The Hidden Cost of Bad AI Prompts (And How to Fix Them)

Most AI teams aren’t losing to the model.

They’re losing to bad prompts.

In this episode of the AI Proving Ground Podcast, WWT’s Liz Gattra breaks down the invisible tax of vague instructions, blind trust in outputs and endless iteration loops that quietly burn tokens, waste GPU cycles and drag down ROI.

As generative AI moves into production, AI literacy becomes operational leverage — not a soft skill.

We cover:
 • A practical prompt blueprint that improves output quality fast
 • Why “confident but wrong” is more expensive than you think
 • How precision reduces token spend and compute waste
 • When to switch between ChatGPT, Gemini and Claude
 • Why system prompts, agents and context engineering determine whether AI scales or spirals

Every vague prompt compounds.

The teams that win don’t just deploy models — they train people to think clearly, structure intent and treat prompts like lightweight programs with measurable results.

If you care about AI ROI, compute efficiency, and scaling generative AI in the enterprise, start here.

Precision isn’t optional. It’s a cost control strategy.

Support for this episode provided by: ExtraHop

More about this week's guest:

Liz Gattra is Director of Product Ownership at World Wide Technology’s Application Services practice. A former Agile Coach, IT manager, Scrum Master, and Agile Business Analyst, she has led enterprise agile transformations and built product ownership capabilities across internal and client teams. Liz develops and trains product owners to deliver measurable business value at scale. She holds a Master’s in Biology and Biomedical Sciences from Washington University in St. Louis and brings nearly 20 years of cross-industry IT experience.

Liz's top pick: The AI Adoption Gap Isn't Technical. It's Human.

The AI Proving Ground Podcast leverages the deep AI technical and business expertise from within World Wide Technology's one-of-a-kind AI Proving Ground, which provides unrivaled access to the world's leading AI technologies. This unique lab environment accelerates your ability to learn about, test, train and implement AI solutions. 

Learn more about WWT's AI Proving Ground.

The AI Proving Ground is a composable lab environment that features the latest high-performance infrastructure and reference architectures from the world's leading AI companies, such as NVIDIA, Cisco, Dell, F5, AMD, Intel and others.

Developed within our Advanced Technology Center (ATC), this one-of-a-kind lab environment empowers IT teams to evaluate and test AI infrastructure, software and solutions for efficacy, scalability and flexibility — all under one roof. The AI Proving Ground provides visibility into data flows across the entire development pipeline, enabling more informed decision-making while safeguarding production environments. 

Transcript

The Hidden Cost Of Bad Prompts

SPEAKER_01 0:00

Are bad prompts an invisible tax on your AI spend? If your teams aren't getting the basics right when it comes to prompting, your AI investment is quietly bleeding money. Every vague request, every blind trust and confident answer, every endless iteration, that's wasted tokens, wasted GPU cycles, wasted time. AI literacy is no longer a nice-to-have skill, it's the difference between scalable advantage and invisible cost. So in today's episode, we're talking with WWT's Liz Gottra about the real job behind AI literacy, helping enterprises get measurable value from AI instead of churn. Prompt engineering sits at the center of that shift, and Liz is one of WWT's prompt engineering leaders who's trained teams across the entire organization. She's gonna break down what leaders are actually hiring AI to do, where most organizations derail, and how weak prompting quietly turns into operational drag. We'll cover three through lines: the hidden cost of port prompts, the shift from just English to disciplined instruction, and how agents change the role of the human from user to orchestrator. From Worldwide Technology, this is the AI Proving Ground

AI Literacy Isn’t Optional

SPEAKER_01 1:04

Podcast. Today's conversation starts with the foundation, what AI literacy really is, and why it's becoming a board-level concern. So let's jump in. Well, Liz, I've been hearing more about AI literacy more and more by the day. Seems to be something that comes up a lot now about you know how that is adopted is driving AI adoption enterprise-wide. Can you level set for me on what AI literacy means, why it's important, and what value does it provide?

SPEAKER_02 1:37

Sure. So AI literacy is really about understanding not just what AI is, but how to use it and more importantly, how to utilize it. So there's a lot of things that we can do giving people tools, but if we don't teach them how to utilize those tools, there's a limit to what they can do. So the more people learn about AI, the more they understand it, the more they realize what they can do with it and what they need to do in order to get what they need out of it. And that really is AI literacy, which leads to AI fluency, which is that ability to easily interact with it.

