Learning Library

← Back to Library

AI Prompting for PowerPoint Mastery

Key Points

  • The speaker outlines a quick 10‑15‑minute method for using AI to create enterprise‑grade PowerPoint decks, emphasizing that the process is repeatable for any organization.
  • They introduce five core prompting principles discovered through trial‑and‑error, starting with “workflow enforcement,” which requires explicitly telling the AI which tools (e.g., Claude’s HTML‑to‑PPTX skill) to use for reliable slide generation.
  • A contrast is shown between a deliberately bad prompt (shared but omitted) and its corrected version, illustrating how precise, system‑oriented prompts eliminate hallucinations and produce clean, design‑ready files.
  • The speaker quantifies the productivity boost: roughly 1/5 of a knowledge worker’s time is spent in PowerPoint, with about half of that on design, so automating decks can save significant hours, especially in large companies that produce many presentations.
  • The overall lesson is that AI‑driven PowerPoint creation is “production ready” in 2025 when users think in terms of constrained workflows and tool selection rather than treating it like generic text generation.

Sections

Full Transcript

# AI Prompting for PowerPoint Mastery **Source:** [https://www.youtube.com/watch?v=S-g778Djaas](https://www.youtube.com/watch?v=S-g778Djaas) **Duration:** 00:21:54 ## Summary - The speaker outlines a quick 10‑15‑minute method for using AI to create enterprise‑grade PowerPoint decks, emphasizing that the process is repeatable for any organization. - They introduce five core prompting principles discovered through trial‑and‑error, starting with “workflow enforcement,” which requires explicitly telling the AI which tools (e.g., Claude’s HTML‑to‑PPTX skill) to use for reliable slide generation. - A contrast is shown between a deliberately bad prompt (shared but omitted) and its corrected version, illustrating how precise, system‑oriented prompts eliminate hallucinations and produce clean, design‑ready files. - The speaker quantifies the productivity boost: roughly 1/5 of a knowledge worker’s time is spent in PowerPoint, with about half of that on design, so automating decks can save significant hours, especially in large companies that produce many presentations. - The overall lesson is that AI‑driven PowerPoint creation is “production ready” in 2025 when users think in terms of constrained workflows and tool selection rather than treating it like generic text generation. ## Sections - [00:00:00](https://www.youtube.com/watch?v=S-g778Djaas&t=0s) **Fast AI PowerPoint Playbook** - The speaker delivers a 10‑15‑minute guide showing why PowerPoint consumes a huge share of knowledge‑worker time, outlines five prompting principles, demonstrates a terrible prompt and its fix, and explains how to quickly implement AI‑driven slide creation across an organization. - [00:05:13](https://www.youtube.com/watch?v=S-g778Djaas&t=313s) **Constraint-Based AI Deck Design** - The speaker explains that using constraint‑based prompts and a multi‑chat architecture lets AI models efficiently create coherent, high‑quality PowerPoint decks by separating planning, generation, and assembly to overcome token limits. - [00:09:07](https://www.youtube.com/watch?v=S-g778Djaas&t=547s) **Resolving Data Conflicts via Prompts** - The speaker explains that by using precise prompts to have AI reconcile contradictory data and expose clean processing logic, teams can reliably generate executive‑grade PowerPoint decks at scale, provided they collect unambiguous data and adopt systematic thinking. - [00:12:35](https://www.youtube.com/watch?v=S-g778Djaas&t=755s) **Model Fails at Visual Formatting** - The speaker critiques an AI‑generated report, highlighting unreadable fonts, poor contrast, confusing charts, and misplaced recommendations, indicating that the model struggled with visual layout despite correctly handling the underlying data. - [00:15:41](https://www.youtube.com/watch?v=S-g778Djaas&t=941s) **Evaluating Slide Accessibility Improvements** - The speaker reviews how refined color palettes, contrast, and typographic hierarchy improve slide readability and explains that shorter prompts allow more effective visual validation of the layout. - [00:18:53](https://www.youtube.com/watch?v=S-g778Djaas&t=1133s) **AI-Driven