So What? Agentic AI Prompting

Katie Robbert – 00:41

Well, hey everyone. Happy Thursday. Welcome to So What, the Marketing Analytics and Insights live show. I am joined by both Chris and John through the miracle of space and time. Chris just got off a plane, John has internet, and we’re all here, so it’s perfect.

This week we are talking about how to prompt agentic AI. Everyone’s talking about AI, agents, and building your own army of robots. All that stuff is great, but not if you don’t know how to talk to them and tell them exactly what it is you need them to do.

Chris, I know you ran an experiment over the weekend, which I was just looking at and unsurprisingly, I have questions, but I’ll save them so I don’t get ahead of us. Before we get into how to prompt agentic AI, do you want to set the stage for what it is we’re even looking at?

Christopher Penn – 01:38

Sure. I apologize, I’ve had about two hours of sleep entirely. If this comes out as garbled delirium, just let us know in the comments.

Katie Robbert – 01:48

It’s what John and I are here for. We’re just the comic relief and we’ll keep the show moving. When Chris just kind of keels over, we’ll be like, “Hey, so did you catch the news today?”

Christopher Penn – 02:00

If we think of an AI model as an engine—GPT-3.5, GPT-3, or the brand new GLM-5 that came out yesterday right in the middle of my workshop. I was saying GLM-4.7 is the latest model, then never mind, I have to change that slide. It literally dropped in the middle of the workshop, which was hilarious.

Claude, Opus 4.6—these are models. Models are the engines of AI. The way that we’ve been using them for the last two or three years is inside a model harness interface like ChatGPT, Google Gemini, or Claude. By the way, funny enough, folks who are native French speakers pronounce it totally different. They call it “Claude,” and during the workshop I was like, “Who? Oh, Claude.”

They provided these web interfaces with a lot of back-end stuff like safety rules. That’s how we come to use AI. You open up ChatGPT, you talk to it, it does stuff, and then you copy-paste and go on about your day. Agentic AI is all about saying, “Well, you’ve got this great engine. What about the rest of the car?”

Maybe I don’t want to be the copy-paste monkey. A lot of people have been talking about how there’s no ROI of AI, depending on which study you read. A good part of that is because if you are just in vanilla ChatGPT and you’re copying and pasting everything in and out, of course you’re not going to get any massive savings because you’re still the bottleneck.

Agentic AI is when you say, “Let’s take the engine, build the rest of the car around it, and connect it to things.” Let’s connect it to our Google Workspace so it can read my Gmail and my calendar. Let’s connect it to Asana, our task management system. Let’s connect it to HubSpot, our CRM, and let’s give it a different interface where now it has more autonomy. You could say, “Pull the last 20 opportunities from HubSpot and tell me how well they fit our ideal customer profile.”

As long as you’ve provided the pieces, the model harness that you’re using—like a Claude Code, Claude Co-work, or Google Anti-gravity—can say, “I’ll go and do that, I’ll come right back and tell you what I found.” That’s the difference between the AI that we’ve been using from 2023 to 2025 and the AI today of 2026.

Katie Robbert – 04:37

When we first started talking about generative AI, we were giving examples of best practices for prompting. We have things like the REPEL framework, which you can get at TrustInsights.ai/repel. We have things like the RACE framework: Role, Action, Context, Execution—the basics.

Obviously, as the technology becomes more sophisticated, the way we need to think about interacting with it becomes more sophisticated. The whole point of today’s episode is how to prompt agentic AI. Do the frameworks that we educate people on—such as RACE, REPEL, and PARE—still apply, or have things changed? What do we need to know today?

Christopher Penn – 05:33

Andrej Karpathy, the co-founder of OpenAI, said in 2023 that the hottest programming language was English. In 2026, the hottest programming language is project management skills. When you are prompting agentic AI, you are no longer giving step-by-step individual instructions to a machine. You are instead giving the machine a plan. You are giving it a project, a PRD, or a requirements document that it can autonomously pick up, run with, and go and do something with.

