So What? How to refresh your Generative AI for 2026

Katie Robbert – 00:00

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.

Christopher Penn – 00:46

Happy New Year.

John Wall – 00:47

Thanks.

Christopher Penn – 00:48

Yes. Happy New Year.

Katie Robbert – 00:49

We got jazz hands from John.

John Wall – 00:51

Jazz hands for the new year. That’s my revolution.

Katie Robbert – 00:56

We wanted to kick off the new year with some housekeeping because it is a good time to reset and refresh—to see what you’ve been working on, what you’re carrying into the new year, and what you’re ditching.

Unsurprisingly, let’s start with generative AI. On today’s episode, we are going to talk about how to refresh your generative AI for 2026. This sounds like a broad topic, but I am fairly certain what we want to do is narrow it down and try to walk through where to start and what things you should be putting on a regular maintenance schedule. When you need to think about whether the model changes, do you need to change your instructions or can you just leave it? Chris, where are we starting today?

Christopher Penn – 01:40

We’re going to start with three things: prompts, environments, and models. Those are the three things that we want to talk about today. Regarding prompts, we have some ways to tidy up around the house on that front. For environments, we have some of 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. Finally, regarding models, we will discuss which of the models to use.

A lot of this deals with the fact that it is 2026 and things have changed. One of the things that I have seen in a lot of other places is people stuck in 2024 in terms of how they’re using generative AI.

Christopher Penn – 02:33

Let’s start with prompts and a couple of things that I wanted to remind folks about. In your prompting techniques these days, one of the things that has proven most effective is to establish what are called first principles instead of a giant long list of everything to do and not do. This is a way of saying, “Here are the three most important things.”

Let’s say we were trying to refresh our sales handbook. Let me ask you both this: what would be the top three things that you would want AI to know as cardinal rules about the way Trust Insights evaluates a potential sales prospect or how we do business?

Katie Robbert – 03:27

I guess this goes to knowledge blocks. A number one cardinal rule would be that in order to appropriately update the sales playbook, you have to refer to our “About Us” knowledge block, which would need to be updated. Then our Ideal Customer Profile (ICP) knowledge block would also need to be updated. You can only refer to those two knowledge blocks in terms of how we conduct business and for whom we conduct business. You cannot look outside of those datasets to figure out what Trust Insights does.

Katie Robbert – 04:15

The onus is on us to make sure we provide that context. That is a guardrail that I would want to set right away. This is who we are, this is what we do, and this is who we do it for. If you start going outside of that, then you have broken the rule.

Christopher Penn – 04:33

I think that list—this is who you are, this is what we do, and this is who we do it for—constitutes your sales first principles. If you don’t have one of those three things, you can’t do the sale, and you certainly shouldn’t be using AI to do it.

John, when you’re looking at how you would give instructions to a new sales team member, what would be the three rules that you would say are essential?

John Wall – 05:03

First, you have to find out if they have normal BANT—do they have budget, authority, need, and timeline? Then you do some exploration to figure out if there’s actually a deal there and determine what they can actually work on. Finally, you address all the process and procedure as far as getting them a scope, what goes on in the sales process, and where everything goes. I think between that, the past data on deals we’ve done, and information on everything else that has come before, that is pretty much the whole process.

Christopher Penn – 05:41

If you distilled that down to a set of rules, what would you tell a new person named Bob who joins our sales team this year? What are the three rules that you tell Bob he must do every day as his first principles?

John Wall – 06:00

Clean out the inbound queue, review the existing deals, and then explore for new business. I guess those would be the three process jobs.

Christopher Penn – 06:10

What you’ve both done is defined a set of first principles. When you’re building things with generative AI, that is how you lead. When you’re prompting these things, if you were saying we need to build an ideal customer profile, those principles should be part of the ICP construction process. There is no point in having an ideal customer profile if it isn’t aligned with how you want to sell things.

It is like if you were to just set off a research agent and say, “Go find us customers that are of this size, budget, and industry,” but you didn’t provide the rules. It is going to bring back stuff that’s not on target. It might say people who sell biowarfare weapons are your ideal customer because they meet the budget.

