So What? How to Refresh Your Generative AI for 2026 📱

Well, hey everyone, happy Thursday. Welcome back to So What the Marketing Analytics and Insights Live Show, our first show of 2026. We are here. Hey guys. Happy New Year.

Thanks. Yes. Happy New Year. We got Jazz Hands from John. Jazz hands for the new year.

That's my revolution. Um so we wanted to kick off the new year with some housekeeping because it is, you know, a good time to reset and refresh and just sort of see what you've been working on, what you're carrying into the new year, what you're ditching. So unsurprisingly, let's start with generative AI. So uh on today's episode, we are going to talk about how to refresh your generative AI for 2026, which sounds like a broad topic, but I'm fairly certain what we want to do is sort of narrow it down and try to like walk through where to start, what things you should be putting on a regular maintenance schedule, um, you know, when you need to think about if the model changes, do you need to change your instructions or can you just leave it? That kind of stuff.

So Chris, where are we starting today? We're gonna start with three things. The three things are things around prompts, things on environments, and things around models. Those are the three things that we want to talk about today. Um prompts, some ways to to tidy up around the house, if you will.

Uh on that front. Environments, some of the the newer generation of tools that you probably want to at least consider installing if you're allowed to do so. Please check with your IT department if you have one of those. And uh models, uh, which is the models. And a lot of this stuff deals with, like you said, it's 2026.

Things have changed. And one of the things that I know I've seen in a lot of other places is people are kind of stuck in honestly, 2024, um, in terms of of how they're using generative AI. So one of the um let's I guess let's start with prompts and a couple things that I wanted to remind folks about. Um, one, in your prompting techniques these days, one of the things that's proven most effective is instead of a giant long list of everything to do and not do, uh, is to establish what are called first principles, which is a way of saying here are the the three most important things. So let's say we were trying to, I don't know, uh refresh our sales handbook uh or something like that.

What would be the and let me ask you you both this what would be like the top three things that you would want AI to know as as cardinal rules about the way Trust Insights evaluates, you know, uh a potential uh sales prospect or how we do business and what we don't do for business. I mean, I guess uh I don't know if this is the answer you're looking for, but I guess this goes to knowledge blocks. Um, and so like the number one cardinal rule, I guess, would be in order to appropriately update the sales playbook, you have to refer to our about us knowledge block, which also would need to make sure that it's updated. Um, and then our ICP, uh, our ideal customer profile knowledge block, which would we would also need to make sure is updated. But I think it would have to be like you can only refer to those two knowledge blocks in terms of how we conduct business and who we conduct business for.

You can't look outside of those uh data sets for lack of a better term, to figure out what trust insights does. So that's you know, it's the onus is on us to make sure we provide that context. But that would want to, that's a guardrail that I would want to set right away is this is who we are, this is what we do, this is who we do it for. If you start going outside of that, then you've broken the rule. And I think that list you just made this is who we are, this is what we do, this is who we do it for, those are your sales first principles, right?

You are that those those lunatic, if you don't have one of those three things, you can't do the sale, right? Um, and you certainly shouldn't be using AI to do it. John, when you're looking at how you would give instructions, maybe to a new sales team member, what would be the the three rules that you would say that this is it? If you don't do this, we we've lost it. Yeah, well, you've got to, you know, first screen, you've got to find out if they've got normal band, you know, do they have budget authority need timeline?

And then answering any questions, you know, doing some exploration to figure out if there's actually, a deal there, um, you know, figure out what they can actually work on. Um, and then all the process and procedure as far as you know, getting them a scope and you know what goes on in the sales process and where everything goes. Um, I think that between that and all the past data deals that we've done, you know, information on everything else that's come before, uh, that's pretty much the whole process. Okay. So if you distilled that down to a rule or a set of rules, what would you tell?

Well, let's say this year uh we have a new person who joins our sales team named Bob. Um what are the three rules that you tell Bob? Like, don't this is what you must do every day, Bob. These are your first principles. Yeah, clean out the inbound queue, you know, review the existing deals and then explore for new business.

I guess those would be the three, you know, process jobs. Exactly. So what you've both done is defined a set of first principles. And when you're building things with generative AI, um that's that's where you that's how you lead, right? So when you're prompting these things, if you're saying, uh, you know, we need to build an ideal customer profile, those first those those principles should be part of the ICP construction process, because there's no point in having an ideal customer profile if it isn't aligned with how you want to sell things, right?