Slide Accessibility Validation** - The speaker describes using AI to generate and validate slide designs, checking contrast ratios and readability at various zoom levels, and converting HTML slides to PowerPoint, while noting the significant effort required for both humans and AI. ## Full Transcript
0:00I solved AI for PowerPoint and I'm going 0:02to show you how to solve it for you and 0:04your organization, too. And we're not 0:05going to take very long to do it. It's 0:06going to be like 10, 15 minutes max. So, 0:08dig in. I am going to go through the 0:10problem and why it's a big deal. I'm 0:12going to show you the principles I use 0:14to prompt. Well, five of them. And then 0:16I'm going to show you a bad example. You 0:18guys have always wanted to see a bad 0:20prompt by Nate. You're going to see one 0:21of my bad prompts, one of the ones that 0:23did not make it to my Substack post, 0:25right? Like, that's too terrible. We're 0:26not going to do it. And then I'm going 0:27to show you how I fixed it. I'm going to 0:29show you how I fixed it and made it 0:30good. Because the larger lesson for 0:32PowerPoint is that unlike producing 0:34text, unlike producing numbers in Excel, 0:37now PowerPoint is really finicky with 0:40AI, but it is production ready if you 0:44know how to think in systems. And so the 0:46challenge, and this is for everybody, 0:48right? I looked it up about 1ifth of our 0:51working time as knowledge workers is in 0:54PowerPoint. So, you know, take a day out 0:57of your week. That's PowerPoint. And 0:58apparently about 40% of that day, so 1:01half the day is just design. I'm 1:04terrible at PowerPoint design. Are you 1:05bad? I'm bad at it. And so, we are 1:07talking about lifting a tremendous 1:10amount of work. If you have a 10-person 1:14company, how much are you saving every 1:16week if you're not doing PowerPoints? 1:18Now I will tell you most 10person 1:20companies try to do as few powerpoints 1:22as possible but the big companies the 1:24thousand person companies the 50,000 1:26person companies which I have worked at 1:28they do a lot of decks that is where 1:30those hours go they can be so much more 1:33productive so let's get into it what are 1:35the five principles that I discovered 1:38through bad prompting to get to better 1:40prompting that actually shape how you do 1:44PowerPoint in 2025 with AI. Okay, number 1:48one, if you want to generate 1:50enterprisegrade PowerPoint decks with 1:51AI, you must think about workflow 1:56enforcement. And that sounds like a 1:58super technical term, and it is a little 2:00bit technical, but not scary technical. 2:02You need to tell AI which technical 2:06tools it should call to consistently 2:08create clean PowerPoint files. Right now 2:11in October of 2025, if you are prompting 2:14Claude, which is far and away the best 2:17at creating PowerPoint files right now, 2:19you want to tell it to use the HTML to 2:23PPTX skill. I am not making that up. I 2:26wish it was not that hacky, but it has 2:29multiple tools to call from. I have seen 2:32it admit to me that it did not use HTML 2:362 PPTX and that that is why it could not 2:40figure out how to measure pixel pixel 2:43overhang correctly and then I've seen it 2:45use it and it works. So it's not just 2:47hallucinating that explanation. So the 2:49larger principle here like let's go by 2:51beyond like let's say next week somebody 2:53else releases a great PowerPoint model. 2:54What's the takeaway for for everybody 2:57regardless of the model you're on? AI 2:59needs those workflow constraints because 3:01it is executing specialized spatial 3:04outputs. So make sure that you take the 3:07time in your AI, maybe it's chat GPT, 3:09maybe it's copilot, figure out what 3:12tools it is calling and make sure that 3:15you enforce the tools that are best for 3:18that particular workflow because and 3:21this is a larger insight. This is not 3:23just for PowerPoint. Any AI system I 3:25have used has the tendency to silently 3:29degrade tool calls and not tell you. And 3:32the reason why is they're trained to be 3:34helpful. And if something goes wrong and 3:36they forget the skill or they can't call 3:38the skill reliably or there's some kind 3:40of connection error to invoking 3:42something in the cloud for that skill, 3:44they will just