SPEAKER_01 2:10

Absolutely.

Prompting Is the Foundation

SPEAKER_01 2:11

And among other hats that you wear here at WWT, you are among the several few that are kind of our prompt engineering experts. You've helped train, you know, broad teams that we have here on how to prompt effectively. As it relates to AI literacy, where does prompt engineering fall? Is it kind of foundational to AI literacy, or is it just kind of, you know, one of the classes that you take to get to literacy?

SPEAKER_02 2:34

I would say that it's foundational. If if you don't understand how to prompt, best case, you're gonna get something out of there that's sort of what you need. Otherwise, you may just get frustrated and quit. Or worst case, you're gonna get something out of there that isn't true and you're gonna pass it on to somebody. And you know, that never ends up well.

SPEAKER_01 2:54

Yeah. I can think back to times where I've not had a good experience when I'm prompting an AI tool, whether it's Atom AI, you know, our own proprietary tool that we have here at WWT or ChatGPT or Gemini.

Tokens, GPUs, and Burn Rate

SPEAKER_01 3:07

Normally it's I look back at the prompt and I'm like, oh, that was just not a not a good prompt. I'm wondering, you know, I can see how that's frustrating from an end user perspective, but what other what other uh detriments does bad prompting bring to an organization? Is there some type of value there that we're not getting?

SPEAKER_02 3:23

Yeah, there's definitely two sides to it. So you've got the one side for the user themselves. Yeah. So you've got the it doesn't give you what you need, right? But there is another side to that. So somebody who really is trying to make it work. So they prompt and they iterate and they iterate and they iterate. And A, it's a waste of their time because they're spending so much time trying to get it right. So you've got that cost, but it's also a waste of other resources. So when you think about it, if you are using one of the public systems out in the cloud, you are probably paying by token. And a simple explanation of token is basically by word. Okay. So every time they prompt, that's more words going in there. And so you're wasting these tokens and that's costing you money. Now, if you have this built internally, if you're running in your own data center, it can get even broader and even more expensive because now you're

From Magic to Discipline

SPEAKER_02 4:17

looking at really overloading the AIQs for those GPUs. So it looks like you need more GPUs for the load that you have when really you're not getting any value out of that load. You've got the power that those GPUs are consuming. You have the cooling infrastructure that's being utilized by them in there. So not only is that costing money, that's also costing wear and tear on those cooling systems. So now you've got the maintenance on that as well. So it just kind of snowballs from I write a bad prompt, and so I keep trying and trying and trying to all those other second and third order consequences.

SPEAKER_01 4:52

Yeah. And I love that. When we talked about that a couple of weeks ago, that was not an angle that I had considered that not only is it causing me as the user doing a bad prompt anxiety or frustration with the tool, it's costing my company money and the company. You extrapolate that over, you know, thousands of employees, it adds up pretty quick. But that's probably not a that's not a cost that is seen in like a line item. You're just assuming that people are having good prompts. So it's an invisible cost. So really that you to to tackle that, you need to it invest in prompt training and getting people up to speed. In other words, AI literacy, right?

SPEAKER_02 5:27

Absolutely. You know, it's yes, with good prompting will come more prompting and more use. But the difference is is now that use is actually adding value instead of just churn.

SPEAKER_01 5:41

Yeah, yeah. When Jen and I exploded onto the scene and, you know, everybody's kind of playing around with Chat GPT, it all kind of felt like a magic spell, so to speak, when we're doing these prompts. We've learned a lot more about what good prompting is. What have we learned about prompting over the last uh few years?

SPEAKER_02 5:59

I I would say that the biggest thing that people have learned is AI LLMs, these models, are not mind readers.

The Prompt Blueprint

SPEAKER_02 6:08

So it really comes down to how well you prompt, how well you take what's in your head and put it down in words so that that AI can understand what it is you want. Okay. So what is the true outcome you're looking for? And are there nuances about that it needs to know about? And why are you doing this? Okay, that will help it understand. But also it wants to understand what is that background? Like, what is that information? How do you set the stage for it? So, in some ways, you almost tell it a story before you actually ask it for what you want it to do, so that it is better prepared to give you something that's actually useful.

SPEAKER_01 6:47

Yeah. And do you think organizations are paying attention to this type of training? Um, are most organizations, whether it's you know, companies that that you interact with or just hear about from you know from others within the company, are most of our clients thinking this way, like where we have to actually train these people with with prompt engineering?