It's it's kind of like if you were to just set off a research agent, say, go find us customers that are of this size of this budget of this industry, but you didn't say like these are the rules, it's going to bring back stuff that's maybe not on target. But say, hey, you know what? Uh, people who sell uh bio warfare weapons are your ideal customer because they meet the budget though, and we're all like no. So when I think about when you said when you know, when you say people are stuck in 2024, which you know, last week was fine because it was only a year ago, and now it's two years ago because that's how math works. Um, you know, when we're talking about prompting, you know, for a lot of us, we were taught you start with you are a super experienced, you know, ultra, you know, high performing salesperson, you know, and so you sort of set that context first, that expectation.

Is that no longer something that we have to do in 2026 with uh prompting with generative AI? Do you still have to do that or no? You can, and it still does work. Um, but the models have changed so much to uh in terms of their ability to do reasoning that in many cases they can sometimes start to infer things. So you would still want to say, um, you know, perhaps you don't say you are this, but you're saying today we are working on building ideal customer profiles for trust insights.

Here's who we are, here's what we do, here's we do it, who we do it for, and you give it those those first sets of three principles. Now let's go and start building out what these things are. And here's the the next twist on the ideal customer profile. It's good to have it be like this is who this person is, and and and you know, needs pain points, goals, motivations, etc. All that stuff.

You want to do a second version of it. And the second version is you want to turn it into a checklist. You want to turn it into a checklist of when you are for any given task, this is these are the criteria to check off against this task. Does this match our profiles needs? Does this match our pain points?

And you might have that in the checklist a list, uh, a more a granular list of these are the pain points, does it meet this yes or no? Because what's turning out to be true across the four different model families for reasoning models is that telling them check your work or be thorough or whatever is not as impactful as here's the list, go down the list and check your work against this list. So where does that come from? Because that feels like a new thing. Um, you know, because again, I'm sort of coming from the less technical lens of we're doing our best with prompting, and this is how we were taught to do it, and we were taught to be like uh check your work or you know, go back and you know validate what you've done.

So it sounds like the structures moved from uh validate that to checklist. When did that happen and how do we make that happen? Uh last year. So as reasoning models have evolved, uh, and there's three different families of reasoning models. There's the formal names of Monte Carlo Tree Search, uh, PRMs, and backtrack, but from uh the at a at a non-technical level, the models are getting much better at figuring out guardrails for a task, but they still struggle to do it accurately.

So when you say think this through, you know, are you you remember back in 2023, you would say, tell a model, think this through step by step? That is too vague. But if you give a model, and this was this came out in a paper in June of last year. If you tell a model, here's the checklist. Verify your work step by step off of this checklist.

It goes, okay, customer needs. They uh they need uh good analytics software, they need good data. Does does this customer have this? Yes or no? Check, check, check, check, check.

It's kind of the same way, Katie, that you would delegate to a junior employee. You wouldn't say, Oh, go check your work, Chris. You'd say, Chris, these are the five things I'm looking for. And of these five things, you did two of them. Which feels like common sense.

No, no, no. And I and I mean that in all seriousness. It feels like common sense because that is how I operate. But I think again, where a lot of people are still trying to figure out how do I get the best result out of generative AI. You mentioned it came out in a paper in June.

I'm gonna go on a limb and say majority of people aren't looking to the academic papers to figure out what to do. They're looking to companies like ours, you know, uh experts like you, Chris. So to be fair, this is not the first time, but this week is the first time I'm hearing about the checklist, but you're saying it's been around for six months. You know, we can sort of in a different episode get into how can your average users stay up to date on the best practices. But it's I want to acknowledge that this feels like new information, even though it it's also common sense.

It is it is newer, yes. Uh, it is newer. And uh it was really brought to light by um the folks at Alibaba who make the Quinn family of models. Um, they won some awards in December, uh actually late November, early December, uh, at the NERUPS conference, specifically for how they've trained their own models to recognize checklists. So it is it is newer to the family.

But this is why this is why we have our live stream, so that we can go through this stuff. Well, and I want to acknowledge that, you know, the way John that you spoke about, here are the three things that you would tell, you know, who the new sales associate is very much in that checklist. So it sounds like that very much aligns with what you're describing, Chris. So, you know, so John, you're saying, you know, number one, basically check the backlog, make sure you know everything's getting responded to. Number two, check, you know, the pipeline, make sure all of that's squared.

And then number three, you know, go searching for the new business. Like that's very much the checklist. Yep, exactly. So that's part one of four in prompting um that of your generative AI refresh for this year. Second, make sure you you do regular maintenance on your prompt library.

