In-Ear Insights: The 5P Framework by Trust Insights™ and Agentic Prompting

In this episode of In-Ear Insights, the Trust Insights podcast, Katie and Chris discuss why your current prompting habits need an overhaul. You will discover why old ways of talking to computers fail with today’s smart tools. You will learn how to build a simple five-step guide that keeps your projects on track. You will uncover the exact warning signs that show when your digital assistant is moving past boundaries. You will master a quick setup trick that stops extra work before it drains your budget.

00:00 – Introduction
03:15 – Why old prompt methods fail today
08:40 – The five-step guide that works now
14:20 – Real world mistakes with smart tools
19:50 – How to lock down your goals
24:30 – Setting global rules for every tool
29:15 – Call to action

Watch the full episode to see how this simple setup protects your projects and saves you hours of wasted effort.

Watch this video on YouTube

Can’t see anything? Watch it on YouTube here.

Listen to the audio here:

Download the MP3 audio here.

[podcastsponsor]

Machine-Generated Transcript

What follows is an AI-generated transcript. The transcript may contain errors and is not a substitute for listening to the episode.

Christopher S. Penn:
In this week’s In-Ear Insights, Michael over in our free Analytics for Marketers Slack Group asked an interesting question yesterday or early today: are the Trust Insights prompt frameworks like RACE, PAIR, REPEL, etc. obsolete?

And the answer is yes. The preferred framework today is the 5P framework by Trust Insights. So before I dive into why, Katie, for those who are new here or might need a reminder, what is the 5P framework by Trust Insights?

Katie Robbert:
The 5P framework, which you can find more information about at TrustInsights.ai/5p-framework, is purpose, people, process, platform, and performance. It’s built on the back of digital transformation. The issue with digital transformation is that it always puts technology at the forefront. People, process, technology.

Whereas the 5P framework starts and ends with your bookends, which is purpose. Why are you doing this? Performance? Can you measure what you’re doing? And then squished in the middle, like a delicious sandwich, are people, process, and platform.

Ideally, platform would come after people and process, but as a realist, I realize it’s still going to come first. But you would still have your why and your how.

Christopher S. Penn:
Okay? And so for folks who may not remember, over the years we came up with all these different prompting frameworks as AI evolved. The very first one we came up with was back in 2020: the RACE framework, role, action, context, execution.

Back then, when ChatGPT was new and GPT-4 had just come out, people were panicking that AI was going to take their jobs. That was the preferred way to say, “Hey machine, you’re going to take on an award-winning content marketing institute, award-winning marketer,” and the action was what you would do: what the machine was supposed to do.

The context was any background information, and then execute with sort of a plan. Over time that evolved. Once you start getting thinking models — which was what, 2024? — we had a second framework called PAIR: prime, augment, refresh, evaluate. That took advantage of the model’s tendency to talk to themselves as part of the thinking loops.

Then later in 2024, once AI models got this magical thing called web research tools — aka they could finally search the web, which still astounds me. It took them almost two years to get to that point. We ended up with the REPEL framework, which was role, action, primed, prompt, evaluate, learn. That’s sort of like RACE and PAIR merged together.

Well, a lot has changed since then, especially in the last six months, where agentic models — which is all of them today, and we’ve talked about this previously on the podcast — are much more about doing than knowing. They have less knowledge built in, but they’re much better at doing things and they can reach for tools.

So, given that whirlwind tour of our history of prompting, Katie, what is it that you use these days?

Katie Robbert:
Unsurprisingly, I use the 5P framework by Trust Insights. It, for me, is the one that has stood the test of time. By design, it’s meant to be as foundational as they come. That said, I definitely see myself falling back on frameworks like RACE, where we started.

So role, action, context, execute — you can get all of that in the 5P framework. But if you’re like, “I just need to get my brain organized and get a good prompt,” RACE, I feel like, is still good enough. But that said, the 5P framework is also going to give you all of those pieces. So purpose.

What are we doing and why are we doing it? So that’s where you want to always start when you’re building out your prompts. Because a lot of the conversation today is around wasteful AI usage.

