Beyond prompt engineering best practices: how to write AI prompts for agents now

Transcript

Prompt engineering is a 2025 conversation. Prompt engineering is dead. Prompt engineering is not what we need to be talking about now. And I say that and people say, Nate, you told us so much about prompting. You said it would be future-proofed. Yes, absolutely true.

It’s table stakes now. It is table stakes now. You got to be good at it. You don’t get credit. Sorry. Welcome to the way AI moves.

Now you need to be thinking about what comes on top of prompt engineering. And I can see in people’s eyes, I talk with people about prompting. I can see the eye roll in person because people are like, why am I doing this? Prompting is such a 2025 thing. I can just ask the AI and it works and I can get what I want. So what, why does this matter?

I want to tell you that we have had updates in how AI agents work, updates in the impact and power of agentic workflows in the last 2 months, specifically with 4.7 and 5.5. So 4.7 from Opus, 5.5 from OpenAI, where we are now able to have an entirely different kind of interaction with AI than we used to have. It is not the same. And it gets so tired when I say people aren’t talking about this, but it’s so true. People aren’t. People are usually saying prompt engineering is something that we don’t need to talk about because you just ask AI for what you want, which is frankly what I hear from a lot of the folks who are at these labs, just ask AI for what you want.

We trained it so well, you can just ask AI for what you want. That’s only true if people know what they want. And what I find is the more complex the agentic workflow, the more it’s hard for people to ask for what they want. And so this video is actually about helping you to understand how to ask for what you want when you have an agent who is, I kid you not, 100 times more powerful than the agents that you had 6 months, 7 months, 8 months ago. And I say that because look at the way they call tools, look at the way they call data, look at the way they’re able to do work for longer and longer periods of time, and the amount of impact that that opens up to you inside the corporation. They are 100 times more powerful.

We have not evolved our prompting 100 times, right? That’s kind of a problem. So this, this is what I want to introduce for you. I want to give you the AI question method, because I think we need something that isn’t called prompting anymore. I think prompting assumes you ask and you get an answer. We are really in a question world, not a prompt world.

And I want you to just fundamentally reshape that. I’m going to tell you a story about my manager years and years ago that will help you make this, this jump. So if you’ve ever had a good manager, this will feel familiar. And if you never have, you’ll probably feel sad and be like, why haven’t I had a good manager? That’s okay. Your time will come.

AI now is like a senior partner on your team, not like a junior partner. That is the biggest mental model shift that you need to make between last year and this year. Last year, part of why prompt engineering mattered so much is we had to talk with AI like it was a junior partner on the team, and we had to be very specific and careful. This year, that’s not true. This year, you have to talk with AI like a senior partner to get the most out of it.

And so when I was on a marketing team years and years ago, I had a fantastic manager who would do exactly what I’m advising you to do with AI with me. And what she would do is she would say, hey, I need you to dig into this problem set with me. I don’t fully have the answers. This is a collection of CSVs and Excels that I have. I’m trying to solve a marketing attribution I need this to turn into a deck that tells a story to leadership and also a doc so that I can actually write down a narrative that makes sense. It was Amazon.

We definitely are going to have a doc, not just a deck. And I need it to be clear and incisive and tell a real story and not get lost in the individual details. But I’m going to leave a lot of the rest of it to you. And I want to ask you some questions along the way to guide your thinking. I want to ask you to think about— and then she would go into the key things that she would want to ask and think about. Ironically, that same approach now looks like a pretty good agentic prompt in 2026.

It does. And I call it the question method because it invites us to think about the way we partner with AI as a series of questions where our job as the manager with a senior partner is to form questions that enable the agent to do its best work.

Now, pause for definition here. Agent is such an overused word. I have to be very clear about what I mean. I am talking about work in Cowork, in Claude Code, in Codex, heavy, heavy knowledge work with frontier models. This, by the way, also works if you’re planning a coding project and you’re in these tools, but your eventual output is code. Similar dynamic.

I, I am referring to the act of sitting down and doing deep partnership thinking with AI to get real high-leverage work done. This is different than if you use the word agents to refer to a pipeline that, say, takes customer service tickets and deals with them, or that takes invoices and deals with them. Those— that’s not how you interact with it. It may be called agents. It is an agent. It does work with tools to get work done, but it’s a defined workflow.

It doesn’t have heavy knowledge work at the beginning that’s unique and custom. It’s actually supposed to be really buttoned up. And it’s supposed to be very predictable. I call that an agentic pipeline. That is certainly something you can build nowadays that has lots of power behind it. If you want to check out more about how to build that, I talk about how to do that on a lot of my YouTube videos in the last few months.