Katie Robbert – 07:07

When you say people are stuck in 2024—which was only a year ago last week and now it’s two years ago—for many of us, we were taught to start with context. You say, “You are a super experienced, high-performing salesperson.” Is that no longer something that we have to do in 2026 with generative AI prompting?

Christopher Penn – 07:49

You can, and it still works, but the models have changed so much in terms of their ability to do reasoning that they can sometimes start to infer things. You would still want to say, “Today we are working on building ideal customer profiles for Trust Insights. Here is who we are, what we do, and who we do it for.” You give it those first sets of three principles.

Now, here is the next twist on the ideal customer profile. It is good to have it define pain points, goals, and motivations, but you want to do a second version that turns it into a checklist.

Christopher Penn – 08:37

You want to turn it into a checklist for any given task. This checklist contains the criteria to check off against the task. Does this match our profile’s needs? Does this match our pain points? What’s turning out to be true across the four different model families for reasoning models is that telling them to “check your work” or “be thorough” is not as impactful as providing a list and saying, “Go down this list and check your work against it.”

Katie Robbert – 09:22

Where does that come from? That feels like a new thing. Coming from a less technical lens, we were taught to say, “Check your work” or “Validate what you’ve done.” It sounds like the structure has moved from validation to a checklist. When did that happen?

Christopher Penn – 09:57

It happened last year. As reasoning models have evolved—and there are three different families of reasoning models with formal names like Monte Carlo tree search, PRMs, and backtrack—the models are getting much better at figuring out guardrails for a task, but they still struggle to do it accurately.

Back in 2023, you would tell a model to “think this through step by step.” That is too vague. If you give a model the checklist—a concept that came out in a paper in June of last year—and tell it to verify its work step by step off of that checklist, it performs much better. It goes through customer needs, analytics software, and data, checking them off one by one. It is the same way you would delegate to a junior employee. You wouldn’t just say, “Go check your work.” You would say, “These are the five things I’m looking for.”

Katie Robbert – 11:13

That feels like common sense because that is how I operate, but many people are still trying to figure out how to get the best result. Majority of people aren’t looking to academic papers; they’re looking to experts like you. To be fair, this week is the first time I’m hearing about the checklist, and you’re saying it has been around for six months.

Katie Robbert – 11:54

We can get into how average users stay up to date in a different episode, but I want to acknowledge that this feels like new information even though it is also common sense.

Christopher Penn – 12:10

It is newer. It was really brought to light by the folks at Alibaba who make the Qwen family of models. They won awards in late 2025 at the Neurips conference specifically for how they’ve trained their models to recognize and generate checklists natively as a way of making the model smarter. This is why we have our livestream, so we can go through this stuff.

Katie Robbert – 12:55

The way John spoke about what he would tell a new sales associate is very much in that checklist format. John, you’re saying number one is to check the backlog, number two is to check the pipeline, and number three is to go searching for new business. That is the checklist.

Christopher Penn – 13:28

Exactly. That’s part one of four in prompting your generative AI refresh for this year. Second, make sure you do regular maintenance on your prompt library. As mentioned at the top of the show, any time your workhorse model changes, you should be revisiting it for optimization.

For example, we’re a Google shop and have been a Google Workspace customer for a while. Gemini is our workhorse model. When Gemini 3 came out, both Pro and Flash, it was a big change. The model’s behaviors changed.

Christopher Penn – 14:17

Our team member Kelsey mentioned in late November that her Gems started going off the rails because the underlying model totally changed. As part of your 2026 refresh, if you have not put your prompts through an optimizer, this is the time to do it. Every company that makes AI tools has some form of prompt optimizer. ChatGPT has a prompt optimizer in the OpenAI platform. Claude has a “generate prompt” feature in the Claude developer platform.

Christopher Penn – 15:15

You put in your current Claude prompt, and it will improve it. Google puts their prompt optimizer in a more obtuse way inside of a Vertex notebook that you have to run inside Google Cloud. However, it is the most thorough way to optimize prompts because it actually does internal testing to see which prompt delivers the best result.

For casual use, it is just as easy to ask the model, “What are the five things you would do to improve this prompt? What are the three questions you have for me about my intent?” Then have it rewrite the prompt.