Uh, Katie, you mentioned this at the top of the show. Um, every so often you should, or anytime your workhorse model changes, um, you should be revisiting it with uh optimization for that. So, for example, in December, uh, we're a Google shop, we've been a Google Workspace customer for a while. So, Gemini is our workhorse model. It's the one that we use, the one that we pay um as a company to use.

When Gemini 3 came out, both uh Pro and Flash, it was a big change. The model's behaviors change. So our gems, and in fact, we ran into this in late November. Uh, our uh our team member Kelsey was saying that suddenly all of her gems just start going completely off the rails. That's because the underlying model totally changed.

And so that's you know, as part of your 2026 refresh, if you have not put your prompts through the optimizer for your system, this is the time to do it. And there's um every every company that makes AI tools has some form of prompt optimizer. So, for example, if you are a chat GPT user, um, in chat GPT, there's the uh OpenAI platform, and in here in the platform, they have a prompt optimizer that will allow you to optimize specifically for uh Chat GPT. If you are on Claude, Claude has a generate prompt feature in the Claude Developer platform. You put in your current Claude prompt and it will fix it and improve it.

Google puts their prompt optimizer in the most obtuse way possible, uh, which is inside of a vertex notebook that you have to run inside Google Cloud to optimize. But it is also the most thorough way to optimize prompts because it actually does internal testing of its own prompts to see which prompt delivers the best result. It's a pain in the butt to use. Um, and you know, for for your casual non-technical user, it's just as easy to say, hey, what are the five things you would do to improve this prompt? And what are the three questions you have for me about my intent for this prompt?

Ask me and then rewrite this prompt. But at a high level, if your workhorse model changes, it's time to clean house and revise all those prompts. So, John, when you do, well, first of all, I don't know actually how often uh you're creating prompts for generative AI, but when you do, are you asking it to check its work and revise it, or are you using more checklist style? You know, as long as I get the output that seems to make sense, like I've I'm fine. I don't spend a lot of time massaging it.

You know, a lot of the stuff that I like to use it for is take existing work and just say, hey, what was missed in here? You know, and it usually comes back with great stuff. But yeah, I haven't reached it, it's funny. I haven't reached anything with prompts where going back and using old gems come through as broken, but I'm also the user that, you know, I'll fire up an app and it'll say, hey, this was deprecated two years ago. You need to go get a new tool.

So, you know. So perhaps you're not the one we should be asking. Well, I'm the worst case scenario user, that's for sure. When we're talking about UX, but and so that's why, you know, I was psyched to hear about today, because like what should I do? So I've got my stack of, you know, 45 different prompts that I randomly roll out for something.

I mean, should I just go run all of them and see how they work? Or is it more wait till I I hear an alarm bell that like, okay, oh, everything changed this week, it could be broken. Like what what would be the right way to handle that? In general, um when uh when uh uh the your workhorse model has a big number change, like Gemini 2.5 to Gemini 3, that's when it's time to do some QA to to test the prompts to see if they still work. Um, because one thing that, for example, with Gemini 3 in particular, it is very sensitive to conflicting instructions.

If you give it um a prompt that has conflicting instructions, it just goes just goes crazy. Um, and and and just can't keep up. Um so when a model changes, that's when it's time to take you know your top five or 10 prompts that you use the most. Test them and see like, am I still getting a good result? I was planning on uh, regardless of this episode, I was planning on uh doing that with my system instructions for one of my most used gems.

I'm assuming the same kind of maintenance is needed for those as opposed to just prompts because it's essentially the same thing. Exactly. So that's techniques, library optimizers. The last thing is to start thinking about this. Although it is something that folks who are more savvy are starting to start thinking about what skills your AI should have.

Uh and a skill we covered this last year in our show on Claude Skills, which was uh announced uh last year. Since that came out, Anthropic has made public the skill uh protocol to all AI makers. And so other companies can pick this up and use it in their software for free. And so just like with MCP, uh which uh Anthropic made public what in 2024, uh, and other companies have started picking up using that skill format is something that is uh going to become popular. Now, here's how to know that you're using a skill that you need AI to learn.

You keep retyping the same instruction over and over again, right? So you say, oh, by the way, go through this document and remove all instances of passive voice and converter to act to voice. If you find yourself typing that over and over and over again all the time, in most systems this year, you're good, and and in Claude now in particular, you can say, Hey, Claude, let's build a skill that converts passive voice to active voice. And then as you're prompting you to say, Claude, use the passive voice skill. And you won't have to type in that paragraph long and it will know what to do and do it.