I’m going to use an example from this week’s workshop. This is the world-famous Gopher Hole Museum in Torrington, Alberta, which is a museum of 45 dioramas of gophers in common Alberta life. It’s this cute little place about an hour north of Calgary. I think it’s literally a one-person show and they’re only open during the summer.

Katie Robbert – 06:42

I’m not going to lie, I would totally be down to go visit that.

Christopher Penn – 06:46

Oh, totally. Part of the talk was to say that static content, like a standard webpage, is very vulnerable to AI overviews and to AI replacement in search. You don’t need to go to their website anymore, which means the website owner loses the ability to try and convert a visitor because you don’t get the visitor anymore.

I asked: what would it look like if we had an agentic system like Claude Code take a shot at making something interactive that AI can’t summarize—something fun instead of just a static page? Not that there’s anything wrong with a static page. I had to sit down and think about how I would do that in an agentic way.

The single best framework for prompting agentic AI today is the 5P Framework. It contains all the ingredients that an AI agent would need to do the task. What is the Purpose? We’ve got a static website where AI can take traffic away. We need something interactive and dynamic that can’t be summarized to keep traffic on our website. Also, because this is a seasonal tourist attraction, they need to raise revenue in the off-season to pay the bills.

The People involved would be people like me who would be using agentic AI and directing it. This is a critical part: today’s agentic tools allow you to create virtual teams of people. In Claude Code, you can say, “I need a UI designer, a UX designer, a gamification expert, an SEO expert, and a project manager.” You specify in your project plan, based on the 5Ps, the virtual people who belong on this project.

The Process is how you want to approach this. You gather the team and do some brainstorming. What are some ways that we could make the site interactive to give people a reason to come to the website or even a reason to donate?

The Platform is something that you would likely have the agents partially design. You have a conversation just like you would in requirements gathering. It has to be something lightweight, as this company does not have the resources for a massive server farm. It has to be client-side and super non-technical.

Performance asks: can you actually build the thing? When you’re done, does it work right? Does it meet those requirements? I was doing this on the plane ride to Alberta on Sunday. I wrote out the 5Ps and gave them to Claude Code. We went back and forth for about 15 minutes, and then I said, “Go.” It spun up the agent teams.

I’m going to share my screen here. This is hosted on our website at TrustInsights.ai/interactives/taya-gopher.

Katie Robbert – 10:48

You know I love surprises, Chris.

Christopher Penn – 10:50

It consumes no resources because we put it on the client side. You can choose your diorama, pick the gopher you would like to add, and give it cute outfits like a Royal Mountie’s uniform or a prop like a broom. At the bottom is a button that says you can order a print, which ties into the Gelato API, a print-on-demand shop.

They allow a company like the Gopher Hole Museum to essentially be an affiliate. If someone makes a cute diorama and hits “order print,” the museum would get about $2 off an $8 print. There are also ways to visit the museum, go to the gift shop, and donate. All of this came together in about 45 minutes.

However, it needed the 5P Framework as an agentic prompt to pull it off. If you specify the Purpose and the Performance—telling an agentic system what success looks like—it will figure out most of the rest.

Katie Robbert – 12:14

I love this because it shows the flexibility of the 5P Framework. I was having a conversation with Kelsey, our account manager, the other day. She asked, “What if I don’t know all of the Ps?” I said that’s okay, because this framework is meant to highlight where you don’t have enough information. You can go get it before you start building things and racking up bills.

I really appreciate that you feel the 5P Framework is right here. You knew a lot of information going in, but you probably iterated with Claude Code to answer questions where you had gaps. You can use this as an anchor to make sure you have all that information before you start building. If I haven’t defined my Performance, can you help me figure out my measure of success? If I know what I want to build—that’s my Purpose—is that clear or do I need to refine it? I’m very jazzed about this.

Christopher Penn – 14:06

Even if I have a lot of information, the 5P Framework is still important because I don’t ever have all of the information. As I’m doing more advanced projects, the number of blind spots I have is increasing.

For example, I wanted to make a new kind of video game, something simple like Wordle. I asked the AI what technologies in the platform are available. It suggested a technology called PocketBase, which takes a SQLite database and wraps it inside of its own harness to have an API. Where has this been for the last year while I’ve been making things with SQLite databases?