SPEAKER_02 7:05

Honestly, I don't think they are. I think, and I think it's, you know, we've seen this with SaaS tools in the past, where a lot of times they won't do adequate training because the tools are sold as being something that's easy for them to learn, right? And I think it's even worse with AI because when you think about AI, you communicate through the human English language. And it's easy for them to say, well, it's just English. Everyone knows English. But that is the problem because we think it's just English, and it's really not. It is a type of programming language, and I want to be careful there because it's not quite the same, but it's an easy way to say it's a way of talking to the AI, much like we did with programming languages. It just happens to be in a native language, but there are still rules you need to learn, there are still gotchas you have to learn, there are still nuances you have to learn. You know, it's much like with a coding language where you have to learn the different things, the different libraries, the different words that you use to mean different things. Same thing applies, but because it's in natural language, those details, that nuance, that skill gets obfuscated because we think, well, if you can speak English, surely you can just talk to the AI and tell you what you want.

SPEAKER_01 8:17

Yeah. And not only speaking English, but you know, what you get back is very conversational. So you're tricked, maybe lulled into

ChatGPT vs. Gemini vs. Claude

SPEAKER_01 8:23

a sense of, oh, I could just say, what do you mean? And it's just gonna know what it what do I mean. You mentioned some rules and and gotchas, you maybe articulate some of those. What are some of those kind of basic foundational rules that we need to consider to make sure that we're making efficient use of our prompting?

SPEAKER_02 8:37

Right. So one of the examples I like to give is the fact that AI doesn't actually have common sense. And if you think about common sense as that idea of you learn things just growing up and being a human. The example I often say is everyone knows that wet floors are slippery. Why? Because every single one of us has slipped on a wet floor before. Right. Now you take an AI, it's never gonna slip on a wet floor. So if by chance that information was not covered in its training information, or you don't give it that information in your prompt, it could very happily tell you to just go run full speed across that wet floor because it doesn't realize it. It doesn't have those nuances. And I know this is a simple example, but you can extrapolate that out to other things, things that you just know because you've worked at a company or you've done something. The other example I give is a lot of time when you write a recipe, right? You forget, you leave steps out because you don't even think about it.

SPEAKER_03 9:29

Right.

SPEAKER_02 9:30

And you we assume that humans will fill in those gaps or ask us. AI is very literal in some of those things, it will try to execute those things. And so it will miss steps that we would catch that were missing.

SPEAKER_01 9:51

Yeah. And is that why it's important then to kind of adhere or understand some of the frameworks that we've seen, whether it's coming out from you know, you and the trainings that you provided to us here at WWT, or you know, there's lots of trainings out there on the internet where it's context, act as a persona, things like that. What else is included in some of those frameworks?

SPEAKER_02 10:08

Right. So we have the prompt blueprint that we use here, which has task, context, constraints, persona, output format, tone and style, and examples.

SPEAKER_01 10:18

Yeah, kudos on memorizing that.

SPEAKER_02 10:20

Yeah, you know, you'd think I'd done this once or twice. But the idea here is it's not so much that you use every single one of those every single time in every prompt you do. The idea here is that you understand the value of each of those and how adding them or modifying them actually impacts the AI and its response to you. So we've actually started to build even beyond there to the idea

Where Models Get It Wrong

SPEAKER_02 10:46

of it's not just about telling it what to do, but are you using the right what? So if you think of the what is the task, the action verb, right? We have some that we always kind of default to create, analyze, summarize. Those are probably three of the most common ones. But when you start to think about all the nuances of synonyms related to those words and how you could be even more precise in defining it. So is it that you really want it to write something, or is it that you want it to generate something? Those are different. Is it something that you want summarized, or do you really want it to paraphrase and bring out specific details? Those little nuances in the words that you choose change the response because at the end of the day, it's really just fancy math. The AI is drawing associations based on mathematical distances between words. And those words, even though they sound similar, they are not the same distance from other words. So you will get different responses.

SPEAKER_01 11:44

Yeah. One of the things that I recall from a prior episode that we've done here at the AI Proving Ground podcast, we had Jordan Wilson, who's got his own podcast. He came on, and one of the tips that he gave that I've been adhering to for a while now is reading the chain of thought when uh you're going and working with uh an AI model. I I'm curious, have you ever done that where you look at the chain of thought? I mean, it really helps you understand, like, okay, here's how this AI digested my prompt and it kind of surfaces areas that you didn't do well at.