But more importantly, as you start using agents within that system, the agents will know the skill exists and will do it autonomously. So you might be saying, I did this recently for fun. Uh let's write a trashy romance novel, and then you would say, Hey, editor agent, make sure that you use the anti-passive voice skill, and it will automatically go through and do that. You won't have to spell it out anymore. So, in the same way we talked about checklists, starting the show, skills are ways to embed the checklist into a repeatable process so you don't have to copy and paste it every single time.

It's like gems or GPTs on steroids. So I know everything you just said, and I was struggling to follow because you just said a lot. So let me just sort of back up a little bit. Um, as you mentioned, we did cover skills in a previous episode of the live stream. You can go to trustinsights.ai/slash YouTube, go to the sew up playlist, and you can find that um skills live stream.

Basically, what you're saying is, so let's say, you know, to your point, if you always add in the instruction, clean up passive voice, or clean up this transcript so that it reads with this tone, or apply my writing style. Those should be skills within your um generative AI system. Basically, it starts to take it like quote unquote automates things a little bit more. Um, you know, so it's not something that you necessarily have to be um, hi bud. You necessarily have to be, you know, developing things.

This is for any user, like a John Wall who, you know, may not have opened the system for two years. Uh, but when he does, those instructions are already written. He doesn't have to remember, wait, how was I correcting for passive voice? I want to do it in a consistent way. Uh okay, great.

Now I have the skill all set. So this is this is that we're talking. So you started talking about agents and everything, and it's, I just want to make sure that we're not losing where we're going with this. Correct. Okay.

In John's case, think about those first principles. If you use, you know, Bant as a framework, you know, budget authority need time frame, and you have a checklist that you've spelled out, you might, if you're using Claude, create a Bant skill. And then when you've connected Claude to your HubSpot, you'd say, hey, this new deal, this new opportunity came in. Claude, run Bant. And it will go and know what that is, what that means, pull the record from HubSpot and run it through the criteria and say, Hey, this is how this scores on a you know uh zero to 10 on budget authority need and time frame, and push that and tell you that, okay, like, yeah, this is not worth the time.

As we saw at the end of 2025 with Google Workspace Studio, when we did that in the live stream, you're gonna see more and more of these capabilities become autonomous, which means that if you get a skill well thought out today, when it becomes available in your ecosystem either as a feature in the chat or as part of an agent, you won't have to reinvent the wheel. When let's say uh Gemini uh workspace studio connects to HubSpot and can support skills. John might have uh an agent that knows the Bant skill, and every deal, every opportunity that comes in the HubSpot, it automatically applies the Bant skill. Well, I love that. Kelsey, put that on the backlog.

So those, so in prompting, those are the four things that we should be doing as to refresh our generative AI for 2026. And are we, and so we're talking about maintenance. Is this something that we should be doing at a minimum quarterly? Ideally, yes. Ideally.

Um, at least just checking in um to see are have there been major shifts and major changes. For example, um Nano Banana Pro came out last year. Nano Banana Flash will be coming out in the next few weeks. And so if you're doing image generation, you'll want to know uh when that model changes because suddenly your images may look very different, or they may be faster or cheaper, or something along those lines. So just keep something to keep in mind.

And I recommend that folks join the Trust Insights newsletter, uh, stay in the Slack community and stuff because we share all this information in those places as we see it happen. People are, I know, probably tired of hearing from me about it, but um, but these are the things. So the next pillar for refreshing your AI is the environment that you use it in. And there's three places that four places that you use AI. Number one is the obvious one.

Go to ChatGPT. I'm in ChatGPT. That's the environment everybody knows, right? That's, as we said, 2024. Nothing wrong with it.

Still useful, still important. But there's three more places that you'll want to be thinking about this year and maybe doing. The first is command line tools. So these are things that exist in your terminal or in your command line. And there's four of them that what that correspond to the different systems.

There is clawed code, if you're an anthropic customer and you're on you're on a paid plan for them. There is OpenAI Codex, which is from the makers of ChatGPT, and it is it allows you to use the GPT models there. There's Quen code, if you're in the Alibaba Quen system, which is uh a very, very powerful system, and of course the Gemini uh coding app. They all four of these if you install them on your computer, the big advantage they have over regular web-based chat is they can access local files with your permission, and they can write local files with your permission. So you if you're sitting there, if you're saying to your your co-CEO, hey co-CEO, keep a diary, help me keep a diary, and I'm gonna tell you what happened every day at the company, and it'll write it'll take notes and have a conversation and write things down in little files in a folder on your computer, and then you can say, Oh, co-CEO, you know what?