Christopher Penn – 15:55

At a high level, if your workhorse model changes, it’s time to clean house and revise those prompts.

Katie Robbert – 16:05

John, when you’re creating prompts, are you asking the AI to check its work or are you using a checklist style?

John Wall – 16:21

As long as I get output that makes sense, I’m fine. I don’t spend a lot of time massaging it. I like to take existing work and ask, “What was missed in here?” It usually comes back with great stuff. I haven’t reached a point where old Gems have come through as broken, but I’m also the user who will fire up an app and only then realize it was deprecated two years ago.

Katie Robbert – 16:52

So perhaps you’re not the one we should be asking.

John Wall – 16:55

I’m the worst-case scenario user when we’re talking about UX. That’s why I’m excited to hear about this today. I’ve got a stack of 45 different prompts I randomly roll out. Should I run all of them and see how they work, or should I wait until I hear an alarm bell?

Christopher Penn – 17:22

In general, when your workhorse model has a big version change, like Gemini 2.5 to Gemini 3, that’s when it’s time to do some QA and test your prompts. Gemini 3, in particular, is very sensitive to conflicting instructions. When a model changes, take the 10 prompts you use most and test them to see if you’re still getting a good result.

Katie Robbert – 18:01

I was planning on doing that with the system instructions for one of my most used Gems. I assume the same maintenance is needed for those because it is essentially the same thing.

Christopher Penn – 18:19

Exactly. The last thing to start thinking about—which is something more savvy folks are doing—is what skills your AI should have. We covered this last year in our show on Claude Skills. Since then, Anthropic has made the skill protocol public to all AI makers so other companies can use it in their software. Using that skill format is going to become popular.

Here is how to know you’re using a skill that you need AI to learn: you keep retyping the same instruction over and over again.

Christopher Penn – 19:20

If you are always telling it to remove passive voice and convert it to active voice, you can now say, “Claude, let’s build a skill that converts passive voice to active voice.” Then, as you are prompting, you can just say, “Claude, use the passive voice skill.” You won’t have to type that paragraph anymore. More importantly, as you use agents within that system, they will know the skill exists and will use it autonomously.

Christopher Penn – 20:08

I did this recently for fun while writing a trashy romance novel. I told the editor agent, “Make sure you use the anti-passive voice skill,” and it automatically went through and did that. Skills are ways to embed the checklist into a repeatable process so you don’t have to copy and paste it every time. It’s like Gems or GPTs on steroids.

Katie Robbert – 20:37

You just said a lot, so let me back up. We covered skills in a previous livestream, which you can find on the Trust Insights YouTube channel. Basically, you’re saying that if you always add instructions to clean up a transcript or apply a writing style, those should be skills within your generative AI system. It automates things more so that a user like John Wall doesn’t have to remember exactly how he was correcting for passive voice.

Christopher Penn – 22:07

Correct. In John’s case, think about those first principles. If you use BANT as a framework and have a checklist spelled out, you might create a BANT skill in Claude. Then, when you’ve connected Claude to HubSpot, you could say, “A new opportunity came in. Claude, run BANT.” It will pull the record from HubSpot, run it through the criteria, and tell you how it scores on a scale of zero to 10. You’ll know immediately if it’s worth the time.

Christopher Penn – 22:44

As we saw at the end of 2025 with Google Workspace Studio, more and more of these capabilities will become autonomous. If you get a skill well thought out today, you won’t have to reinvent the wheel when it becomes available in your ecosystem as a feature or part of an agent.

Katie Robbert – 23:40

Regarding maintenance, is this something we should be doing at least quarterly?

Christopher Penn – 23:47

Ideally, yes. Just check in to see if there have been major shifts. For example, Imagen 3 came out last year, and newer models will be coming out in the next few weeks. If you’re doing image generation, you’ll want to know when that model changes because your images may look different, or be faster and cheaper. I recommend folks join the Trust Insights newsletter and Slack community because we share this information as it happens.

The next pillar for refreshing your AI is the environment you use it in.

Christopher Penn – 24:40

There are four places where you use AI. Number one is the obvious one: going to ChatGPT. That’s the environment everyone knows. That is very 2024—still useful, but there are three more places to think about this year.