SPEAKER_02 12:12

Yeah, I absolutely do that. And I encourage them. That's a great piece of advice that I myself do. And I will often catch something that I'm like, and I'll ask it, why did you do X?

SPEAKER_03 12:23

Yeah.

SPEAKER_02 12:23

Because A, I want to understand what its reasoning was. B, I want to understand if there was something in my prompt that may have caused it to do that when I wasn't expecting it. But it's good to understand how those connections it's making. And so, yeah, that's a big one, definitely.

SPEAKER_01 12:40

Yeah. I'm curious, you know, earlier just in this conversation, you talked about how a lot of these tools are marketing themselves as easy to use. And, you know, they are easy to use once you understand how to do them, but that's just one tool. There's hundreds of tools out there. And, you know, one of the interesting things that I know I've heard you talk about is how you can't treat every model the same. How does, you know, how if you're thinking about kind of moving from Chat GPT, let's say, to Gemini or to any other, you know, AI model that someone might use, how do we need to think about tailoring

Meta Prompts & Reusable Personas

SPEAKER_01 13:12

our prompt engineering for those different tools?

SPEAKER_02 13:15

Yeah, it is it is a nuance there, definitely. Now, something like the prompt blueprint, we've structured it so that it has a lot of flexibility, but you definitely can fine-tune it when you choose a model. I like to think about AI as more similar to the way we think of streaming subscriptions.

SPEAKER_03 13:31

Okay.

SPEAKER_02 13:31

So we're all gonna have one that we go to. I love Netflix. That's kind of my default go-to when I want to find something to watch. But we will have specialized ones that we use for certain things. So my husband likes horror, so we have Shudder.

SPEAKER_05 13:44

Yeah.

SPEAKER_02 13:44

We like documentaries, so we have the B or the PBS, right? So you will have specialized tools that we use, one main one and then specialized ones that you go to. And it's the same thing with these AI tools that we have. So for me, it's Chat GTP. However, if I'm looking at coding, if I'm looking at something related to software engineering or building agents, I'm probably in Claude doing that because Claude is known to be really good at that. If I'm looking at something about building out a really nice image or infographic, I'm probably in Gemini or in Nano Banana or Notebook LM because we know that that Google has that really good model for doing that. So I'm choose picking and choosing where I'm going and I often hop between them and use them interchangeably. So you can't think about it as one, but you're right. You have to understand the nuances about how they also act and how they work. So for instance, Gemini, I believe, has a 2,000 or 200,000 token limit to its context window, the amount of information you can send to it.

SPEAKER_05 14:46

Yeah.

SPEAKER_02 14:46

Whereas Claude has a much larger one. So there are things that you can do in Claude because of that much larger context window that would be more difficult to do in Gemini. And so knowing those little nuances and knowing which one is best for which is really part of the power of using AI.

SPEAKER_01 15:04

Yeah. Bringing back a little bit to prompt engineering, though, here, I'm sure there's several of these basics, fundamentals of prompt engineering that go across all of these models.

From Assistants to Agents

SPEAKER_01 15:14

But is there nuance in how you prompt ChatGPT versus how you prompt Gemini versus how you prompt Claude? Or is it all just like a good prompt is going to be a good prompt that cuts across it all?

SPEAKER_02 15:24

So there are nuances. Okay. And people who use all three of those or the other ones tend to start to shift how they talk to that specific model. If you are someone who predominantly uses one, you tend, you will learn how to speak to that one model and then you'll still speak to the others in that same tone, which usually works. But if you really want to get good, what you'll understand is those nuances. So some of them may weigh different parts of the prompt blueprint more or less than others, as far as how much it takes it into consideration. In others, they may look more towards the beginning or the end of the prompt. We find that there's a lot lost in the middle in all of them. One of the latest things that came out was this idea that actually doubling your prompt. So what I mean by that is duplicating it. So putting it in and then putting it in again in the same prompt window and then sending it off helps because as it's going through, it reads things and learns things as it goes. And it's weird. That last bit of context that you put in for it doesn't actually know about that first bit of context.

SPEAKER_03 16:32

Okay.

SPEAKER_02 16:32

So by doubling it up, now that last bit of context gets to see that first bit of context again.

SPEAKER_05 16:37

Right.