It’s a big theme in 2026. But that’s not this video. This video is about the heavy knowledge work that we do with agents and how we need to think about a revolution in prompting that I don’t think enough of us are doing intuitively. And when I say that, I’m not just saying vaguely because the internet thinks it. I’m saying because I watch, I get the privilege, right? People talk to me about AI all the time.

They ask me to check their prompting. They ask me to look over their shoulders. They show me prompts. They send me emails with prompts. They ask me questions however they can get to me. And what I see is that most people have not fully transitioned from 2025 to 2026 in prompting terms.

They are still thinking in terms of, I live in a world where AI is something I have to give a task to. And I have to define that task and ask it to do it well, which was very much best practice as late as November of last year, which is not that long ago. It’s like 6 months and not anymore. And in fact, I would argue we have seen a second acceleration since the launch of 4.7 and 5.5 in the last month or so. I, what I am describing to you is sort of the question method or how you work with AI as a senior partner. To me, intuitively, it feels like something that is more doable, particularly with 5.5 in Codex, than it has ever been with any other model before.

And so your mileage will vary. If you were on a free account, if you were on a paid account and run out of tokens quickly, if you are using an older model, this will not work for you. But if you are doing heavy knowledge work on a cutting-edge model, then this is how you need to do it. And this is where you need to evolve. You need to first and foremost change the mental model so that AI is a senior partner, and then you need to learn the art of asking questions that open up the scope of the problem without being too open-ended. And I want to talk about that a little bit.

If you want to dive into that more, I have a whole series of prompts. I have a deep dive on the Substack on how to think about questions and changing your mindset to a question approach, how you talk to your AI about this and say, I want to start to ask more questions because it has memory now. It’s going to remember that, right? So you can get all of the quick start guides and everything on the Substack for that. But, but here for you to take away today, what I want to give you is 3 key principles, and we’ll go through each in a little bit of detail that help you understand how to ask those kinds of questions because I can’t assume you’ve had a good manager. Not everybody has.

In fact, I talked to a lot of people who are like, wow, you’re comparing this to a good manager? I’ve never had one, Nate. Good luck with that. Great. Okay, we’re going to go through this, and this will have the benefit of teaching you what good management looks like at the same time as it helps you learn to communicate with AI, which is one of the larger themes we’re seeing in 2026, is that the communication patterns we use with AI happen to help us with our people and our teams as well. So, you’re trying to ask questions that open up the problem.

Principle number 1, a good manager will understand how to shape the questions to get you thinking about where all you could go to discover their perspective in more detail. And I— we’re going to unpack that. But the first thing I want to call to your attention is that you need to be clear about your perspective and the questions that you’re asking. And so if you say, help me learn more about the Mona Lisa, you are not giving your perspective. If you say, I want you to learn more about the Mona Lisa from the perspective of da Vinci’s later life and how the Mona Lisa shaped his relationships with his peers, because I have a thesis that the painting of the Mona Lisa actually shaped da Vinci’s relationships late in life. Now you’re conveying your angle, right?

And I gave a painting example because the Mona Lisa is a known painting, but you can do that with business problems too, right? You can say, I have a thesis that our marketing attribution is broken because we don’t have our Google organic bucketed out correctly and we need to, right? And therefore, I want you to start to investigate the data with that in mind. I might be right, I might be wrong, but that’s my thesis. And so a good manager conveys that directionality of opinion sort of like the center of a beam of light from a flashlight. And there’s a wide range around it that you explore, but you kind of know where their eyes are and you know where they’re looking.

Your questions to AI need to have that kind of intent behind them so that you can be clear, this is where we’re going. You have room to explore and partner with me in that space, but I want to give you a center of the I want to give you a sense of the bounds or the edges to go after. I’ll give you an example of how you ask questions that also illustrate the boundaries and the edges. Let’s say you’re talking about meeting notes, and what you want to do is take meeting notes, combine them with some files, and start to put together a document that’s a report. What you want to do is you want to convey both what you want as the central message of that, and that’s something that’s like the center of the flashlight and the questions you’re asking. And then you want to say also, hey, we spent 15 minutes of the meeting talking about something entirely different.

It was a project we were greenlighting, and we don’t need to have any of that included in the report. Please excise that and drop it out. And then you ask the AI to factor that in, make sure that it does not include this, this subtracted portion of the context window, and that it does fully consider and wrestle with the larger file structure that you’re giving it. So it understands what to put in the report. Again, your job is to convey in your question set where your focus is and give the AI space to explore and also some hard edges where needed. So that’s the first principle.