Two months ago, we were talking about uh this idea. Did we ever do anything with that? And it can search through those files on your computer and say, yes, Katie, we did talk about that. We didn't do anything with it. Which is an interesting use case, but I'm I'm a little confused.

So do I have to be coding in terminal to make that happen? You have to use terminal. It's you do not have to code. You do have to write a single line of code. You chat with it like you chat with chat GPT.

Okay. I mean, so that's that's interesting. And I feel like that's something, you know, we could probably do a whole episode on that because when I think terminal, I'm like, oh crap, I gotta write code. All right, let me either, you know, go ask Chris to do it or find it a different way that doesn't involve talking to Terminal because it always gives me an error. Right.

Um, which by the way, your screen is very tiny and intimidating. Um so I just turned on Claude code, and you can see that it's it's got its uh cute little icon there. All these, once they're installed, you either say you put them in it, you figure out what folder you want to work in. You say Claude or Codex or Quen or Gemini, and you know, and you get the the lovely uh little eight-bit greetings. There's Quen, there's Gemini, uh, and everybody loves their little decor.

And then codex. Yes, that's fine. Um, codex, not nearly as attractive. This is still Claude. So you could say something along the lines of, I want to start keeping a diary.

Um, can you ask me questions about my day and write a date stamped file for each day's entries? Something simple like that. No coding, no, nothing complicated. And Claude would be like, sure, I can help you do that. And then you come back each day in this folder and say, hey, let me tell you about my day.

And just hey, I'd be happy. How was your day overall? Uh and some specifics. What was the highlight of your day? Do you think frustrating or challenging happen?

What'd you accomplish today? Is there anything on your mind right now? I assume you can say and ask me these specific questions instead. Yes, exactly. Um, but because it has access to your the file system that you put it in inside whatever folder, you can have it write things down.

And that is the big big secret to these tools that makes them better than a web-based tool. Is you could say, let's write this down, let's take some notes. Let's uh and then you can have a go back and review those notes and review things. You can also, Claude in particular supports this really well. Remember how we were talking about skills earlier.

You might have a say, like a not just a co-CEO, you might have a uh a set of skills for board members. You might have the skeptical board member, the you know, cry the you know, the the crass board member, the the optimistic board member, and you might say, I want to talk about my Q1 plans. Uh, let's have all three board members weigh in, and they it invokes this three separate skills, like a miniature role play, and you have a conversation with them in the in the chat. But I would have to do that through terminal. Yes, you'd have to do that in here.

Um, because you'd want to again have then say, Claude, let's take write down the minutes from this board meeting. And you would have your your virtual board meeting. Claude would write down the notes, and then in you have enough, you would have your your folder, here's the minutes from the January board meeting, the January virtual board meeting that Katie, co-CEO Katie, uh, and the three you know board members all had together and you could read it like real board minutes. And I didn't even attend it. They just all talk to each other.

Exactly. Now remind me, because it's been a hot second. So obviously, I know you're on a Mac, so you're on an iOS operating system. Can you also do this on a Windows machine? You sure can.

Uh in Windows, it's called the command prompt. In Mac, it's called Terminal. Gotcha. I just wanted to make sure, because I know we talk about terminal a lot, but that's because we all use Macs. I didn't want to exclude the Windows users because there are a lot.

Yes. Um it's called command prompt. Uh and you can in Windows, if you go to the start menu and type search as start typing the word in the search box, you can just start typing the word command and it will help you find it and get started right away. Gotcha. And um configuration of this though, is this using an API to call their web servers, or have you had to install the models locally?

No, this uses their online models. You have to have now this is a that's an important clarification, John. You have to be a paying customer. These are not available for the free users for the most part. Like there's like a little time to be useful, you have to be on what the at least the $20 a month plan.

Gotcha. That's very helpful to know. Yeah. Second thing, if you just you're like, you know what? No, I refuse to use terminal.

Sorry, not gonna happen. There's two additional environments you could install that are full-fledged applications that would allow you to do this. Uh one is called anti-gravity by Google. Um, it is a it is their desktop uh AI environment. And anti-gravity uh functions very, very similarly, but it's a it instead of it being um a terminal, it's just an app.