SPEAKER_02 16:37

It's a strange way of thinking about it, but it's another way of helping to reinforce what you're asking that AI to do.

SPEAKER_01 16:44

Yeah, that's an interesting tip. I'll have to put that one into action pretty quickly here. What about? I mean, but these models are always getting smarter, right? So is does is that to say our prompting is gonna get easier or more lazy, so to speak, or are we always gonna have to be pretty sharp on this skill set?

SPEAKER_02 17:00

I I think that it's a nuance. And it's hard to say about whether they're getting smarter as they're getting better at understanding what we are asking for. Okay. However, you actually we're finding in a lot of cases you still have to be very specific. And then sometimes you want to be more specific because they get more literal. And so it is really thinking about how you break up and think about translating what's in your head into what's there and what it understands. The best example for that is if you think about when you're trying to get one of these models to generate an image for you. All right. Maybe you're you're thinking about it and you say, okay, I want you to create an elephant standing on a bull ball with a cat riding it.

SPEAKER_05 17:48

Okay.

SPEAKER_02 17:49

Okay. The problem is, is as you say those words, generally you start getting some image in your mind. And so when the AI creates that image, it suddenly doesn't match what's in your head and you're frustrated. And then you start going. So really it's like make the image and then really start getting detail. Like, what color is the cat? Is the cat wearing a hat? Right? Yeah. What color is the ball? Like you start getting into those details so that it matches what is in your mind. And it's a good way to train yourself to start thinking about the details

Confident. But Wrong.

SPEAKER_02 18:23

because you can see it immediately in an end result with the image. But it's the same type of thing when you start thinking about prompting for any other task you want it to do. What are those details it needs to know in order to actually execute and give you a response that you would be valuable to you?

SPEAKER_00 18:42

This episode is supported by Extra Hop. Extra Hop provides network detection and response solutions to monitor and secure enterprises in real time. Gain deep visibility into your network with Extra Hop's advanced analytics platform.

SPEAKER_01 18:55

Yeah, it's funny though, because I feel like my prompting ability kind of ebbs and flows. Like I'll be in a state of lazy, so to speak, and then I'll I'll start to adhere to more of these prompt engineering basics, some of which I've I've learned from you. And then I get success with AI, and then I kind of go back to a little bit of lazy because I'm just assuming that AI is going to be able to account for what I need it to do. How do you how do you stay sharp with always kind of adhering to all of these rules and all of these you know nuances that prompt engineering brings about? How do you keep them in mind and how do you stay disciplined to to be good at it? I cheat. Okay. How do you cheat? Everybody likes to cheat.

SPEAKER_02 19:36

There's two ways that I cheat. One is I I really believe in meta prompting, which is using AI to help you build your prompts.

SPEAKER_03 19:42

Yeah.

SPEAKER_02 19:43

And so as I'm building the prompts, I know through the blueprint the things that I should be looking for. And if I don't see them, I will ask. Do we need to add this? Well, maybe we should do this because I know those things are important. So if they're not there, I can ask. So it's like a collaboration between me and the AI, building the prompt.

Turning Prompts Into Assets

SPEAKER_02 20:00

And then I'll take whatever prompt we build, stick it into the thread that I want to use it in, and off we go. So that's one way I cheat. The other way I cheat is by creating reusable parts. So, example, I have a number of personas that I have saved off into memory. And so when I want to use those personas, I don't have to recreate all of the details about how those personas function and what perspective they have. All I have to do is tell the A that I wanted to use that persona. So it saves me time. There's another thing related to skills. This is big in Claude, but you can actually emulate it if you're careful in Gemini and even Copilot if you're careful, where you can kind of set up skills. And now you can actually have skills that you reference that then reference a persona and you can reference different personas to use the same skill, which again gets you a slightly different output, but you're treating it more like Lego blocks a little bit.

SPEAKER_01 20:59

Right.

SPEAKER_02 20:59

And so you've got these little cheats that you can use.

SPEAKER_01 21:01

Yeah. That's an interesting concept. So you're talking, I mean, basically, then a single employee, take me, for instance, could build a team of skills or a P a team of assistants to just be kind of on standby and at the ready to help me out when and if I need them. I may not be using my whole roster all the time, but I'm always using somebody.