Your questions need to have that center of the flashlight intent behind them. Otherwise, you, whether you’re working with humans or agents, you’re not giving them what they need. That’s, that, and, and that’s something that I see people miss a lot. They will ask overly open-ended questions, they will ask overly closed-ended questions. Having the art to have a center to the flashlight and also this sort of larger edge to the light, that’s really important. And it is an art, and I have lots of examples on the Substack of what that looks like and how you can ratchet through that, so you can dive into that if you Principle number 2, ask the AI questions that invite the AI to consider what good looks like in an outcome in a more open-ended way than just by writing an eval.

I love talking about evals. I talk about them a lot. They’re really important in the kind of agentic pipelines I discussed earlier in this video. You have to have them to make sure you get work done. But increasingly, I think that people are mistaking the idea that evals are great and useful in many cases for the idea that you can’t ask the AI to explore what good looks like for an outcome that you are targeting with you. I’ll give you an example here.

Let’s say you’re writing a PR FAQ with your AI. Amazon document, press release first, FAQ next. You know what a good PRFAQ looks like if you’ve read one, but it is hard to write an eval that captures that. It’s very difficult. If you were to say, I want to work with you to write a PRFAQ, to your AI assistant, the best way perhaps to get that tough document standard met is to ask it questions that force the AI to contend with the kinds of outcomes you’re looking for and to synthesize an answer that meets them. And so you could say this PR FAQ is about, you know, a new launch that we’re going to have at Prime Video.

It’s going to be absolutely— I’m making all of this up, by the way. It’s going to be absolutely incredible. And we’re going to have sort of 3-dimensional sports figures for the World Cup, and they’re going to be on your living room floor, and they’re going to be kicking the ball around, and it’s going to be amazing. And your job in the PR FAQ is to do the following. I need you to think with me about how this customer experience is something that is accessible whether you’ve had a 3D experience or not, whether you’ve worn 3D glasses or not. I’m not sure how to convey that, but I want you to think about it from the customer’s perspective.

And can you think about how to weave that into the press release? And then you say, well, we’re not done yet. We have another hard question. I want you to think about how you convey the interrelationship between the software experience and the hardware experience, both in the press release and in the internal FAQs, because you need to both convey the emotion of it for the customer in the press release and also in the internal FAQs. You have to really list out how this comes together and how this is seamless for the customer. Did you see how in this made-up example I called out multiple different difficult question types?

I did not say this is how I want it woven together because maybe I don’t know. I’m trying to think through it with the AI. Instead, I ask the AI to look at a set of files and wrestle with me on it and think it through and come back. That is so much more useful now than it was 2 months ago, 3 months ago, let alone 6 months ago. And when I look over people’s shoulders, I’m not seeing people doing it. I’m not seeing people actually say, ah, let me ask multiple open-ended questions and let the AI synthesize across.

And if you’re like, I don’t know how to ask those kinds of questions, Nate, how did you learn to do that? One, I’m going to put together sort of a question prompter guide starter for you so you can get a quick start pack to start to learn to move your muscles that way. You can even have a conversation with AI where you start to learn to ask these questions. I’ll give you a prompt for that. It’s totally possible. But the other thing I would call out is part of how I learned to do this is by leaning into my curiosity and that, that mental model that AI is a senior partner.

And so I talked to it a little bit like AI is a senior partner. I think that’s something I’d love for you to take away from this video is that that is something that has changed. And that we need to start thinking about. Okay, third thing, last thing I want to call out in this video as like the top takeaways for you, and you can always go and grab more from the link below the video when you want it.

Okay, principle number 3 for asking questions. We’ve talked about the idea that you want to ask questions that help the AI explore both the focus of your intent and the edges of it. We’ve talked about asking questions that enable the AI to synthesize across multiple different complex question directions. Third principle, you want to ask AI questions that enable it to wrestle with both the data and file inputs you give it, as well as these softer and more implicit inputs that you may have as opinions that lay across that data. And that’s, again, it’s not an easy skill, but good managers do that. Good managers convey both the files and the hard data you have to wrestle with, and also their thesis or their And so when I am looking at Codex right now, one of the things I love to do is I love to take Codex and I love to have it reorganize and copy over files into a working context folder.

And I love that because I can get formal files like doc files or PowerPoints or Excel or code files, whatever I want. And then I can also have informal files like meeting transcripts and they’re all in one place. It’s very clean, easy place where I can point at the folder and work. And Codex has enough context window that it’s very easy for it to go and grep the files, grab the files and stick them into the folder for me, and then start to work on it from there. It’s not bad at all. But in that world, if you imagine the structure that the AI is working with as a folder with a bunch of files of varying different types in it, your job in the prompt is to make sure that you are conveying enough in your questions that the AI has good reason and good angles to dig in across your point of view framed as a question on all of those files.