So again, you could say, yeah, I'm gonna have this folder and I want to start a diary, and you would turn on anti-gravity's agent manager and have those conversations to say like I I want to do this, and you could choose you know uh a few different models, uh, and you could say, like, I want to have a conversation about my day. Let's take some notes, or I want to do a simulated board meeting. Now, this does not have built-in skills, but it does have the the agent that can help you do stuff. And is since it's Googles, does it only work with Gemini? It works with Gemini, Claude, and uh the OpenAI GPT OSS for the clawed ones, you need a paying subscription for this.

Okay. Um, and the Gemini ones also. So that's one potential uh way of doing it. If you I just like you know, I absolutely positively refuse to use um a terminal. The second is an app called VS Codium or Visual Studio Code, uh, and an extension in it called Kilo Code that again ties to the different models.

You can choose uh gazillions of different models in this one and do very similar things. So when if you wanted to say edit the trust insights newsletter, um you could have it do that in this environment. But those environment, those two environments are are if if you're like, you know, I just can't do the terminal thing. I refuse to, you know, I refuse to look like I'm it's 1980. I want to I want something a little more modern.

I'm gonna be honest, these don't look more modern. But it explains a lot about why you create the UIs that you do. Um exactly. Yeah, no, I mean, I I understand the importance of it, but I'm looking at these and they all today, you know, ask me again next week, but today it all feels very overwhelming and intimidating. And so this is gonna be a tough question.

Maybe I'll ask uh John. Um, John, what do you think is the barrier to entry for using terminal or for using these standalone apps that sit outside? Like, do you look at this and go, Oh yeah, I can totally do that, or do you feel like, nah, that's way too technical. I still can't do that. Yeah, well, I mean, I don't I don't see there's no need that would require me to do that.

I mean, I do see the value of if you're if you've got scripts or you're doing other stuff, now you've unlocked that as a tool that you can run wild with. But yeah, I don't there's nothing uh, you know, I'm usually not having a script fire and have it run something every week for me and put the output in a folder or anything like that. Um, you know, so you know, it's good to know it's there, but yeah, I don't know. I don't have anything off the top of my head that I would need it for. So again, John is the worst case scenario of of end users.

If we think about the ideal customer profile, you might have um, you know, an ICP skill in one of these environments, and then anything you're working on, you would say, check this ICP check this out, or use the ICP skill on this. So you're working on a uh Q2 promotion, you're working on an event, you're working on a pitch uh to a sponsor for marketing over coffee. Um, if the ICP is built, then you would create the skill for it and say, ICP, what do you think? Right. And so it's always it's one of those things where it's a shortcut on tap.

Now you don't have to do that. You can still copy and paste out of your prompt library, but having it built into the into your ecosystem means it's faster for you to get to there. Gotcha. Yeah, because I either copy it out of our prompt library or I just you know attach the physical PDF of the ICP and say, make sure it aligns with this criteria, which doesn't feel uh cumbersome to me. And again, this is something that you and I talk about uh effective.

I think we talked about on the podcast not too long ago. A lot of the more um infrastructure-heavy uses of AI like terminals, stuff like that, are not mandatory, they are faster, they are more efficient, but they are not mandatory. You can still get the same result. In fact, we did talk about this on uh this week's podcast. You can get the same result by doing it manually, you know, copy-paste at a time.

Uh, what these tools allow you to do is to scale that much faster. So, for example, if I had an entire quarter's worth of promotions, I could use the ICP skill inside Claude and have it spawn an agent to look and review each one. And so instead of taking six hours to copy and paste each one individual one at a time, 22 agents could launch all at once and use the same ICP skill and be done in 10 minutes. Provided you know what you're doing, because for someone like me, you're saying be done in 10 minutes, but it's gonna take me probably a whole week to get that uh working correctly versus the six hours of copying and pasting. Or we yeah, we just go to the local Duncan with your laptop and set it up.

I mean, you we say that, and then the second I go home. And I think that this is really the clarification is so you really need to think about how you need to be using these things. So, you know, obviously you have you know the three of us, and we use it all basically, we kind of cover the gamut of the way people use generative AI. Chris, you're the power technical user. I'm sort of the middle of the road, I could kind of go either way.

And then John, you aren't using it as often as other people. It's just not what you need to be doing right now. It's not a good use of your time. And so when we're talking about these different solutions, when Chris is saying more efficient, it's more efficient for someone who's doing work that he's doing. And so I think when people are watching these, you know, live streams or listening to us talk, those are clarifications that we need to be making because we were talking about this in our uh Slack community uh this week, which you can join for free at trustinsights.ai slash analytics for marketers.