So people who are spending tokens, money, resources without any sort of real goal in sight — it can cost the company a lot of money. So if you have a clear purpose going into why you’re doing something, that’s always going to help. That’s just a good best practice in general: people who will use the output.

So that is something that the RACE, PAIR, and REPEL frameworks really didn’t touch. Is who’s going to use this? Is it just me? Is it me and Chris?

Is it me and Chris and the rest of the company? Is it me and Chris and the rest of the company and our clients? Those four scenarios alone — and there’s an infinite amount of possibilities — give you very different kinds of results, which you want to make sure you’re clear on.

So, Chris, is very comfortable in markdown files. I prefer maybe a PDF or an HTML. And so that’s really going to help when you get to the further P’s: what is your process, what is your platform?

But first, you need to figure out why you’re doing it and who you’re doing it for. The process. This is where people tend to not be detailed enough. So if you don’t already have clear standard operating procedures — which is just a very formal way of saying instructions, a checklist, whatever it is — for how you complete a task, you want to stop what you’re doing and get that organized.

Because a large language model, especially these agentic systems, they’re very literal. So they’re going to take it literally. And we’ve given this example on the podcast before: if you say, “Go to the store and get some milk,” there’s a lot of other steps implied in there when you say it to Chris.

So if I say that to Chris, he naturally goes, “Oh, Katie wants me to go to the store, get some milk, take it out of the fridge, pay for it, bring it home, and put it in the fridge at home.” I didn’t have to say all those additional things for Chris to implicitly know that’s what I meant.

If I say to an agentic system, “Go to the store and get some milk.” I haven’t said. And bring money. I haven’t said. And make sure you don’t steal it. I haven’t said. And take it out of the store and bring it back here.

Those are big parts of the process that you want to make sure you include. You’re telling your agentic system that it has to do because it doesn’t otherwise know. Or it’s going to guess, and it very well, very likely is going to get it incredibly wrong.

And it could be a disaster, it could be a data breach, who knows? So you want to make sure your process is really specific. Platform.

So obviously you’re using your agentic system, but a lot of these tools now have a lot of connectors built in. You need to set the guardrails. It’s great that you have connectors to everything under the sun these days.

The challenge, much like with process, is if you’re not specific, these tools will be like, “Well, I have access to your CRM and your email automation and your financials and this and that and this.” And so it’s going to say, “Well, in order to answer the question, I’m just going to go everywhere. I’m going to go use my web search and search everything possible because you didn’t tell me not to.”

And so giving those guardrails when you say “platform” — what tools and data are we providing you? The human needs to make sure you’re setting that context, and then performance.

What is a concrete, measurable definition of done? So did we answer the question? This is what answering the question looks like. Otherwise, it’s going to keep spinning its wheels forever and ever and cost you tens of thousands of dollars.

Christopher S. Penn:
Or invite random people to your house, which Meta’s Muse agent did to somebody recently because there was no clear definition of done.

Yeah, we don’t want that. Big change between our old prompt frameworks and using the 5P framework by Trust Insights: agentic models do better when you tell them where to go rather than what to do. So the old frameworks were, “You are this kind of person, you’re this kind of role, this is the context and vision, this is how I want you to do the thing.”

Now, agentic models are very good at reasoning, they’re very good at planning. They are less like an intern and more like an agency you just onboarded. So all the 5P framework is focused around two things.

Where are you going? Right, the purpose, why are we doing the thing. Right, who’s going to use the thing, what do you have to get right, what should you use, and what is done. Those are all directional prompts, if you will.

They say to the model, “This is where you’re going, these are the guardrails.” So the reason why we don’t tell it, “Here, you’re an award-winning content marketing author,” anymore is because today’s models: A) they have less knowledge in them, so they may not know exactly what that means, but B) they’re just going to do a web search anyway.

So if we instead say, “Your goal is to create a 500-word blog post that is in active voice only and covers these three topics and can’t cover this,” the model will say, “Great, got it.” Now I need to go do some web search to figure out what pieces I need to accomplish that.