So it doesn’t just pick one. And I think, find that that’s one of the things that people tend to mess up is they tend to look at an AI with a bunch of files and they assume it will universally like dig into each of them clearly. Sometimes it does. Sometimes you will see that your question inadvertently angled the AI into a particular folder or file and just dove really deep. And you may not have wanted that. You may have wanted to look across all the files in the folder instead, or everything in the context window if you attach those files.

Cowork likes you to attach those files. And so what I do is I sit there and I say, okay, I need to ask a question that invites the AI to look across those files and dig in with my opinion in the back. And so let me give you an example here.

Let’s say I’m trying to understand a particular MRR problem from a product management perspective. And so my job is to increase the monthly recurring revenue of the product. And I have 2 or 3 product levers to do that with. And I know the previous experiments. And I have lots of voice of customer. First task, get Codex to organize the voice of customer transcripts.

I get it to organize all of the existing customer support ticket copies. I get it to organize everything I know about each of those products, maybe previous PRDs, launch announcements, whatever it may be. And then I make sure that I have it understand the analytics and the quantitative data that helps me understand my MRR position. I get that all into one folder. And that can be an organizational task that you can also give to Cowork. It can sort of work in folders and reorganize files to some extent too.

So I do that. Then I come to the AI and I don’t just say, here’s the center of my intent and here’s the edges. I don’t just say, here’s some complicated questions to synthesize. I specifically ask questions that get at that breadth of data and make sure that it knows my opinions across the breadth so it engages with all of it. And so I’ll say something like this. I’ll say, you know, my thought is that the product-led growth angle that we have been following for the last 2 years is broken.

I don’t think we’re getting good margin out of our product-led growth. I think that shows up in flatter MRR growth in the last 6 months. I think it shows up in somewhat dead cat bounce launches over the last 3 product launches we’ve done. I don’t think that they’ve launched the way we want them to. And I think it’s shown up in our own thinking in the PRDs. You see how I’m naming data artifacts in this question?

And it is a question. I want you, AI, to look into the data with that perspective in mind. Look across all of the data sources I’ve given you, and please come back with your thesis on how you would engage with this problem space. You don’t have to agree with me. Yeah. You can come back and say, no, I think product-led growth is not the right angle here.

But whatever angle you come back with, it needs to be the cleanest, most elegant, most explanatory thesis that you can find on all of the data that I have given to you so that we can partner together to think about what is the next product to launch and why. That is a long question. It took me a while to speak and say. But you notice it also challenges the AI both with an intent and also with the need to get into the data and really understand it and, and understand all of the pieces of the data and not just some of them. I don’t see that happening enough in prompting. I see too many cases where the prompt comes back and it’s like either very flat and it just lists the data and it doesn’t tell you like any sense of intent or question, or if it gives your opinion, it is just a wild rambling opinion where the goal of the AI is probably going to be to mirror the opinion back rather than to invite the AI to examine the data and push back as a senior partner would do.

And so those are my three tips for you: think about that flashlight and how you convey intent. Think about how you ask a series of questions that invite the AI to generalize and synthesize across a very complex space. That was my PR FAQ example with the 3D soccer players for, for the fictional Prime Video launch. And then think about how you invoke the data and how you convey intent in your questions. If I am laddering up some of the things I am intuitively doing that I don’t see other folks I’m doing actually asking when they work with AI on this intent knowledge work, those are the 3. And you can absolutely get better at that.

That’s what the Substack is for. You can dig into that. I have a quick start guide. I have work you can do to work with your AI so it will remind you when you are not asking these questions. Yes, you can actually prime AI. It has memory now and it will ask you, hey, you’re not asking me, you know, thoughtful questions.

Can we work on that together? You can do that. I’ll give you, I’ll give you a note for your memory that you can use.

But regardless, if you step back, I want to put this down as a marker in the sand. We have left behind in so many ways the world of prompt engineering. The eye roll that you feel and see that I see in person is real. And it’s not because prompt engineering didn’t have its place or that it doesn’t still matter. It’s table stakes now, right? You don’t get credit for it.

We need to find a new word for the future of prompting. And if you’re doing heavy, intense knowledge work with agents that are 100x more powerful, you need to treat them as senior partners. And you need to move from prompting the way we prompted junior partners in 2025 to prompting the way you prompt senior partners. And so I hope that these 3 principles, one, help you know what a good manager looks like, and 2, help you think about prompting the AI differently. Prompt it like a senior manager. Use an AI question method instead of just prompt engineering to get that work done.

The words were never the things that mattered the most in prompt engineering. The, the intent was always what mattered. The intent now is best expressed as a really, really sharp series of questions that help the AI explore the problem space and do really meaningful creative work with you. So take that away with you. Take this transcript, chat with AI about it, see if that sparks something, and have fun. I’ll see you next time.