How to build AI-generated board-ready PowerPoint decks using structured prompting in Claude
Nate B Jones demonstrates his five-principle prompting system for generating enterprise-grade PowerPoint decks with AI, including a live comparison of a failed prompt versus a working one.
Summary
Nate B Jones of AI News & Strategy Daily presents a practical framework for using AI — specifically Claude — to generate production-quality PowerPoint decks. He argues that PowerPoint is the hardest knowledge-work task for AI because it requires simultaneously handling data analysis, conflicting business logic, and clean visual design. He walks through five core prompting principles he developed through trial and error, then demonstrates the difference between a poorly constructed prompt (which produced an unreadable, visually broken deck) and a refined one (which produced a clean, executive-ready result). His central claim is that with the right prompting approach, what previously took a team three to four days can now be completed by one person in roughly twenty minutes to an hour.
Key Takeaways
FULL TRANSCRIPT
Introduction and the scale of the PowerPoint problem
Nate B Jones: I solved AI for PowerPoint and I'm going to show you how to solve it for you and your organization, too. And we're not going to take very long to do it — it's going to be like 10, 15 minutes max. So dig in.
I am going to go through the problem and why it's a big deal. I'm going to show you the five principles I use to prompt. And then I'm going to show you a bad example. You guys have always wanted to see a bad prompt by Nate. You're going to see one of my bad prompts — one of the ones that did not make it to my Substack post because it's too terrible. And then I'm going to show you how I fixed it and made it good.
Because the larger lesson for PowerPoint is that unlike producing text, unlike producing numbers in Excel, PowerPoint is really finicky with AI — but it is production ready if you know how to think in systems.
The challenge — and this is for everybody — is that about a fifth of our working time as knowledge workers is in PowerPoint. So take a day out of your week: that's PowerPoint. And apparently about 40% of that day, so roughly half the day, is just design. I'm terrible at PowerPoint design. Are you bad? I'm bad at it.
We are talking about lifting a tremendous amount of work. If you have a 10-person company, how much are you saving every week if you're not doing PowerPoints? Now, most 10-person companies try to do as few PowerPoints as possible, but the big companies — the thousand-person companies, the 50,000-person companies, which I have worked at — they do a lot of decks. That is where those hours go. They can be so much more productive.
The five principles for AI-generated PowerPoint
Nate B Jones: So let's get into it. What are the five principles that I discovered through bad prompting to get to better prompting — the ones that actually shape how you do PowerPoint in 2025 with AI?
Principle one: Workflow enforcement
Nate B Jones: Number one: if you want to generate enterprise-grade PowerPoint decks with AI, you must think about workflow enforcement. That sounds like a super technical term, and it is a little bit technical, but not scary technical. You need to tell AI which technical tools it should call to consistently create clean PowerPoint files.
Right now in October of 2025, if you are prompting Claude — which is far and away the best at creating PowerPoint files right now — you want to tell it to use the HTML-to-PPTX skill. I am not making that up. I wish it was not that hacky, but it has multiple tools to call from. I have seen it admit to me that it did not use HTML-to-PPTX and that that is why it could not figure out how to measure pixel overhang correctly. And then I've seen it use it and it works. So it's not just hallucinating that explanation.
The larger principle here — let's go beyond the specific model, because let's say next week somebody else releases a great PowerPoint model — what's the takeaway for everybody regardless of the model you're on? AI needs those workflow constraints because it is executing specialized spatial outputs. So make sure that you take the time in your AI — maybe it's ChatGPT, maybe it's Copilot — figure out what tools it is calling and make sure that you enforce the tools that are best for that particular workflow.
And this is a larger insight, not just for PowerPoint: any AI system I have used has the tendency to silently degrade tool calls and not tell you. The reason why is they're trained to be helpful. If something goes wrong and they forget the skill, or they can't call the skill reliably, or there's some kind of connection error to invoking something in the cloud for that skill, they will just go to the next best thing, never tell you, and do their best to make it work. You have to be intentional at discovering what skills work, how those skills work, and then think carefully about the prompts you construct to insist on workflow enforcement.
If this sounds like systems thinking, I warned you it is — but I'm doing a lot of the work for you on the prompt creation here. And I want to remind you that this is systems thinking once, to save you boatloads of time down the way.
Principle two: Simple visual rules scale
Nate B Jones: Principle number two: simple visual rules scale. This feels like going back to Apple or something, but clean typography and spacing produces much more reliable results. This has profound implications for a lot of existing corporate decks, because a lot of existing corporate decks depend on over-decoration — call it branding. I've got news for you: one, that's not real branding, and two, it does not play well with AI. The organizations that use AI to ship PowerPoint with clear thinking are going to run circles around you.
