I want it structured. Four to seven sentences on how the thing performed.

new videos to film, topic one, topic two, and you can see, like, this is a pretty long report that I get. I get it once a week, that's why. And I just want to see, how are we performing, what things should I do this week? Now, I get one every day, too, that kind of keeps me very, close to what seemed to have worked the other day, so I can kind of fuel that more. And then I get this once a week, which can also… that gives me more of a plan of what I should focus on, moving forward. User input, this is where you put in the… the input, the context it needs in order to do that job, okay? So we're going to talk about that with Margaret's thing right now. So this is our prompt writing prompt, okay? If the link hasn't been shared, you can share the link here. And the input, you're going to write a description of what the prompt you should receive. So let's go to ChatGPT here.

You can see I just copied and pasted the prompt in. I'm gonna replace this section with Margaret's stuff. So, let me go back to her original, this is what I wanted. I wish I had a better prompt to find out if it's the right timing. Okay. The goal of this prompt is to figure out if it's the right time to reach out to a specific target account. To try and book a sales meeting. I'll leave that for now. And then I use, Super Whisperer to do talk text, by the way. Just so you know, my brain break there, I was talking what I wanted. I always forget about that. Here are the main signals that I look for when I'm trying to determine if I should reach out to a target account now or not. And I'm gonna copy, and I'm past… and you can see, like, this is sloppy. I'm literally just copying and pasting…Why did that not copy and paste in?

Oh yeah, copying from Zoom chat is some…

Does it work? Oh.

It'll work, and then it doesn't.

break. I might just have to speak it out. Pace, let's see. Got it, okay.

I don'.

manually click copy and whatever. Did I forget? Okay, I offer, yeah, yeah. My offering is around data analysis strategy and direct marketing services. The input of this prompt should be the website of the target account. Okay, so that is all you do. You can see, like, pretty sloppy, I just spoke it out. I pressed send, and now… you guys know this part, it's gonna start thinking Okay, so it's going to create a prompt following my prompt writing prompt, which is in a structure that I already know and love. So here we go.

Objective, and what I'm gonna do now is we're gonna go through this. I won't read the whole thing out loud, but I'll read kind of, like, some stuff that I think is important, but I'm also gonna highlight areas where we could run into trouble, so you can see through, like, see how I go about getting to a first draft, because I treat this as, like, a first draft, then my human eyes go through it, I make some tweaks, and then Margaret, you're going to take this prompt, I'm gonna give it to you. you're gonna take this prompt, and you're gonna use it, and when it's wrong, you're gonna go, huh, why was it wrong? And you're gonna go back to the prompt, and you're going to make some changes to it. And I'll show you where you would do that. Okay, so the objective, determine whether now is a strong time to reach out to a specific target account and attempt to book a sales meeting. Okay, important. This is, where you'd probably make some changes. either remove stuff or add stuff. If

You don't like the outputs. This is part of refining a prompt. So keep the following details in mind. To this account be contacted now?

I do like the ChatGPT lets you select and add ChatGPT around certain things as well. Yeah. That's pretty helpful.

Great. Okay, I think that's fine. Now are the instructions. These are the step-by-step instructions that you would take, right? That you would follow. Identify the organization type. And these might not be right for you. They might be right for you, right? I think it's probably from the list you sent, so they are, but this is where you might make some tweaks. Evaluate offering fit, So here's an example. I would absolutely change this before running it. Offering fit. Analyze whether the organizations appear to likely benefit from this. What is likely bet… likely to benefit? Like, under what context? Right? Everybody, you ask 10 people, hey, would this website, would this company benefit from data analysis strategy? You're gonna get 10 different answers.

well, actually, it would just be yes or no, but you're gonna get different answers. It'll be like 50-50, right? Like, I don't know, probably, or this is so broad, I don't… yeah, I guess, like, doesn't everybody just kind of benefit from this kind of type of stuff? Right? So this is where I would spend a lot of time on this, Margaret. And what I would do is I would say, like, what does benefit mean? Like. How do I know it better… they would benefit from one of these things? And then let's look at this, so, like, right here, it's trying to do some of this. Look for evidence such as donation pages, annual giving programs, campaigns. I would spend a lot of time… step two would be a very important step to me. And I would put in some of the things that you do to decide if they're likely to benefit. It might be signals, like maybe they post something specific, maybe they're doing an action, maybe they have a specific role in the company, right? I would spend more time on Step 2. I think step two is a little weak here.