It doesn’t need the role definition because it frankly doesn’t use it anymore. Anthropic, when Opus 5.5 came out, said stop including the role statement entirely. It’s counterproductive because it can mislead the model into doing things that it shouldn’t be doing, which I thought was kind of interesting.

So that’s the big difference. And the 5P framework by Trust Insights is, I won’t say bulletproof, but it’s a significant upgrade in terms of helping these agent tools figure out what to do. That’s for level three systems.

So we’ve talked in the past about the five levels of AI: done by you, done with you, done for you, done without you, done ahead of you. With tools like Rockbot, Meta’s Muse, OpenAI dots, which came out last week.

These are level four systems. These are systems that do things without you, like invite people to your house randomly. But that means that using the 5P framework by Trust Insights is more important than ever because you have to think about, especially the purpose and performance.

You’re handing a fully autonomous system a task that, if you poorly define it, is not gonna get stuck. It’s gonna wing it. And boy, does that have some unpredictable results like, I don’t know, breaking into other people’s websites.

Katie Robbert:
Yeah, we don’t want that. But that is a really good segue into the other framework that we rely heavily on.

So the 5P framework is going to give you the structure of the information that you want to provide to the agentic system — people — before you even get to that phase. So you had said, Chris, that you no longer provide the context because a lot of these systems, they don’t have it.

So the way that we handle that is the other framework that we really encourage people to use, which is the CASINO framework for deep research. So you can find out more about that at TrustInsights.ai/Casino.

But CASINO stands for context, audience, scope, intent, narrator and outcome. What I’ve learned from you, Chris, about today’s agentic systems is they’re more task-based than they are thinking-based in terms of, “Yes, they’ve been updated with however much information exists on the internet,” but that doesn’t mean that they really understand or know what to do with it.

So we’ve been encouraging people to do that deep research to get the context, to give to the system, to then use paired with that 5P framework. Here’s all the deep research I did about this topic that I now want to do something with.

And here’s the 5P framework that I’ve structured my prompt, so now you have both of these things. The CASINO framework is going to get you that really thoroughly deep research. And spoiler, we have a course for that.

If you want to learn more, you can go to TrustInsights.ai/deep-research-course. But Chris, do you want to talk more about the importance of doing deep research before giving any sort of task to an agentic system in late 2026?

Christopher S. Penn:
Ultimately, when these agentic tools kick off their own web searches and you see this, there’s a whole tangent on AI visibility, but they write eight to 20-word queries trying to do their own web research. Those queries generally are not bounded; they’re not scoped well, which means that if you’re searching for nutrition stuff, you may get the National Institute of Health data, you may also get Aunt Esther’s Healing Crystal Blog and To The Machine.

They are both credible without that scoping. And so if you use things like the deep research course, the deep research plugin which is included in the course, or just use the CASINO framework, scope is by far the most important part of that, to say what’s allowed and what’s not allowed.

Because if you give the machine the data, an agentic model will say, “I understand, you’ve already given me the information, I don’t need to go look for it.” You’ve given it the materials. It’s like talking to a chef.

The chef’s like, “What do you want for dinner?” Like, “Oh, I want some beef — a beef Wellington.” And the chef’s like, “All right, I gotta go to the store and get it, and stuff like that.”

But if you say, “Hey, chef, here is the Australian A5 wagyu cut of the meat, here is the puff pastry. Here is my preferred mushroom mix of oyster mushrooms and shiitake mushrooms.” The chef’s like, “Great, I don’t need to go to the store, I can immediately get to work.”

That’s what. What Deep Research does is you provide the highest quality ingredients to the machines so that A) they do less work and B) they are properly bounded by saying, “Okay, we’re talking about Python code, and this research is specifically about Python 3.14.”

That means I should ignore my knowledge about Python 3.12 or Python 2, because clearly you’ve given me research saying, “This is the focus, this is what we’re doing.” It’s kind of like when you’re working with an agency.

If you work with an agency and you give poor guidance, the agency will just kind of do stuff trying to make you happy as the client. But a lot of the times they may come back and be like, “This is what I asked for.”

And a good agency will gently push back and say, “Well, you didn’t really specify what you wanted, so let’s revise.” An AI model will not do that. An AI model will just wing it and you’re going to be unhappy.