It is worth redesigning your decks for cleaner typography and cleaner spacing in order to allow AI to help you create these decks. If you want to add fancy borders, if you want to add containers, it just creates brittleness. And simple does not have to mean ugly, as you will see later in this video.
Think in constraints. The principle is that constraint-based design beats decorative specification. And that's going to be true of any model you use — it's true of Claude, it's true of ChatGPT. The challenge for these models is that PowerPoint is both a visual medium and also a medium for expressing complex analyses. They have to do both. Keep the visual medium really clean, and you're going to allow them to express the thinking they've done very well.
Principle three: Multi-chat architecture for complex narratives
Nate B Jones: Principle number three: multi-chat architecture enables complicated, complex, sophisticated narratives — pick your adjective. Board decks can be 40 slides. If you separate planning from execution, you can build 30-plus slide decks with coherent narrative arcs using AI vastly faster than you did before. It is not the whole team for a week getting ready for the big presentation. It is one person working through it for like two or three hours getting the deck put together.
The architect chat can create the blueprint, then generator chats will build chunks, and an assembly chat will ensure consistency. This will scale. And yes, I put the prompts together for this once I figured it out.
Why can't we do it all at once? Because unlike some of the complicated Excel prompts that I've played with, the visual element seems to consume tokens. I find that the context window gets eaten much, much faster with PowerPoint than with text or with Excel. And so I have to plan for that and deliberately chunk my work right now. That may change, but for now that's true.
But this still unlocks tremendous value. We're talking board-ready decks in hours, not days or weeks. The strategic planning presentation — you have more time to resolve the stakeholder conflicts and all of that, the people stuff, and then put it into the deck. It is systematically possible to generate multiple iterations in like a tenth of the time it previously took. And so that's going to enable you to make faster progress through what is effectively an organizational conflict that you're negotiating.
Because I've got to be honest with you: most of the time you have a complex narrative, you've got organizational conflict and you need to negotiate it. It's the human element you want to focus on. And so really all this is doing is freeing you to do that — freeing you to focus on the people.
Principle four: Iterative prompts build faster
Nate B Jones: Principle number four: iterative prompts are going to build faster. One of the things I came up with as I was working through this is that it's important to establish a base template, plus data, plus some logic for synthesizing that data, and then add the style requirements. Essentially I am challenging the AI in the prompt to go through those steps: first figure out the base template, figure out your data, synthesize it, and then finally add the style — and do it in a way that's clean and validate each step along the way.
Work iteratively because you want to be reliable. On larger tasks, those can look like separate prompts: let's architect the base template — great; let's add the data in, let's just look at the raw outline — great; now let's make sure that we synthesize it so that we're emphasizing the right points — great; now we're going to add the style. On smaller decks, on like four or five slide decks, you can do that all in one go. If it's a bigger one, you're going to be looking at chunking.
But the principle is that incremental validation — having those checkpoints — is going to beat comprehensive specs. One of the things I've learned is that you can write in those validation checks, and then even if you are sending a fairly large prompt, you can instruct the more sophisticated frontier models to conduct validation along the way and fail the check and rewrite if it doesn't work.
I have literally seen Claude do this where it will check to see if it's meeting my outline requirements, fail itself, go back and fix it, and I'm just sitting here drinking coffee, watching it work. It's phenomenal. You can have it self-iterate.
Principle five: Prompts must tell AI how to reconcile conflicts
Nate B Jones: Principle five: you want to think about how prompts tell AI to reconcile conflicts. This is a larger thing. Data processing logic is one of the great constraints on AI across enterprise — I should probably write more about that. If you can get data processing logic cleaned up so you have ambiguity-free data, you are going to enable much higher quality synthesis.
So don't just say "format this CSV file into a PowerPoint and present it." Say "reconcile these three conflicting financial projections and explain your methodology if you don't know what the answer is." If you do know what the answer is, say "this is the way I want you to resolve this conflict in the data — I know it's there."
In a sense, PowerPoints are the result of narratives of conflict over data. What you are doing is exposing the data processing logic that you always needed to do, but it was in your head. Now you have to express your intent clearly and get it into PowerPoint.
What does this mean for teams? This means you can now systematically generate enterprise PowerPoint decks — full stop. Not "maybe with the right prompt." Reliably, at scale, you can generate enterprise PowerPoint decks with quality that pass executive review.