go to the next best 3:46thing, never tell you, and do their best 3:48to make it work. You have to be 3:51intentional at discovering what skills 3:54work, how those skills work, and then 3:57think carefully about the prompts you 3:59construct to insist on workflow 4:01enforcement. If this sounds like systems 4:04thinking, I warned you it is, but I'm 4:07doing a lot of the work for you on the 4:08prompt creation here. And I want to 4:11remind you that this is systems thinking 4:13once to save you boatloads of time down 4:16the way. Principle number two, simple 4:19visual rules scale. This is I feel like 4:22this is going back to Apple or 4:23something, but clean typography and 4:25spacing produces much more reliable 4:28results. This has profound implications 4:31for a lot of existing corporate decks 4:34because a lot of existing corporate 4:35decks depend on over decoration and call 4:38it branding. I've got news for you. One, 4:41that's not real branding. And two, it 4:44does not play well with AI. And the 4:45organizations that use AI to ship 4:47PowerPoint with clear thinking are going 4:49to run circles around you. It is worth 4:51redesigning your decks for cleaner 4:53typography and cleaner spacing in order 4:56to allow AI to help you create these 4:59decks. If you want to add fancy borders, 5:01if you want to add containers, it just 5:03creates brittleleness. It just creates 5:05brittleleness. And simple does not have 5:07to mean ugly as you will see later in 5:09this video. And so think in constraints. 5:11The principle is that constraintbased 5:13design beats decorative specification. 5:15And that's going to be true of any model 5:17you use. It's true of claude. It's true 5:18of Chad GPT. The the challenge for these 5:21models is that PowerPoint is both a 5:24visual medium and also a medium for 5:26expressing complex analyses. They have 5:29to do both. Keep the visual medium 5:31really clean, and you're going to allow 5:33them to express the thinking they've 5:35done very well. Principle number three, 5:37multi- chat architecture enables 5:40complicated narratives and complex 5:42narratives, sophisticated narratives. 5:44Pick your adjective. Board decks can be 5:4640 slides, right? If you separate 5:49planning from execution, you can build 5:5130 plus slide decks with coherent 5:54narrative arcs using AI vastly faster 5:58than you did before. It is not the whole 5:59team for a week getting ready for the 6:01big presentation. It is one person 6:03working through it for like two or three 6:05hours getting the deck put together. The 6:08architect chat can create the blueprint 6:10and then generator chats will build 6:12chunks and an assembly chat will ensure 6:14consistency. This will scale. And yes, I 6:17put the prompts together for this once I 6:18figured it out. Why can't we do it all 6:20at once? Because unlike some of the 6:23complicated Excel prompts that I've 6:25played with, the visual element seems to 6:28consume tokens. I find that the context 6:31window gets eaten much much faster with 6:33PowerPoint than with text or with Excel. 6:36And so I have to plan for that and I 6:38have to deliberately chunk my work right 6:40now. That may change, but for now that's 6:42true. But this still unlocks tremendous 6:44value. We're talking board ready decks 6:47in hours, not days or weeks. And the 6:49strategic planning presentation, you 6:51have more time to resolve the 6:52stakeholder conflicts and all of that 6:54and the people stuff and then put it 6:56into the deck. it is systematically 6:59possible to generate multiple iterations 7:02in like a tenth of the time it 7:04previously took. And so that's going to 7:05enable you to make faster progress 7:07through what is effectively an 7:09organizational conflict that you're 7:11negotiating. Because I got to be honest 7:13with you, most of the time you have a 7:14complex narrative, you've got 7:16organizational conflict and you need to 7:17negotiate it. It's the human element you 7:19want to focus on. And so really all this 7:21is doing is it's freeing you to do that. 7:23It's freeing you to focus on the people. 