The first is command line tools. These exist in your terminal or command line. There is Claude Code if you’re an Anthropic customer on a paid plan. There is OpenAI Codex from the makers of ChatGPT. There is Qwen Code if you’re in the Alibaba Qwen system, and of course, the Gemini coding app.

Christopher Penn – 25:42

If you install them on your computer, the big advantage is they can access and write local files with your permission. You could have a “Co-CEO” help you keep a diary. You tell it what happened every day, and it takes notes in little files in a folder on your computer. Two months later, you can ask, “We were talking about this idea—did we ever do anything with that?” It can search through those local files and give you an answer.

Katie Robbert – 26:30

Do I have to be coding in Terminal to make that happen?

Christopher Penn – 26:39

You have to use Terminal, but you do not have to code. You do not write a single line of code. You chat with it just like you chat with ChatGPT.

Christopher Penn – 27:14

I just turned on Claude Code. All of these, once installed, work in whatever folder you want to work in. You get your greetings, and then you could say, “I want to start keeping a diary. Can you ask me questions about my day and write a date-stamped file for each day’s entries?” No coding, nothing complicated. Claude will help you do that.

Christopher Penn – 28:17

You come back each day and tell it about your day. It will ask specifics: “What was the highlight? Did anything frustrating happen? What did you accomplish?”

Katie Robbert – 28:34

I assume you can tell it to ask specific questions instead.

Christopher Penn – 28:38

Exactly. Because it has access to the file system, it can write things down. That is the big secret that makes these tools better than web-based ones. You can also invoke separate skills. You might have a set of skills for virtual board members—the skeptical one, the crass one, and the optimistic one.

Christopher Penn – 29:35

You might say, “I want to talk about my Q1 plans. Let’s have all three board members weigh in.” It invokes three separate skills for a miniature roleplay. You have a conversation with them in the chat, and you can tell Claude to write down the minutes from that virtual meeting.

Katie Robbert – 30:24

Can you also do this on a Windows machine?

Christopher Penn – 30:35

You sure can. In Windows, it’s called the command prompt. In Mac, it’s called Terminal.

Christopher Penn – 30:50

In Windows, you go to the Start menu and type “command” in the search box to find it.

John Wall – 31:02

Is this using an API to call their web servers, or have you installed the models locally?

Christopher Penn – 31:11

No, this uses their online models. An important clarification is that you have to be a paying customer. These are not available for free users. You have to be on at least the $20 a month plan to be useful.

If you refuse to use Terminal, there are two additional environments that are full-fledged applications. One is called Anti Gravity by Google. It functions similarly to Terminal but it’s an app. You can use their Agent Manager to have those same conversations and take notes.

Christopher Penn – 32:44

Anti Gravity works with Gemini, Claude, and the OpenAI GPT models. For Claude and Gemini, you need a paying subscription. The second app is called VS Codium or Visual Studio Code, with an extension called Kilo Code. You can choose gazillions of different models in that one. If you want to edit the Trust Insights newsletter, you could do it in that environment.

Katie Robbert – 33:25

I’m going to be honest, these don’t look more modern, but it explains a lot about the UIs you create. Today, it all feels very overwhelming and intimidating. John, what do you think is the barrier to entry for using Terminal or these standalone apps? Does it feel too technical?

John Wall – 34:05

I don’t see a need that would require me to do that yet. I see the value if you’ve got scripts running, but I’m usually not having a script fire every week to put output in a folder. It is good to know it’s there, but I don’t have anything off the top of my head I would need it for.

Christopher Penn – 34:41

Think about the Ideal Customer Profile. You might have an ICP skill in one of these environments. When you’re working on a Q2 promotion or a pitch to a sponsor for “Marketing Over Coffee,” you could just say, “ICP, what do you think?” It is a shortcut on tap. You can still copy and paste out of your prompt library, but having it built into your ecosystem is faster.

Katie Robbert – 35:27

I usually just attach a PDF of the ICP and say, “Make sure it aligns with this criteria,” which doesn’t feel cumbersome to me.