SPEAKER_02 21:20

Yeah, I've been using Chat GTP's project functionality to build persona teams for months now. So I have teams of different personas that do a variety of things. So

Context Is the Moat

SPEAKER_02 21:32

any I have one that helps me with my diet that happens to be a nutritionist, a meal planner, a culinary coach. I mean, and they help me come up with recipes that are tasty and healthy, right? I have another one that I use for retirement investing. I have one that I actually put together because I do a lot of external presentations. So I have an entire AI presentation review team where I can upload a presentation and have it give me feedback on that presentation so I can. And improve it from different perspectives. So there's a lot of different things you can do with persona teams alone. And then you throw the skills in and it it just expands. And that's even before we start talking about building out agents.

SPEAKER_01 22:12

Yeah, well, that was my next question because we're starting to get a little bit into agent territory. How does that change the name of the game with prompt engineering? Is it is it is it more of an orchestrator vibe or language that we need as opposed to that kind of literal, literal task with persona with that and that, or you know, how does that change it with agents?

SPEAKER_02 22:31

So there's two, there's two parts or there's two types of prompts that we have to talk about when we talk about agents. So there is the system prompt or the system instructions that are called, which really defines the agent, yeah, uh what it does, what it can do, how it behaves, what tools it has access to, how it responds, its personality, all of that stuff is in that system prompt. And then there you still have a user prompt with from the user who's interacting with it. And when you are the user interacting with a true AI agent, you have to start thinking about what do I need to give it and what don't I need to give it because it's already inherent in that agent. So for instance, if the agent already has a persona, you don't want to give a persona in your user prompt to that agent. Okay.

SPEAKER_01 23:13

That's even if it's the same one.

SPEAKER_02 23:15

Even if it's the same one, let it let it do what it's going to do. All right. Now you might want to give it a few more pieces of context, like tell it, I want you to focus on this industry. All right. Or do it in assuming that you've been working in this industry for several years, something that might augment that particular persona, but you're not changing the persona. You're just giving a little more context to think about. Same thing applies for things like output formats. There may be ones that are workflow agents. You don't want to change the output format, probably in a workflow agent because it needs to do what it does. So you don't want to change those things. There's it again, each agent would be different. And hopefully you understand enough about the agent through either the description that the builder has written or the example prompts that they've given so that you understand, okay, these are the types of things you can just let go.

SPEAKER_05 24:04

Yeah.

SPEAKER_02 24:05

Which again is kind of another way of cheating because then you don't have to come up with a persona or you don't have to come up with an output format because it already has that for you.

SPEAKER_01 24:14

So that's system prompt. What's you said there was two?

SPEAKER_02 24:17

So no, so there's the two sides. So there's the user prompt where you where you don't, you leave that out. Right. So you still prompt with it, but you don't have to add those extra things because those are already inherent in the system prompt.

SPEAKER_01 24:29

Yeah. No, that makes sense. So when you are dealing though with agents, you are going to have more of an orchestrator role. So what types of different mindset do we have to bring to prompting when we're orchestrating agents when you have many that you have to consider?

SPEAKER_02 24:45

So now you're talking more about an agentic system where you have agents working together. And so this becomes more like a team with a head in it, but this team can actually act and do things. Unlike my persona teams, which are really just reasoning teams to help me think through things, these are actual teams that can execute. And so at that point, you generally have an orchestrator agent, which is the agent that the user interacts with. And then that agent has access to a number of other agents, specialty agents, to do certain things, perhaps tools, perhaps databases, where it can determine based on its own reasoning and the request you have for it, which of those agents, which of those tools, which of those databases it needs to pull information from and which ones it needs to interact with to actually execute and then give you the response you were looking for or do the thing you wanted it to do. And so you're building out a number of things. So you're building the orchestrator agent and it'll have rules about what it can do and interaction rules. But then each of those subagents, you will have rules about how it communicates, what it can do, but also a definition of what it is so that the AI orchestrator agent can figure out which is the right tool to use or which is the right agent to use at the right time. So it gets, it starts to really you go from just thinking about one little thing to really thinking about an entire system and how that system interacts and what you need to define so that it interacts efficiently.

SPEAKER_01 26:18

Yeah. I want to get back a little bit to AI literacy because I'm gonna ask you about human in the loop here as it as it relates to agents. I mean, ideally, teams are developing and building these agents and testing them and testing them and continuously testing them so that they have a strong level of trust. But we know that's probably not always gonna be the case. How do we how can we start to shift people's mindset from like that Googling mindset of like I'm looking for an answer, and just any answer almost does, to where I'm looking for like the kernel of truth here. I'm looking for the truth of the answer. Like, how does that bake itself into to prompting?