7:25Principle number four, iterative prompts 7:27are going to build faster. And so I one 7:30of the things I came up with as I was 7:32kind of working through this is that 7:33it's important to be able to establish a 7:35base template plus data plus some logic 7:39for synthesizing that data and then add 7:41the style requirements. Essentially I am 7:44challenging the AI in the prompt to go 7:47through those steps. So, first figure 7:49out the base template, figure out your 7:51data, synthesize it, and then finally 7:53add the style and do it in a way that's 7:55clean and validate each step along the 7:57way. Iteratively work because you want 8:00to be reliable. And so on larger tasks, 8:03those can look like separate prompts. 8:05Let's architect the base template. 8:07Great. Let's add the data in. Let's just 8:09look at the raw outline. Great. Now, 8:12let's make sure that we synthesize it so 8:14that the we're emphasizing the right 8:15points. Great. Now, we're going to add 8:17the style. Now, on smaller decks, on 8:19like four or five slide decks, you can 8:20do that all in one go. If it's a bigger 8:22one, you're going to be looking at 8:23chunking. But the principle is that 8:25incremental validation, having those 8:27checkpoints is going to beat 8:29comprehensive specs. And so, one of the 8:31things I've learned is that you can 8:33write in those validation checks and 8:35then even if you are sending a fairly 8:39large prompt, you can instruct the more 8:42sophisticated frontier models now to 8:44conduct validation along the way. and 8:47fail the check and rewrite if it doesn't 8:50work. I have literally seen Claw do this 8:52where it will check to see if it's 8:54meeting my outline requirements, fail 8:56itself, go back and fix it, and I'm just 8:59sitting here like, you know, drinking 9:00coffee, watching it work. It's 9:02phenomenal. You can have it selfiterate. 9:04Principle five, you want to think about 9:07how prompts tell AI to reconcile 9:09conflicts. This is a larger thing. Data 9:11processing logic is one of the great 9:13constraints on AI across enterprise. I 9:15should probably write more about that. 9:17If you can get data processing logic 9:19cleaned up so you have ambiguityfree 9:22data, you are going to enable much 9:24higher quality synthesis. So don't just 9:27say format this C CSV file into a 9:30PowerPoint and present it. Say reconcile 9:33these three conflicting financial 9:35projections and explain your methodology 9:37if you don't know what the answer is. If 9:39you do know what the answer is, say this 9:42is the way I want you to resolve this 9:43conflict in the data. I know it's there. 9:45And so in a sense, powerpoints are the 9:48result of narratives of conflict over 9:51data. And what you are doing is you are 9:53exposing the data processing logic that 9:56you always needed to do, but it was in 9:58your head. And now you have to express 10:00your intent clearly and get it into 10:02PowerPoint. So what does this mean for 10:04teams? This means you can now 10:05systematically generate enterprise 10:07PowerPoint decks. Full stop. Not maybe 10:09with the right prompt. So it's reliably 10:11at scale you can generate enterprise 10:13PowerPoint decks with quality that pass 10:16executive review. Now are you going to 10:18have to collect your data? Yes. Are you 10:19going to have to prompt well? Yes, we'll 10:21get into that. Are you going to have to 10:23make sure that you are systematic in 10:24your thinking? I just went through that. 10:26All of those things are true. But this 10:28changes the economics of workloads for 10:31entire companies. Status reports can be 10:34automated. You can have sales decks that 10:36are templ templatized with AI 10:38customization. You can have board 10:39updates that really you're just obsessed 10:42with the right message for the people 10:43and you don't have to think about how 10:46you get that message into PowerPoint. 10:47It's just done. And you in fact have 10:49time to take in multiple AI 10:52perspectives. Think about them. Think if 10:54they're correct, refine the deck, 10:55iterate on the deck in ways that you 10:57never had before. And you also have the 10:59ability to generate decks on the fly, 11:00which has always been a struggle for 11:02people. I want this deck by tomorrow. 11:04Well, now I'm staying up all night. 