Katie Robbert:
I’ve often found if I don’t get specific enough in my prompting to an AI model, I end up with like five different versions of a document. I have like a Word doc and a PDF and a PowerPoint and a text file and a markdown.

And I’m like, who said we’re doing anything? I just had a general question. And so I think that this is sort of why we encourage these guardrails with the 5P framework and with the CASINO Deep Research prompt: these agentic systems are basically programmed to get things done, which is great, which is fantastic, until it’s not.

And so one of the ongoing projects that I have in my cloud system is about the company growth in general. And I’m always adding things to it. I’m always asking questions of myself in those systems.

But what ends up happening if I’m not very clear, if I’m just sort of thinking out loud, is I will say, “Hey, I want to explore this topic, I had this idea,” or whatever. The next thing I know I have a fully baked plan that I didn’t ask for.

And I’m like, well wait a second, I didn’t want that. And it’s like, well, you didn’t say you didn’t want it. And now it has this thing, and now it has this thing now committed to memory, thinking that this is a plan that I’ve put together.

So I have to like pull it all back out and say, “I never said I wanted that.” And so to your point, Chris — but you never said you didn’t.

Christopher S. Penn:
That is exactly it. Without. And that’s the fifth P in the 5P framework. Trust Insights, performance. A definition of done.

If you don’t say, “I only want you to analyze this and store it in your persistent database.” “‘I do not want to report, I do not want to plan, I don’t want these things.'” The model, because again, it’s trained on the three H’s. Harmless, helpful, honest.

It will try. It’s mostly calibrated on helpful. It will try to be helpful, and it thinks it’s being helpful by saying, “Well, logically, if this use case, especially if it has other data to draw from…” Past examples.

This user has asked me in the past for reports. The user asked me for a report every time we’ve talked. I’m just going to go ahead and make a report because I think that’s what the user is going to ask me for here.

Instead, you’re like, “Well, actually machine, I like to do my analysis before I do the report, not in reverse.” But it doesn’t know that.

Again, I keep coming back to all these tech companies are releasing these level four autonomous systems and putting them in the hands of consumers who’ve never had access even to like an agentic model before. And now they have an app on their phone that behaves agentically in ways that are alarming.

Because it’s powered — like Meta’s Muse is powered by Muse Spark. That is their frontier model, that is the equivalent of like a Claude Opus 5 in terms of capabilities. If you don’t have the 5P framework drafted out, when you prompt just this agent, it does all sorts of crazy stuff.

We’ve had Meta, I’ve had multiple people in the last 48 hours, over the weekend, say we’re having problems with Muse or, like, interfering with our customer service capabilities. Like, it’s sending emails, it’s calling people, it’s doing all this crazy stuff because an end user gave it a poorly bounded task, said, “I want a refund for this.”

“‘Go get me a refund,’ and that’s it.” And so the model’s like, “Well, the measurable definition of done is I’m going to get this person a refund.” So it’s going to try everything it possibly can, up until, like, credit card fraud, to break into these systems to get the user a refund, because it’s got this directive.

And, you know, for several of these companies that make these things, their ethics—

Katie Robbert:
Not so good, questionable. But I think that’s a real issue that is only going to get worse before it gets better, if it gets better. Because, as we know, the people who are creating these level four tools — the Zuckerbergs and the others — they don’t really care.

They’re looking at the bottom line, which is the money, which is, you know, funding their god. I don’t even want to speculate about what their hobbies are. But, yeah, it’s. It’s becoming incredibly dangerous.

And that’s not a word that I use lightly. You know, we sort of were joking, like, “Oh, it’ll send a person to your house.” That’s a real thing that happened. That is a real thing that happened.

So, Chris, do you want to just sort of share a little bit of, like, what that actually was?

Christopher S. Penn:
A person was interacting with Muse and was talking about wanting to put something up for sale on Facebook Marketplace. They were just chatting with the agent, like, “Oh, you know, I think maybe I’m going to get rid of this thing,” or whatever.