In another case for a client project, V1 was not going well. I said, “Let’s do a V2.” I told the AI to assume everything I wrote in the requirements document for V1 was wrong. It told me I picked the wrong database and suggested one that supports vector search. If we want to use generative AI to understand thematically what’s in the database, PostgreSQL with PGvector will do that, whereas my old choice would not.

Part of the 5P Framework when prompting agentic AI is to ask: “What are my blind spots? What don’t I know? Ask me questions about what portion of the Purpose I did not think through.” In the game example, I asked about monetization, and it told me not to do it a certain way because it would ruin the game. It suggested building in sponsorships in a way that’s not crappy for players.

One of the keys of agentic prompting is that you need to spend about three times as much time planning as you do anything else. If you get the plan right, you can hand that to the agent and it just goes and does it. If your plan sucks, you’re going to get a steaming pile of crap.

Katie Robbert – 17:31

It’s almost like someone has been saying this forever. Even before generative AI upended our lives, someone has been preaching that if you just got your planning straight, you could execute things so much faster. I am relishing the validation and taking full credit for being the one yelling to no one.

Hopefully people are listening now because the stakes are a lot higher. There is a very tangible cost associated with agentic AI tools. You can’t be spending a ton of money and getting no results. That’s what happens when you don’t do your planning.

Christopher Penn – 19:38

Right. And I would suggest that within your agent system, you build it into the system. In my version of Claude Co-work and Claude Code, I actually created a Trust Insights 5P plugin. Inside is “Co-CEO Katie” to help with it, along with a 5P Framework skill and a fact-check skill. When I say let’s do this thing, Co-CEO Katie will say, “I’m going to give this a read-through and force you to think through all parts of the five Ps.”

Or, if you want to go crazy, you could create an agentic version of yourself to talk to the agent version of Katie and the two of them can debate. You can come back later and see how things turned out.

Katie Robbert – 20:40

Well, that doesn’t really work for me because I am Co-CEO Katie. When I talk to the Co-CEO, I’m talking to myself and I already find that very odd. I got stuck in this weird visual of Katie talking to Co-CEO Katie and them coming to a standstill because they’re both stubborn.

Christopher Penn – 21:23

Let’s see how you would put this to use. I’m going to use my Data Diaries project, which is what I use to edit the Data Diaries every week in the Trust Insights newsletter. I’ll drop in my “recipe,” which is the general prompt I use, and then we’ll do it agentically. We’ll say, “Using the Co-CEO and the 5P Framework, help me reformat and rewrite this agentic prompt into the full 5P Framework.”

The current recipe basically says this is an editing task using the writer agent, editor agent, and writing style analysis skills. It reviews the rough transcript and transforms it into a polished newsletter piece, cleaning up speech disfluencies and bad grammar. Then it uses the Voice of the Customer agent with the ICP provided to ensure it aligns with what a customer would find helpful.

Katie Robbert – 23:05

Thank you for making that bigger. Whenever my face is down here, it means I can’t read the screen.

Christopher Penn – 23:18

It says the current recipe is a rough conversational prompt that mixes goals, tools, processes, and quality standards in a single run-on paragraph. The structure makes it hard for an AI orchestrator to parse priority sequence dependencies. Let’s now apply the 5P Framework with Co-CEO rigor to do this.

Katie Robbert – 23:50

I appreciate that it’s taking a minute. As simplistic as the 5P Framework is, you really need to think through what goes into each piece. It is an opportunity to think through whether you have enough information. I think of it as the full 360 of business requirements. You can move forward without those pieces, but you’re not going to get the best result.

Christopher Penn – 24:36

It definitely pointed out a bunch of different gaps. Because the original intent is clear and the task is narrow—take this transcript and make it better—it’s not a big strategic or complex piece of code. For larger stuff, you’ll go through rounds.

What the 5P version delivers is Purpose, People, Process, Platform, and Performance. The Purpose answers how to consistently transform a rough voice transcript into a polished educational newsletter piece that meets the Trust Insights brand standard and resonates with our ICP. It solves the problem of speech disfluencies and unstructured logic. Success looks like a finished Data Diaries column that adheres to our writing style and contains no unverified claims.