11:05Right? Anyone who has worked in business 11:07has had that moment. We don't want that. 11:08You don't have to stay up all night 11:09anymore. And so the new workflow is that 11:12you can give the deck what it wants as 11:15long as you own the data and the 11:17narrative requirements and can 11:19communicate that clearly. And then AI 11:21can just generate the deck for you and 11:23you're off to the races. And so what 11:24took three or four days can now take an 11:26hour. It's that fast, maybe less if you 11:29have the data. Probably 20 minutes. And 11:31so my suggestion for you is that we 11:34think about the outcomes we're driving 11:37and the data and business logic that 11:39drive those outcomes and then we build 11:42in between them the prompt intents that 11:46allow us to automate those outcomes. All 11:48of my principles come down to that. And 11:50if that sounds a little bit abstract, 11:52let me make it more specific. I'm going 11:54to show you now a bad I'm going to show 11:56you what I promised the result of a bad 11:59prompt. And I'm going to explain why it 12:00was bad. And then I'm going to get to a 12:01good prompt and explain how it worked 12:03for a good power. All righty. This is 12:05the bad one. Are you braced? You might 12:07think, "Wow, this is not too bad." It's 12:09bad. I would never put this in front of 12:11someone. Let me explain why. I want you 12:13to look here. That 1.2 million addresses 12:15customerf facing systems. I don't care 12:17about the text. It is somehow underneath 12:19this object. It's basically stacked 12:22these box objects stuck text on top. The 12:25text is sliding under the box 12:26underneath. This whole slide is 12:28completely unreadable. I don't care if 12:30it's an emergency board review. I'm not 12:32reading it. This is also sliding out. 12:34Are you starting to see the pattern? It 12:35looks like the model is struggling with 12:37outlines. I thought that, too. We'll get 12:39into that. Over here, you see that you 12:41you think that it might be good at sort 12:43of increasing the size of the font. No, 12:45it's terrible at it. That did not work. 12:47And I looked at that and I was like, 12:48I've seen it do that well. I wonder why 12:50it's really struggling. I noticed that's 12:51in the box. Again, you also notice the 12:54amount of text here. Look at this. It's 12:56got scenario one. It's got two lines of 12:57text. If you look through, this is like 12:5910 lines of text here. This is another 13:02eight lines of text. This pie chart, 13:05where's the I can't read the numbers. 13:06It's black on black. There's clearly a 13:08contrast issue here. Like, that's a 13:10disaster. This is completely unreadable 13:13in tiny font. The bar charts aren't easy 13:16to read either. And at the end of the 13:19day, like it it hides the executive 13:20recommendation in this little box and 13:22it's hard to even read it. And so even 13:24though I think as I look at it that the 13:26content is probably good because I 13:28prepared a data packet. This matches the 13:30data packet, I don't think it's wrong to 13:32say that there's $480,000 in loss deal 13:35value because that was in the data 13:36packet. I think it's getting that right. 13:38What this reads to me like is a prompt 13:41that did not handle the visual element 13:44correctly and that probably overpacked 13:47the instructions. So with that in mind, 13:49let's go back up and see what we got for 13:52the prompt. By the way, look at how much 13:53work it does. You see all this work? 13:55Look at all this work it's doing. Okay, 13:58let's go back up. Let's check the prompt 14:00that I sent. Okay. Wow. This looks like 14:02a Nate prompt. It's so complicated. 14:04They're not all good, guys. So, I give 14:06it all of the input specs here. I give 14:09it the validation requirements, but 14:11you'll notice the validation 14:12requirements at this stage aren't 14:14visual. This is all about data. That 14:16might be a mistake. I give it the data 14:18synthesis challenge. Again, it's all 14:20about data and getting it right. You'll 14:22notice I am doing best practice in 14:23specifying this output structure. I'm 14:26giving it a slide outline here. And then 14:28creative constraints. So this is where 14:30things might have gone wrong. The deck 14:32must feel like a McKenzie crisis 14:34consulting but maintain startup urgency. 