Christopher Penn – 35:44

These infrastructure-heavy uses of AI are not mandatory. They are faster and more efficient, but you can get the same result manually. What these tools allow you to do is scale much faster. Instead of taking six hours to copy and paste promotions individually, 20 agents could launch at once, use the same ICP skill, and be done in 10 minutes.

Katie Robbert – 37:48

You really have to look at yourself as an individual and say, “What is it that I need to be doing?” We want to challenge you in 2026 to figure out what works for you, not just what works for Chris. I will probably take a look at the Terminal stuff because it makes sense, but I’ll also use the systems natively.

Christopher Penn – 38:47

The last thing everyone should have is a plan for a backup that isn’t a “big tech” provider. This means local models. These are models you run on your own hardware using tools like AnythingLLM or LM Studio.

The economics of generative AI today are not sustainable. OpenAI is burning through billions of dollars.

Christopher Penn – 39:43

Google can backstop the funding for a while, but even they cannot burn money indefinitely. There is a non-zero chance that one of the big players does not make it out of 2026 intact. Technology isn’t going anywhere, but providers might. It is a good idea to do some scenario planning.

Katie Robbert – 40:30

Being reliant on technology is always a risk, but the concept of a backup is not new. We have talked before about having Matomo running in case Google Analytics implodes, or backing up your CRM data locally in case HubSpot goes away. The generative AI space is very volatile and the speed of change is giving people whiplash.

Christopher Penn – 41:41

I recommend LM Studio with one of Alibaba’s Qwen models. If you use the model by itself, it is small and prone to hallucinations. If I ask, “Who is the CEO of Trust Insights?,” it might invent a name because it has no grounding.

Christopher Penn – 42:32

However, if you set up the ability for it to do a web search, it becomes reliable. It can find Katie Robbert, the LinkedIn page, and the Trust Insights page. This is a big change from a year ago when local models were not trained to do that.

Katie Robbert – 43:19

I would also add an audit of all the custom GPTs, Artifacts, and Gems you’ve built. You’ve probably built dozens that felt like a great idea at the time but are now collecting dust. Sunset them, delete them, or archive them. Grab the system instructions and put them in a doc. Especially in larger organizations, you want to make sure people are only using the versions you want them to use. Just build auditing into your plan once a quarter.

Christopher Penn – 44:59

We had a question from Brian asking about non-developers doing AI coding and creating apps for clients. About 95% of what we do is for internal use or operated on behalf of a client. We could do 12 livestreams in a row on AI coding practices because there are far more ways for it to go wrong than right.

Katie Robbert – 45:38

Software development is a profession that takes time to learn. There is a reason there is a whole team—managers, architects, and testers—around any individual developer. My unsolicited advice is to build things internally first. Do not put them out for public use because then you have to deal with support, maintenance, and bug fixing. That is a whole different infrastructure.

Christopher Penn – 46:54

Even with AI, there are things that go wrong immediately. To Katie’s point, there is a baseline amount of domain knowledge you need to have about the profession of software development to do it well.

Katie Robbert – 47:21

It is very easy to stand up an application using Claude Code or Google Developer Tools, but those apps can be fragile, riddled with privacy holes, and lacking data integrity. People often don’t think about where the data goes or what happens if someone needs it later.

Christopher Penn – 48:20

Prompts, environments, and models are the three pillars you need to refresh. That is above and beyond all the other data governance and maintenance you need to do. Put those on your 2026 spring cleaning list.

John Wall – 48:59

Using these tools to write code that runs once is one thing, but the hardening of software is a whole different level. I would be interested in a future stream about automated testing and requirements gathering. Having more people “running with scissors” is not a better world.

Christopher Penn – 49:26

If you want an example of how bad the problem is, search for “OpenAI key” with the site restriction for GitHub. You will see so many people putting their private credentials in public because they did not follow basic procedures.

Christopher Penn – 50:04

That’s going to do it for this week’s show. Thanks for tuning in. Be sure to subscribe wherever you’re watching. For more resources, check out the Trust Insights podcast at TrustInsights.ai/tipodcast and our weekly email newsletter at TrustInsights.ai/newsletter. If you have questions, join our free Analytics for Marketers Slack group at TrustInsights.ai/analyticsformarketers. See you next time.