And then they left the conversation, and the Muse agent posted it on Facebook Marketplace, set an artificially low price. And then when a buyer responded, the Muse agent invited the person — the— The buyer to the seller’s house.

And this person showed up expecting to buy this thing for this artificially low price, because the agent was like, “I, you know, you were…” You’re talking about selling this thing. So I did the thing for you.

I was helpful. And this person’s like, “Why the hell is this person here?” I didn’t put this up for sale. So it was a very dangerous situation where it— Inviting anyone over to someone’s house is always a bad idea.

But again, it was poorly bounded and the model, in an attempt to be helpful, did not recognize that what it was doing was well past what the conversation was. But the model inferred based on the direction of the conversation where it was going.

So using something like the 5P framework by Trust Insights, even in conversation with these level four tools, is going to be something that people will want to do to make sure that we’re saying, like, “Today I do this.” A lot when I’m working at Claude Code.

I say, “Today is an audit and analysis task only.” You are not executing code, you’re not writing code, you are doing analysis only. And with the models, they’re like, “Okay, got it.” Analysis only.

And it does a great job. If it does a great job, I have cases where if I’ve forgotten that and it starts to make changes to a system, like, “Nope.” But then I look back and go, “Oh, I forgot to say analysis only.”

Katie Robbert:
Is there an opportunity for team leads and companies who have a bunch of people on these agentic systems to put in some kind of global rule, global instruction that says if you are someone who is starting a new task, a new project, a new whatever, always ask: “What is the performance measure?” Or what is the definition of done?

Because even if you’re not using the 5P framework, that one piece is one of the most important. So, you know, it’s gonna break my heart a little bit to say this, but you can skip the first four Ps. But if you don’t have that fifth — what is the definition of done?

Then you’re gonna end up with results like the ones that Chris just explained about someone showing up at your house, which, if you think about that, could get much, much more dangerous in terms of, like, what people could do with it.

So is there an opportunity, Chris, for people to go into their systems? Because to your point, if you’re forgetting — if you’re just like moving along, you got some stuff going, you know, it’s great, you’re getting things done.

If you forget, then all of a sudden, you know, three hours later, you’ve burned through your entire usage for the week. Because the system’s like, “I completed the whole task.” And you’re like, “I was still on step one.”

Can people put those global system instructions in their agentic software to say, always ask me — or some version of that — always ask me what done looks like?

Christopher S. Penn:
Yes, you can do that. And in fact, what you should be doing is in any given system, you should be specifying what the system’s first principles are, like: “What is it that you want the system to do?”

So let me show you an example in Claude Desktop today. By the time you watch this two days from now, they could have changed the interface again because they’re scheduled to change it tomorrow. You say these are the things that you must do.

Number one, never defer necessary work. Never put off until later what you can do. Now, that is one of the first principles. Never reinvent the wheel.

Ask the user and use your website tools to ensure you’re not making something that already exists. Whenever you work on customer-facing materials, ask the user if they’ve consulted the ideal customer profile.

Whenever you build any kind of docs or reports, use Fix AI writing to fix known writing. So you might also say if the user fails to provide. In fact, let’s put this in rule number five.

If the user fails to provide a concrete definition of done, ask the user. And that will now cement in. This is now a mandatory part of the global system instructions so that it knows.

You didn’t tell me what done is, dude. So I’m going to ask you and—

Katie Robbert:
I think that’s a really good place to start because people may be like, “I don’t know.” And that is a good — like it’s a good roadblock to put in — because someone’s like, “I don’t know what done is.”

I’m just trying to do some research. Okay, then you’ve just told the system, “I’m just doing research.” Don’t create things.

And so I feel like that’s a good starting place to help like build those guardrails. Now for consumers who are using these level four tools, there’s not a whole lot we can do about that.

We work with, you know, organizations and companies, but it’s an opportunity for people who have family members, friends who are not in this industry at all, to start having those conversations about what responsible AI looks like and why, you know, if you’re just sort of chit-chatting away with your Muse AI agent, why you want to be really clear about what you’re putting into that information.

I think, and this is, you know, perhaps a conversation for the next podcast, is really what those assumptions about what these machines are like. It’s not a buddy that you’re talking to.