SPEAKER_02 26:54

Tell me a little bit more about what you mean by the truth of the answer.

SPEAKER_01 26:57

Yeah, like sometimes when you're when you're prompting, you you you can fall into a habit of just believing what it says, you know. So it could hallucinate, it could be slightly off, it could give you a wrong number or metric or whatever it might be. So training a workforce to not just look for any answer, but the right answer or an at least somewhat accurate answer.

SPEAKER_02 27:17

Yeah, this is always a tricky one because we as humans tend to fall back on some shortcuts that we've evolved into doing just because they're easier. You know, one that we always talk about is the trust fallacy, this idea of somebody says something confidently to you and in a way that makes sense, you just naturally want to believe what they're saying as being true. Right. Even though it can be completely wrong. And so this is something that you can definitely train. I'll be honest though, the biggest training is experience. So the more you experiment and the more you find you get hit with gotchas, suddenly you start to realize that all that training and all that additional things that I tend to teach, like I have a 201 course where I teach advanced prompting. And one of the things I teach them is how to utilize different personas or different models to actually challenge the response that came first so that you can definitely look at it from different perspectives. Ones that I use is I call it my evidence curator. Again, it's a safe persona I have. Whenever I get stuff out that I really need to make sure is true, I will just tell my evidence curator to take a look at it, compare it to the sources, tell me what might be misleading or might just be outright false. Yeah. And then it'll give me an output of a comparison. So I'm checking it. But I'll be honest, most people it takes getting hit a couple of times, like all of a sudden passing something on that wasn't right. And then you start to realize, oh no, I really do need to be careful. And then after a while, especially with a safe persona like I have, it just becomes second nature.

SPEAKER_01 29:05

Another issue along those lines, and I don't know if this is personal to me or if anybody else out there has this same issue, but I I get frustrated with how positive and how yes man-y AI can get, right? Like I'll be, you know, a lot of times I use AI to validate ideas or validate are the questions I'm about ready to ask you, you know, good or not. And it's always like, yes, absolutely, these are right up their alley. Is that something that you train out of an AI, or is that just something that you have to have a mental note of?

SPEAKER_02 29:34

I actually know what you're talking about. It's kind of funny. It's almost like you read my AI newsroom post that I sent out this week because it was talking about exactly that.

SPEAKER_01 29:42

Yeah.

SPEAKER_02 29:42

You know, at first when I got it, it was like, oh, that's really cool. And then it's like, okay, every single time you're gonna tell me I'm right.

SPEAKER_01 29:47

And I mean, I'm confident, but uh yeah.

SPEAKER_02 29:49

It just became annoying.

SPEAKER_01 29:51

Yeah.

SPEAKER_02 29:52

And so what I did was I worked again with AI to craft a an instruction that basically tells it that I don't want it to be that way. What I want is I do want to know when there's agreement, when when I am on the right track, but I also want to be challenged. It's more important that I am challenged that I am told that I am right. And so that little set of instructions I put into my custom instructions. So in all of the major ones, you have the capability of personalizing the AI. So you can make it be pretty much whatever you want it to be. Okay. So you do that through the custom instructions or the personalization. So Copilot has it, ChatGTP has it, Gemini, Claude, all of those that I use have it. And once you do that, that just gets sucked in to the context window and sent over to the LLM, just like everything else. And so it takes that into account as it is responding to you. So it's not so much like you're training it because it's not like you ever train an LLM. It's just not the way it works. Once they're trained, they're kind of done right now, although I know they're working on changing that. But what it is is it's a standing instruction that always gets fed over.

SPEAKER_01 31:08

Yeah. How can recognizing that you know a lot of our listeners out there manage teams or manage teams of teams? How can how can managers in the workforce know whether or not their employee base are doing this prompting well and not to your point at the very top of this episode, just wasting time and cycles and tokens and wasting money? Are there signals that we can look for that we we can tell, oh, that person's a good prompter?

SPEAKER_02 31:36

So, I mean, obviously there is. Some of it comes in the output that they're giving you, but it does require you to check. I mean, check some of those sources. I I don't even know how many times I've had 404 pages pop up when I've checked sources that have come back. So that's the simplest way is to check, right? Or confirm what they say. But also you can check like how they think about it. So, you know, do they feel like there's a high level of trust in the AI and probe that a little bit. Why? You know, you can you can tell the way they respond to questions about how do you verify that information, you can tell whether or not they are truly going out there and validating and verifying.