The People involved are documented as AI agents. This contains the “character cards” I made from our ICPs—named personalities that I use every week.

Katie Robbert – 26:32

It’s funny you say that because offline, you gave me research and my first question was, “Who are these people? I’m not paying them. They’re not on the payroll. Get out of here.”

Christopher Penn – 26:48

Yes, they’re all synthetic characters. For the Process, it reads the transcripts, style guide, and ICPs. It cleans up speech disfluencies and maps the logical structure using the “What, So What, Now What” framework. The writing agent, editor agent, and voice of the customer agent all offer their perspectives. The fact-check skill runs to validate claims.

A critical part of agentic prompting is specifying the order of operations. Otherwise, tools will either run sequentially—which is slow and saves no time—or they’ll launch all agents at once. If they run at once, you’ll have race conditions where an agent does its part before the part it depends on is ready.

Here it says phases four and five can run in parallel if the editor’s output is available, but phase six runs after four and five.

Katie Robbert – 28:37

A project plan with dependencies. Who knew?

Christopher Penn – 28:41

Exactly. Here is the Platform—the tools and dependencies required—and the Performance measures. Did it comply with the writing style? Would all five ICPs find it valuable? Is it factually accurate? It defines when the agent teams are done.

This is now ready for prime time. That one-paragraph recipe is now a full-fledged recipe that I can use every week. I can just say, “Here’s the draft, go,” and when I come back, it’ll be done. Because the 5P Framework creates such rigid definitions of success, it allows an agent to iteratively check if it is finished.

Katie Robbert – 30:09

Devil’s advocate question. I love this extra detail, but help me convince someone like John. What is the benefit of doing this versus just taking your transcript and having a copy editor, human or virtual, clean it up so it’s more coherent?

Christopher Penn – 31:12

There are two things. One, there are extra steps here that humans typically don’t do, like the fact check. Your average copy editor is not going to Google every claim and validate that it’s true. Even if I am an expert, I can still be wrong.

Two, I have biases when I write or talk. The Trust Insights newsletter is the company’s publication, not my personal one. A copy editor’s job is not to enforce the Voice of the Customer. Their job is to tell me I have the world’s longest run-on paragraphs. By aligning this with our Voice of the Customer, we refine it to identify what Chris said that is actually valuable to our customer.

Katie Robbert – 32:48

We have an interesting comment from a viewer: “What you just defined is going to use a lot of tokens. It went from being a $50 project to a $150 project.” What do you say to that?

Christopher Penn – 33:12

You have it backwards. In an agentic framework, if you give it a crap prompt, it’s going to spend a lot of time trying to infer what you meant. You burn a lot of tokens for the model trying to figure out what it’s supposed to be doing.

If you have a good, detailed project plan up front, Opus is going to have fewer debates with itself. It will understand the plan and just execute it. Second, if the plan is really well thought out, you can switch down to a lighter model like Sonnet or Gemini Flash to save a ton of money.

Plan big, act small. Use the smartest model for the plan, but then for the typing part, you can use a lighter model as long as the plan is good. A poorly defined prompt makes the model spin more and gives you less repeatable, reliable results. If I’m not specific, I might get three differently formatted columns from the same text. If I specify no listicles and no 28-point bullets, I get a repeatable result, which actually cuts down token usage because the model doesn’t have to wonder what I want.

Katie Robbert – 35:47

That is a great segue into Brian’s question. Do you use Claude to create the agentic project plan and then take that into Claude Code, or do you make the plan directly in Claude Code?

Christopher Penn – 36:03

Personally, I do everything in Claude Code. For the average user who is not technical and doesn’t want to fire up a terminal, I would suggest doing the agentic plan in regular Claude and then executing it in Claude Co-work.

Katie Robbert – 36:18

That’s how I typically work. I’m not savvy in Claude Code, but I have become best friends with Claude Co-work. My brain already works in the 5P Framework. You can create the plan and then give it to the agentic system. You can even make the platform one of the Ps, specifying that it will be run through Claude Co-work or Claude Code.