People get caught up in the AI hype and the engine that kind of keeps like you have to do this and you you're falling behind, you're doing this. You really have to look at yourself as an individual and say, what is it that I need to be doing? And that is also one of the things that we want to challenge you to do in 2026 is figure out what works for you, not do what works for Chris because it works for Chris. Do what works for you. So I'm likely gonna fall somewhere in between Chris and John.

I will probably take a look at the terminal stuff because it makes sense to me. I get it, but then I'll also use the systems natively and work just within Gemini because I can see use cases for both. Yep, exactly. The last thing that I think should be on everyone's list for this year as a you should at least have a plan. You don't have to do it, but you should have a plan, is to have some form of non big tech provider as your backup.

So this means this this is what we when you talk about local models. John, you and I did a live stream last year on small uh language models. Um local models are ones that you run on your hardware, and this is in tools like anything LLM or LM Studio or what have you. When you look at generative AI today and the economics of it, it is not sustainable. OpenAI is lighting like five billion dollars a month on fire.

Um and just you know getting more and more money from investors and things like that. Of all the big tech companies, you know, when you you look at something like uh Google, they yeah, Google can backstop the funding on Gen AI for a while, but even they can't afford to just burn money indefinitely. There is a real non-zero chance that one of the big players does not make it out of 2026 intact, right? They either get acquired or or whatnot. And so the technology isn't going anywhere, but a one or more providers might.

So it is a good idea to do some scenario planning and to say, what would it look like if we had to use local models? Because we've gotten used to using AI in our work, and a provider of choice went boom. Being reliant on technology is always a risk. Um, but it's not a new concept to have backup. So one of the things that we've talked about in numerous shows is the sort of Google Analytics and Motomo conversation.

Motomo being the you know free version, not as robust, but to have it running in the background in case Google Analytics implodes, which is it's done before. Um, you know, uh your CRM, making sure you're backing up your data to, you know, a local database or something cloud hosted, where if Hubspot or Salesforce or whatever goes away tomorrow, uh, you haven't lost all of your contact data that you've been accruing over the years. That's not a new concept. And so I absolutely think that you're right, Chris. When we think about what's going on in the generative AI space, it is very volatile.

I mean, how quickly it's changing is you know, it's giving people whiplash. And at some point, that's going to, they're gonna hit a wall. Someone's gonna hit that wall first. We just don't know who. Right.

We don't know who, but it's it's probable. I mean, yeah, you can't light five billion dollars a month on fire for definitely just it's just not one of those things. Um, the the tools uh the stack that I recommend if you have the equipment to support it, big asterisks there, because Katie is Katie and the company have been very generous in funding the equipment we're allowed to have. Um, uh, is LM Studio with one of Alibaba's Quinn models? And I'll give you an example of how you would tie this together today.

Um, if you use the model by itself, it's a small model. It is like eight made of mostly made of hallucinations, so if I say as an example, who is the CEO of Trust Insights, right? It's going to, it says Mia Whitman. Just completely invented it because it has no grounding. But if you have done your homework and you set up things like the ability for it to do web search, is you can now say who is the CEO of Trust Insights.

It will go, oh, I don't have confidence in my answer. Let me go do a web search. It's Katie Robert. And here's the K the LinkedIn page and and her and the Trust Insights page. So even in the case of a relatively small model that natively hallucinates, because today's local models now understand, oh, there's this thing called the web and I can search it.

They can be very reliable now. And that is a big change from even just a year ago when models were not trained to do that. And yeah, we're hallucinating all over the place. I think um the other thing I would add to this how to refresh generative AI for 2026 list is do an audit of all of the custom GPTs and artifacts and gems that you've built. Because like us, you've probably built a few dozen that at the time felt like a great idea.

And now you're like, oh, I haven't touched that. You know, I've become John Wall and it's collected dust over two years. Uh sorry, John, you're you're great. We love it. I'm not denying it.

You're not the you're not the end user, which is which is totally fine. Um, but I think that you know, we get caught up when we try this thing and then we never touch it again, sunset it, delete it, archive it, whatever makes sense. Grab all the data from it, the system instructions, put it in a doc, and then remove anything you're not using, especially in a larger organization where you have these accounts that are shared. You want to make sure that people are only using the GPTs and gems that you want them using. So if you have stuff that's out of date that you're like, oh, that was like five iterations ago.