14:36That seems like overworthiness. And I'm 14:38I I would be getting headaches if I were 14:40the LLM. Uh narrative arc frame as 14:44controlled crisis with a clear path 14:46forward. I guess that's fine. Nothing 14:48visual here, by the way. It says 14:50creative constraints, but this is all 14:51about story. You'll notice that. And 14:53then it gives you failure conditions, 14:55generic slides, missing synthesis, 14:57charts without clear decision 14:59implications. Do you see how as you read 15:02through this, the LLM could have passed 15:06every validation step in this prompt and 15:10used all of the files I gave it above 15:12and still created the deck we saw. 15:15That's right. We did not do a good job 15:17explaining how to deal with the 40 50% 15:20of our work that is visual storytelling 15:22in PowerPoint. That is a miss on this 15:24prompt. Now, let's see how I fixed it. 15:26Okay, here we are. Immediately, it looks 15:28better. I can read this. I'm not going 15:31nuts with what looks like ambulance 15:32sirens. The numbers make sense. They're 15:35highlighted appropriately. I guess green 15:37is good and red is bad. I don't know. 15:38But at least I can read it really well. 15:41Uh, and I go through, I see narrative. 15:43This works. Is it a simpler slide? Yes. 15:46Does it have no design? Actually, that's 15:48not true. It very intentionally has a 15:50color palette to it. You can actually 15:52see these subtle gray outlines are now 15:54working. Uh so you I don't know if you 15:56can see, but like these three each have 15:58subtle gray outlines around them that 15:59work well. Uh this has a color block 16:01that works well. Also, the graphs have 16:04good contrast unlike last time. We're 16:06not messing around with like black on 16:08black. Um, and I think I had a black on 16:12navy blue at one point. Like there was 16:14some real bad inaccessible color 16:17contrast. So it looks better. It still 16:20looks thoughtful. It's much easier to 16:23read. And actually, I would argue it's a 16:25better communication tool. So now let's 16:28go back up and ask ourselves, what did 16:30we do differently? How did the prompt 16:32change? First of all, you notice it's 16:34now validating a bunch of things that we 16:36would call visual. So it's not color 16:38blocking. It's got high contrast. It's 16:40validating typography hierarchy. I bet 16:42you can guess what's going to happen. 16:44It's going, "Oh, I love this. 16:48I love this because it shows us how good 16:52it is at testing. It is actually 16:54measuring how good the layout is in 16:58detail." So, it keeps doing that. These 17:00prompts are long. This is why it eats 17:02the context window. You're seeing all of 17:03this work it's doing. Go back up now. 17:06Let's look at the prompt. So the prompt 17:09is much shorter. It's so short it now 17:11fits inside the text window and this 17:14helps the system to understand better 17:16what we're doing. So first things first, 17:18we are now specifying the HTML 2 PPTX 17:21workflow. Debug installation issues do 17:23not switch because we've observed that 17:25works better with the PowerPoint skill. 17:28Again, if you were using chat GPT, if 17:31you were using Copilot, you need to 17:32figure out the skills that they are 17:34using by talking to the system and then 17:36start to insist on the ones that work 17:38better. So, we're insisting on the 17:40skill. We are insisting on no border 17:42boxes around text elements because that 17:44was one of the things that we noticed 17:46failed. No outline shapes, no rounded 17:48rectangles because that was one of the 17:50issues. Use clean typography, spacing, 17:52and subtle color blocks. I will say I 17:54have seen rounded rectangles work 17:55sometimes. This was a little bit 17:56overconervative. We specify where text 17:59should sit and then we start to think it 18:01through. Please describe the clean 18:03layout approach without border elements, 18:05the color palette, the typography, the 18:07visual emphasis, and we'll go down and 18:08see how it does this. And then generate 18:10the deck. Here's your inputs. By the 18:12way, this is a subtle thing. I list 18:14three inputs here. I gave it six. I'm 18:17over I'm overshotting the context to 18:20stress it out on purpose for this test 18:21and I think it passed. Then do the 18:23layout. You see how we are actually 18:26acknowledging what we humans have to do 18:27as work. We are acknowledging how hard 18:29it is to do PowerPoint and giving the 18:31system some help here. If you are 18:33wondering, well, how do I get my 18:36particular brand into AI PowerPoint? I 18:40actually wrote a prompt for that. But 18:41the the TLDDR is you have to give it a 18:43slide and tell it to extract the style 18:45from the slide. All right. Then we go 18:47into the slides and what each has what 18:49we're looking for. We're very specific. 