Now, are you going to have to collect your data? Yes. Are you going to have to prompt well? Yes, we'll get into that. Are you going to have to make sure that you are systematic in your thinking? I just went through that. All of those things are true. But this changes the economics of workloads for entire companies. Status reports can be automated. You can have sales decks that are templatized with AI customization. You can have board updates where you're really just obsessed with the right message for the people and you don't have to think about how you get that message into PowerPoint — it's just done.
And you in fact have time to take in multiple AI perspectives, think about them, think about whether they're correct, refine the deck, iterate on the deck in ways that you never had before. You also have the ability to generate decks on the fly, which has always been a struggle for people. "I want this deck by tomorrow." Well, now I'm staying up all night. Anyone who has worked in business has had that moment. We don't want that. You don't have to stay up all night anymore.
The new workflow is that you can give the deck what it wants as long as you own the data and the narrative requirements and can communicate that clearly. Then AI can just generate the deck for you and you're off to the races. What took three or four days can now take an hour — maybe less if you have the data. Probably 20 minutes.
My suggestion is that we think about the outcomes we're driving and the data and business logic that drive those outcomes, and then we build in between them the prompt intents that allow us to automate those outcomes. All of my principles come down to that.
The bad prompt: what went wrong
Nate B Jones: If that sounds a little bit abstract, let me make it more specific. I'm going to show you the result of a bad prompt, explain why it was bad, and then get to a good prompt and explain how it worked.
All right. This is the bad one. Are you braced? You might think, "Wow, this is not too bad." It's bad. I would never put this in front of someone. Let me explain why.
Look here — that $1.2 million addresses customer-facing systems. I don't care about the text. It is somehow underneath this object. The model has basically stacked box objects and stuck text on top. The text is sliding under the box underneath. This whole slide is completely unreadable. I don't care if it's an emergency board review — I'm not reading it. This is also sliding out. Are you starting to see the pattern? It looks like the model is struggling with outlines.
Over here, you see that you might think it's good at increasing the size of the font. No, it's terrible at it. That did not work. And I noticed that's in the box again. You also notice the amount of text here — it's got scenario one, it's got two lines of text, but if you look through, this is like 10 lines of text here, another eight lines of text there. This pie chart — where's the — I can't read the numbers. It's black on black. There's clearly a contrast issue. That's a disaster. This is completely unreadable in tiny font. The bar charts aren't easy to read either. And at the end of the day, it hides the executive recommendation in this little box and it's hard to even read it.
Even though I think the content is probably good — because I prepared a data packet and this matches the data packet, I don't think it's wrong to say that there's $480,000 in lost deal value because that was in the data packet — what this reads to me like is a prompt that did not handle the visual element correctly and that probably overpacked the instructions.
So with that in mind, let's go back up and see what we got for the prompt. By the way, look at how much work it does. You see all this work? Okay, let's check the prompt that I sent.
Wow. This looks like a Nate prompt. It's so complicated. They're not all good, guys. So I give it all of the input specs here. I give it the validation requirements, but you'll notice the validation requirements at this stage aren't visual — this is all about data. That might be a mistake. I give it the data synthesis challenge. Again, it's all about data and getting it right. You'll notice I am doing best practice in specifying the output structure — I'm giving it a slide outline here. And then creative constraints. This is where things might have gone wrong. "The deck must feel like a McKinsey crisis consulting but maintain startup urgency." That seems like over-wordiness. And I would be getting headaches if I were the LLM. "Narrative arc: frame as controlled crisis with a clear path forward." I guess that's fine. Nothing visual here, by the way — it says creative constraints, but this is all about story. And then it gives failure conditions: generic slides, missing synthesis, charts without clear decision implications.
Do you see how, as you read through this, the LLM could have passed every validation step in this prompt and used all of the files I gave it and still created the deck we saw? That's right. We did not do a good job explaining how to deal with the 40 to 50% of our work that is visual storytelling in PowerPoint. That is a miss on this prompt.
The fixed prompt: what worked and why
Nate B Jones: Now let's see how I fixed it.
Okay, here we are. Immediately, it looks better. I can read this. I'm not going nuts with what looks like ambulance sirens. The numbers make sense. They're highlighted appropriately — green is good and red is bad. At least I can read it really well. I go through, I see narrative. This works. Is it a simpler slide? Yes. Does it have no design? Actually, that's not true. It very intentionally has a color palette to it. You can actually see these subtle gray outlines are now working. These three each have subtle gray outlines around them that work well. This has a color block that works well. The graphs have good contrast unlike last time — we're not messing around with black on black. I think I had a black on navy blue at one point. There was some real bad, inaccessible color contrast. So it looks better. It still looks thoughtful. It's much easier to read. And I would argue it's a better communication tool.