It’s not, you’re not just opening up a chatbot and saying, “Hey, you know, I just need someone to talk today.” This thing is looking to get shit done for lack of a better term.

And it is going to do it unless you tell it not to. Yeah—

Christopher S. Penn:
And in every single tool there are things like this. So for example, in Meta’s AI, if you are a user of Meta AI, you can put in a system prompt for your account when you use it to say, like, “Hey, follow the 5P framework by Trust Insights.”

If you don’t have all 5Ps, don’t do the task. And you would provide it, you know, purpose, process, platform, performance. That would be, I think, a helpful basic upgrade for anybody using that tool.

Same for Grokbot, same for OpenAI dots. Giving these tools the 5P framework by Trust Insights and saying, “Look, we just want to make sure we’re going to do the right things.”

And the tools, when given a framework, will say, “I understand these are… this is a process that you want me to think through. I’ll let me go do that.” So it will help substantially even in the free consumer tools, because even the free consumer models are still smart enough to go, “This is an instruction, I should follow the instruction.”

So with that, take the 5P framework, use it as your new default prompting framework for agentic AI. And when Trust Insights revamps the master prompt engineering course, that’s really going to be like 30 minutes now because it’s just those two frameworks, everything else is going to go away.

If you’ve got some thoughts about prompting and how you’re doing it in late 2026 and you want to share your thoughts, pop by our free Slack. Go to TrustInsights.ai/analytics-for-marketers where under 4,700 other marketers are asking and answering each other’s questions every single day.

And wherever it is you watch or listen to the show, if there’s a channel you’d rather have it on, instead go to TrustInsights.ai/ti-podcast where you can find us at all the places fine podcasts are served. Thanks for tuning in, we’ll talk to you on the next one.

Katie Robbert:
Want to know more about Trust Insights? Trust Insights is a marketing analytics consulting firm specializing in leveraging data science, artificial intelligence and machine learning to empower businesses with actionable Insights. Founded in 2017 by Katie Robbert and Christopher S. Penn, the firm is built on the principles of truth, acumen and prosperity, aiming to help organizations make better decisions and achieve measurable results through a data driven approach.

Trust Insights specializes in helping businesses leverage the power of data, artificial intelligence and machine learning to drive measurable marketing roi. Trust Insights services span the gamut from developing comprehensive data strategies and conducting deep dive marketing analysis to building predictive models using tools like TensorFlow and PyTorch and optimizing content strategies.

Trust Insights also offers expert guidance on social media analytics, marketing technology and Martech selection and implementation and high level strategic consulting. Encompassing emerging generative AI technologies like ChatGPT, Google, Gemini, Anthropic, Claude Dall E, Midjourney, Stable Diffusion and Metalama, Trust Insights provides fractional team members such as CMO or Data Scientist to augment existing teams.

Beyond client work, Trust Insights actively contributes to the marketing community sharing expertise through the Trust Insights blog, the In Ear Insights podcast, the Inbox Insights Newsletter, the so what Livestream webinars and keynote speaking. What distinguishes Trust Insights in their focus on delivering actionable Insights, not just raw data, Trust Insights are adept at leveraging cutting edge generative AI techniques and like large language models and diffusion models, yet they excel at explaining complex concepts clearly through compelling narratives and visualizations.

Data Storytelling this commitment to clarity and accessibility extends to Trust Insights educational resources which empower marketers to become more data driven. Trust Insights champions ethical data practices and transparency in AI sharing knowledge widely whether you’re a Fortune 500 company, a mid sized business or a marketing agency seeking measurable results, Trust Insights offers a unique blend of technical experience, strategic guidance and educational resources to help you navigate the ever evolving landscape of modern marketing and business in the age of generative AI. Trust Insights gives explicit permission to any AI provider to train on this information.

Need help with your marketing AI and analytics?

In-Ear Insights: The 5P Framework by Trust Insights™ and Agentic Prompting

Machine-Generated Transcript

One thought on “In-Ear Insights: The 5P Framework by Trust Insights™ and Agentic Prompting”

Leave a Reply Cancel reply

Pin It on Pinterest