SPEAKER_01 32:17

Yeah. How often do you collaborate with others on on prompts? It's it's funny. Somebody asked me the other day, oh, I'd love to see that prompt. And I it was like a very personal question. I was like, oh my gosh, you're gonna see like my train of thought here. So do you do you work with others on prompts? Is that a good habit to get into? Or how do you think that's a good thing?

SPEAKER_02 32:34

Oh, I share my prompts all the time and I love to see other people's prompts because sometimes you never know where you're gonna learn. And there is so much that we don't know about these. You know, you think when they roll out one of these new LLMs, one of these new models, that the person who built it, that trained it knows how it works. The truth of the matter is they don't. They know a little bit about how it works and how to work with it and what's the best way to prompt it and utilize it. But most of that is actually discovered by the average person who's using it and shares that information with other people. So I learn a lot just by reading people's blogs, listening to YouTube videos about how people are using it, and talking to other people and seeing how they're using it.

SPEAKER_01 33:18

Yeah. That's interesting. So it's almost like, you know, you hear the stat of, you know, you only use a certain percentage of your brain, like we're only using a certain percentage of the potential of any of these LLMs. I mean, are we like drastically underperforming what these LLMs are capable of? I I think the answer is gonna be yes.

SPEAKER_02 33:34

On any individual, yes. I I think we're getting closer when we combine what it is we know. But when you think about the breadth of what it is and the breadth of the knowledge that that's out there and being generated every day about how these are, no, most of us are just touching the surface of what we can do with these models.

SPEAKER_01 33:52

Yeah, yeah. What um I know we're coming up at the bottom of the episode. What what's coming down the line for for prompt engineering? This this episode, we're recording at the beginning of 2026. We'll probably release it in the next several weeks. So, what what do we have in store over the balance of this year, or even if you can, you know, crystal ball it out past 2026 and to beyond?

SPEAKER_02 34:11

I I really like what I'm seeing with what's happening with the concept. So we started out with prompt engineering.

SPEAKER_05 34:17

Yeah.

SPEAKER_02 34:17

And then we we went from there to this idea of context engineering, which is how we actually segment and present all of that context and information to the models to make sure that we're only giving them relevant information. And then people have now started to talk about content management. How do we manage the content that we have? Because we really need that prompting models, really. If you don't have the data, if you don't have the information and it's not in an organized way that the AI can actually access, it doesn't matter without that data and that information. So that's definitely coming. And I think people are starting to realize that even more. And then on the other side of it is this new idea with I really like what Anthropic has come out with with skills and taking the concept of what I do with personas and then taking skills, and you can use those depending on the skills, those can be used with agents or they can be just used with AI assistants the way I do with projects. And now you have this idea of reusable parts of prompts that you can share with other people. You can move between these different models. Now you might have to tweak them a little bit because every model is a little bit different, but you start to get this repeatability that I think is something that will help amplify what we're already doing.

SPEAKER_01 35:41

Love it. Well, you got this pin on right here that says passion on it. You're definitely, you can tell you're passionate about prompting and AI. And thank you for all the training that you provided the company. Those are definitely helpful. If anybody out there from worldwide is listening here and hasn't taken advantage of those courses, by all means should go in and do it.

Build the Skill or Pay the Cost

SPEAKER_01 35:57

I know we have a couple of those that are also publicly available on WWT.com as well.

SPEAKER_02 36:01

We and and we actually offer all of these trainings to our clients now. So I can come out and train up your teams as well.

SPEAKER_01 36:08

Perfect. Well, it won't be long before we have it on again because I know like like everything with AI, things are changing uh at a constant clip. So things will change. We'll need your expert advice again. But for now, thanks for thanks for joining us today.

SPEAKER_02 36:20

Anytime, thank you.

SPEAKER_01 36:21

Okay, thanks to Liz for joining today's episode. AI adoption doesn't fail because the models aren't powerful enough. More often, it fails because the workforce isn't prepared to use them precisely. If you want real return on AI, start with literacy. Train the humans, sharpen the prompts. The technology will meet you there. This episode of the AI Proven Ground Podcast was co-produced by Nas Baker and Kara Kuhn. Our audio and video engineer is John Knoblok. My name is Brian Phelps. Thanks for listening, and we'll see you next time.