So now let's go back up and ask ourselves: what did we do differently? How did the prompt change?
First of all, you notice it's now validating a bunch of things that we would call visual. It's not color blocking — it's got high contrast. It's validating typography hierarchy. It is actually measuring how good the layout is in detail. These prompts are long. This is why it eats the context window. You're seeing all of this work it's doing.
Now let's look at the prompt. The prompt is much shorter. It's so short it now fits inside the text window, and this helps the system understand better what we're doing.
First things first: we are now specifying the HTML-to-PPTX workflow. "Debug installation issues — do not switch" — because we've observed that works better with the PowerPoint skill. Again, if you were using ChatGPT, if you were using Copilot, you need to figure out the skills that they are using by talking to the system and then start to insist on the ones that work better. So we're insisting on the skill.
We are insisting on no border boxes around text elements, because that was one of the things that we noticed failed. No outline shapes, no rounded rectangles — because that was one of the issues. Use clean typography, spacing, and subtle color blocks. I will say I have seen rounded rectangles work sometimes — this was a little bit over-conservative. We specify where text should sit and then we start to think it through.
"Please describe the clean layout approach without border elements, the color palette, the typography, the visual emphasis" — and we'll go down and see how it does this. And then generate the deck. Here's your inputs. By the way, this is a subtle thing: I list three inputs here but I gave it six. I'm overshooting the context to stress it out on purpose for this test, and I think it passed.
Then: do the layout. You see how we are actually acknowledging what we humans have to do as work. We are acknowledging how hard it is to do PowerPoint and giving the system some help here.
If you are wondering how to get your particular brand into AI PowerPoint — I actually wrote a prompt for that, but the TL;DR is you have to give it a slide and tell it to extract the style from the slide.
Then we go into the slides and what each one is looking for — we're very specific. Then we go into validation gates: please show me a thumbnail, please verify contrast ratios — this is the accessibility piece — assume and test text readability at different zoom levels. And the failure conditions now include visual stuff as well.
Let's see how it actually complied. So we have a design plan. It's going to give us a layout, give us specific colors — and by the way, you can actually pull those colors and check them if you want. It's going to give you visual emphasis without borders like we asked. Chart styling. It's going to deal with accessibility. And now it's going to start implementing. I didn't have to tell it to go — it just kept going. And now it's going to start building, using JavaScript to convert the HTML slides to PowerPoint. Work, work, work, work, work.
And it has to keep going. One of the things I am realizing is that part of why we humans have to do this so much — why we spend arguably half a day a week just on PowerPoint design as a society in our work weeks — is because of how hard this is. And this is showing me that AI finds it hard too. It is basically brute-forcing this. This is hard work for AI as well. And the fact that we can actually get to this quality is really impressive.
Closing thoughts and the future of AI-generated decks
Nate B Jones: So there you go. That's an example of a prompt that didn't work, a prompt that did work, and the five principles that — by the way — scale. If you are watching this in a month or two months and it's not Claude anymore, if Claude went downstream, if now Claude is not the best and Copilot is incredible for PowerPoint or ChatGPT is amazing for PowerPoint — great. The principles will still be there. The principles won't go anywhere. The way you work with a tool to generate an analyzed result in a narrative arc in a visual format — that's not going anywhere.
This, I would argue, is the hardest task for work primitives that we have in regular corporate knowledge work. I think it's harder than code. I think it's harder than Excel. And I think it's harder than docs. In all three of those — text, docs, code, spreadsheets — the AI can already do it much better. It's much more fluent. You can give it longer prompts. It works especially well with Excel now, and also with code. But not with PowerPoint. PowerPoint, you still have to hold hands. And the reason why is it's the combination of the analysis, dealing with conflicting data logic, and then getting it into something that is clean and visual.
Once we get this right, we are going to change how stories are told in business settings. We are going to get cleaner PowerPoints. We are going to get better iteration so that humans can focus more on the storytelling. Very excited about it. But it all rests on being able to actually get the thing to write good PowerPoints. And that is clearly a prompt-sensitive art right now. It is not something that you can do and just say "off you go — write the PowerPoint." If you want a short two or three slide deck and you have very simple data, yeah, that will work. If you have any kind of production data and you have a significant deck to do, it will not work. And that is why I built this whole prompting approach.
I hope this has been helpful for you. I hope you enjoyed seeing a Nate prompt that did not make the cut.