Why is it still there? It's like any kind of auditing. You still have to do the auditing with these you know tools. So just build it into your plan once a quarter. Everybody sit down in the conference room, as Chris likes to say, over a beverage of your choice, grab some pizza, and then someone screen share.

Okay, here's all the custom GPTs we've built. Which ones stay, which ones are people using, you know, send out an informal survey. Doesn't matter how you do it, you gotta do it. Yep. Uh we got a question from Brian, our friend Brian, uh asking about uh as a non-developer doing AI coding, what things do we need to think about for hoping to create apps for clients to use?

Uh, or are you mostly building stuff for internal use? We uh so I say 95% of what we do is for internal use. Um, or it's stuff that we operate on behalf of a client. We could do an entire week of content, uh like 12 live streams in a row on coding practices with AI, because there are far more ways for it to go wrong than right. Um, so maybe that's a a good topic for a future show.

I mean, the short version is that at the end of the day, it is software development. And software development is not, I don't want to offend anybody, but it's not just like a skill that you can pick up and do tomorrow. There's a reason people, you know, spend a lot of time learning it. There's a reason there's a whole team around, you know, any individual developer because you also then have a manager and an architect and a database architect, and you know, whatever it is that you have testers. Um, I guess I could in I could rant about software development for an entire year because I did it for a very long time, and you know, Chris has forced me to still do it.

Um so my advice, unsolicited advice, is build things internally, don't put them out for public use because then you start to get into the support, the maintenance, the bug fixing, the feedback, the well, I used it and it didn't work, I want a refund. Like that's a headache, and that's a whole different infrastructure if you're not ready for it. All right, I'm gonna settle down now because like I really could go off on this. There's a lot of good things about software development, but there's also a lot of risk. Yeah.

Like I said, I think that is worthy of at least a podcast episode, if not a live stream. Because even with AI, there are a number of very interesting things that that tend to go wrong right off the bat. And there are practices for mitigating them. But to Katie's point, because it is a profession, there is still a baseline amount of domain knowledge you need to have about the profession of software development to do it well. Yeah.

I think uh what we're seeing, and again, we'll go into this in a different episode. What we're seeing is it is very easy to stand up an application using Claude Code or the Gemin the Google developer tools. It's very easy. These apps that people are standing up are so fragile, they're so riddled with privacy holes. There's very little data integrity, and quite frankly, there's a lot of problems with data storage because it's not something that people tend to think about.

They're like, oh, I want this app to do this. Well, where does the data go? What if somebody needs it later? Again, whole rant that we could go on. Get up the bug.

Oh my goodness. Get out the coffee mug. Oh my god. Oh, yeah. I gotta get my uh the ranty pants.

We have the official ranty pants. And I also got sent the ranty pants keychain. We gotta get the Shopify store up and running. Oh my goodness. All right, so yeah.

Prompts, environments, and models, the three pillars of the things that you need to refresh. And that is above and beyond all the other data governance you need to do, all the uh maintenance, all the refreshing of data itself to make sure the data it's all based on is good. But those are the things that I would put on your 2026 spring cleaning list to get done sooner rather than later. Yeah, and if you're in New England, don't wait for spring, it ain't coming anytime soon. Yes, as we look at the uh layers of ice out the window here.

Um any final thoughts, John? You know, I it was just interesting. The this whole idea of yeah, you can use these tools to write code and prop up something that will run once, but the hardening of software is a whole different thing, a whole nother level of stuff. And I'd be interested in a future stream of you know what's happening out there, you know, as far as automated testing and bug fixes and requirements gathering, all that kind of stuff. Like it it's just uh having a lot more people running with scissors is not just a better world.

If you want to speak with me. If if you want a an example of just how bad the problem is, just do a Google search for the phrase open AI underscore key and you know the restrictions of the site github.com, and you will see so many people putting their private credentials in public that you can then you know, if you were an unethical person, go use on their dime because they did not properly follow even the most basic procedures. Excellent. I have the final words. Exactly.

That's gonna do it for this week's show. Thanks for tuning in, and we will catch you all on the next one. Thanks for watching today. Be sure to subscribe to our show wherever you're watching it. For more resources and to learn more, check out the Trust Insights Podcast at TrustInsights.ai slash TI podcast.

At our weekly email newsletter at TrustInsights.ai slash newsletter. Got questions about what you saw in today's episode? Join our free analytics for marketers Slack group at TrustInsights.ai/slash analytics for marketers. See ya next time.