18:51Then we go into validation gates. So 18:53please show me a thumbnail. Please 18:55verify contrast ratios. This is the 18:57access accessibility piece. Assume test 18:59text readability at different zoom 19:01levels. So test that. And the failure 19:04conditions now include visual stuff as 19:05well. Let's see how it actually 19:07complied. So we saw some of the work, 19:08but here now we have a design plan. It's 19:11going to give us a layout. It's going to 19:12give us specific colors. And by the way, 19:15you can actually pull those colors and 19:16check them if you want. It's going to 19:18give you visual emphasis without borders 19:20like we asked. Chart styling. It's going 19:22to deal with accessibility. And now it's 19:24going to start implementing. I didn't 19:25have to tell it to go. It just kept 19:27going here, right? And now it's going to 19:29start building. It's going to start 19:30using JavaScript to convert the HTML 19:32slides to PowerPoint. Work, work, work, 19:34work, work. And it has to keep going. 19:36Like one of the things that I am 19:37realizing is that part of why we humans 19:39have to do this so much. Why we spend 19:42arguably half a day a week just on 19:44PowerPoint design as a society in our 19:46work weeks is because of how hard this 19:50is. And this is showing me that AI finds 19:52it hard to it is basically brute forcing 19:54this. This is hard work for for AI as 19:57well. And the fact that we can actually 19:58get to this quality is really 20:00impressive. So there you go. That's an 20:02example of a prompt that didn't work, a 20:05prompt that did work in the five 20:07principles that by the way scale. If you 20:09are watching this in a month or two 20:11months and it's not Claude anymore, 20:12right? Claude went downstream. It went 20:14from max plans to plus plans. But now 20:16Claude is not the best and co-pilot is 20:18incredible for PowerPoint or Chad GPT is 20:20amazing for PowerPoint. Great. The 20:22principles will still be there. The 20:24principles won't go anywhere. The way 20:25you work with a tool to generate a 20:29analyzed result in a narrative arc in a 20:32visual format that's not going anywhere. 20:34This I would argue is the hardest task 20:37for work primitives that we have in 20:39regular corporate knowledge work. I 20:41think it's harder than code. I think 20:43it's harder than Excel. And I think it's 20:45harder than docs in all three of those 20:47issues like text, docs, code, 20:50spreadsheets, the AI can already do it 20:52much better. It's much more fluent. You 20:54can give it longer prompts. It works 20:56especially with Excel now. And also with 20:58code, but not with PowerPoint. 20:59PowerPoint, you still have to hold 21:00hands. And the reason why is it's the 21:02combination of the analysis, dealing 21:04with conflicting data logic, and then 21:06getting it into something that is clean 21:08and visual. Once we get this right, we 21:10are going to change how stories are told 21:12in business settings. We are going to 21:14get cleaner powerpoints. We are going to 21:16get better iteration so that humans can 21:18focus more on the storytelling. Very 21:20excited about it. But it all rests on 21:22being able to actually get the thing to 21:23write good powerpoints. And that is 21:26clearly a promptsensitive art right now. 21:28It is not something that you can do and 21:30just say h just off you go, right? Like 21:32write the PowerPoint. If you want a 21:34short two or three slide deck and you 21:36have very simple data, yeah, that will 21:37work. If you have any kind of sort of 21:39production data and you have a 21:41significant deck to do, it will not 21:43work. And that is why I built this whole 21:45prompting approach. I hope this has been 21:47helpful for you. I hope you enjoyed 21:49seeing a Nate prompt that did not make 21:50the cut. And uh I'll see you next