Step 3. And if you have questions, hopefully that was clear. If there's questions, type in questions, I'm happy to elaborate more. Identify timing signals, so looking for timing indicators that suggest hour range might be relevant now. Okay, so this is good. This is helping. Half of Step 2… well, that's of the stuff I said about Step 2, I think, got moved into this section, but step two still, my thing applies. What is likely to benefit from this stuff? Like, this stuff is all probably useful stuff for those people. It needs guidance on what that means. Okay, this is good. Assess urgency. I'm not sure I'm gonna like this, just to be clear. Deciding whether to kind of starters…Yeah, there's a couple of ways you can do it. I think, actually, this is okay. If this… if you think this is useful for you, keep it. If not, delete it. I think…

like, since you're doing a one, one thing, I think this will go okay, but basically it's like, hey, is there… are there multiple signals from steps 2 and 3? that means this is more of a higher urgency thing, but if you're putting the lead in there, maybe it's just high urgency. I have this one, agent that I built that, does account research, and part of what it does is determine, does it fit our ICP, and then are there signals to reach out? And it might be, like, they raised money recently, they just hired a new person in this kind of role, like, that kind of stuff. So I built out this whole agent. And this part of it, this urgency rating, we don't do with the LLM. We do this in the spread… it kind of all goes to a spreadsheet type thing. And, it just does… it does math. It just goes, okay, meet ICP, yes, and then based on how many signals it has. Does it have good signals? Does it meet ICP? That's low, medium, high. That's how you hit the urgency. So you can do it without the LLM, but since this is one-to-one, Margaret, I think you can keep this. Assess sales relevance, determine whether…

Okay, that's… that's good. make recommendation, This is good. Identify missing information. I wouldn't do step 8, I would cut Step 8. If you have examples, you can put it here at first drafts, I don't have examples usually. And then here's how it's gonna present the output, and you can play around with that. In interest of time, I'm gonna not go any more in detail there. Okay, so that's, like, how I would go through. So I just told you, like, quickly, like, the tweaks that I would make. The real tweaks, like, heavy stuff. is gonna come from using it, and as I see what worked, what didn't work, I'm gonna make tweaks to that prompt, and then, Margaret, once this prompt gets to a place where it works really well for you, you build this out as a workflow. And then that workflow is, is gonna be one that you, you build in into, like, your CRM, or you use some kind of automation tool to trigger that to take place. That's how, how I would do that. And I'm gonna make this a little doc, and I'll share it with you, Margaret, so you have it.

No, that was awesome, man. We put the prompt writing prompt in there for you guys to grab. I think you can see how much easier it makes to just, you know, come up with a strong prompt right there from day one. And also, like, I'm sure a lot of people now are familiar with creating GPTs or projects in Claude. Another just, like, easy starting point, too, is taking that prompt, loading it into a GPT or a project in Claude. And you can just kind of keep running that over and over instead of having to repaste the prompt in. But we've got 2 minutes left. There was just one question that kept getting asked in different, various forms, Kyle, that I wanted to do last. And I feel like we answer this in every single show. A lot of people would ask Claude, ChatGPT, Perplexity, which one should I use, and why do you prefer one over the other? Do you want to give your quick overview on which LLM to use and why, or…

Yeah, I'll tell you what I use, I'll just break down my personal life. So if I'm doing quick searches, like if Jed and I are hanging out, you know, at a park for some reason, and he's just like, oh, something, and I'm like, oh, let me look it up. I'll use Grok, and it's because Grok has very good recent information because of X, and I don't always ask a question that needs that recency, but when I do, it gets it, so I use that as kind of, like, my daily Q&A thing. For all my work, I use, Codex by ChatGPT. It's their desktop application. It does really good coding, but also does compute… I almost do… my day is almost all prompting Codex to go do shit for me, stuff for me. So that's, the other one. I, I think…if you're using Claude, Claude's really good as well. I think they make the best models,

So, you can stay there. Perplexity I used for a long time, Perplexity Computer, I've now moved back to Codex. I stay pretty flexible, I don't think you do. If you're using, ChatGPT or Claude, I think you're in pretty safe, safe, territory there. And then you don't need to switch.

Love it. Well, good stuff, guys. Again, check out the previous AI Foundation shows. We'll see you all on the next one. Check out, we just dropped Kyle's Instagram. If you guys are there, check out his videos. Help him get to 4 or 5 million next time. Dude, thanks, man. I appreciate it. Always a good show, really tactical stuff, so thanks again.

Cool, yep, good chatting, everybody. We'll talk soon. Bye.