John, you have been a non-stop chatterbox this entire episode, so I’m curious to get your thoughts.

John Wall – 37:24

There is not a huge dividing line for me between straight prompts and calling agents. I’m still trying to get my head around at which point agents are being invoked and how that is fundamentally different from straight prompting.

Christopher Penn – 37:43

In systems like Google Anti-gravity or Claude Code, an agent is essentially another copy of the AI. When I say I want to spin up an agent team of a UI/UX designer, an SEO expert, and a project manager, you will see little “employees” start popping up on the command line. They each get their own context window and working memory. They can trade messages back and forth.

When you start using agents, the window you are in becomes the orchestration window where you are the project manager. Claude is like the factory floor manager, and the agents are like employees on the factory floor. If you’re just typing into a regular AI, you’re basically talking with one employee as opposed to a team.

John Wall – 39:30

How about having this stuff running 24/7? Do any of the platforms have anything where you would leave these running, or is that not where we’re at?

Christopher Penn – 39:46

You can tell Claude to independently run certain tasks with a scheduler. It requires some knowledge about shell scripting and scheduling on your computer, but you can ask Claude how to do that. For example, I could say to Claude every weekday, “Run a script that builds me my daily briefing from Asana.”

I personally like to ask for it because I typically give extra instructions about things I forgot to put on my list. But there is a way to have your briefing ready every morning at 8:00 AM.

Katie Robbert – 40:52

A good example for you, John, is HubSpot. I could see creating a script so that every morning you have a daily briefing of everyone you need to follow up with and who is ghosting you. Agentic AI is helpful here because it can bring in additional context, like prioritizing that list based on who fits our ICP best.

Christopher Penn – 42:11

Think about what you would hire a small agency to do for you—that’s a use case for agentic AI as opposed to a single contractor. If you wanted more corporate sponsors for Marketing Over Coffee, you would put together a 5P plan. The People would include a sales and revenue agent, a project management agent, and a marketing agent.

The Process would be to research what other mid-six-figure podcasts do and build a strategy and tactical plan. The Performance would initially be a plan that works, and ultimately, being surrounded by Benjamins.

Katie Robbert – 44:14

Chip offers the opinion that while agentic teams have big benefits, they also risk complicating projects without creating commensurate value. I hear you, but that is why we hammer home that people should constrain their projects to the 5P Framework. If you don’t have this plan, don’t start building agents.

AI is going to be as helpful as possible and suggest everyone. You, the human, need to push back and ask for the bare minimum of people needed to fulfill the task. You still have to oversee it and be the person who says, “No, don’t do that.”

Christopher Penn – 45:56

When we talked about this on the podcast, we asked: if you were hiring humans, would you hire a 100-person team to revamp the Marketing Over Coffee website? No. You’d hire a small boutique agency with maybe five or ten people.

In the 5Ps, you don’t have to have every agent on deck all the time. When you’re building a house, the HVAC guys are not there while the scaffolding is going up. If you use the 5Ps, the tools will help you figure that out as part of the plan.

Katie Robbert – 47:02

If it’s suggesting a bunch of synthetic people with roles you’ve never heard of, push back. Ask what they bring to the table. Get that plan squared away first.

Christopher Penn – 47:29

Maybe on next week’s show, we should do a part two and deploy an agentic swarm on the Marketing Over Coffee website to build a plan for John to become filthy rich. We can build a swarm with a business manager, a sales manager, and a researcher. We’ll pull industry research from Tom Webster and Edison Research to see why some shows do better than others.

Katie Robbert – 48:43

I like it. I think that’s it.

Christopher Penn – 48:50

That is it. Please try out the 5P Framework. You will save yourself so much headache if you just get good at project management. That’s going to do it for this episode. Be sure to subscribe to our show and check out the Trust Insights podcast at TrustInsights.ai/tipodcast and our weekly email newsletter at TrustInsights.ai/newsletter. Join our free Analytics for Marketers Slack group at TrustInsights.ai/analyticsformarketers. See you next time.