AI for Presentation Decks — A Deck, a Full Transcript, and 13 Voiceover Clips in 30 Minutes

Since AI came along, making a deck has become easy.

But presenting still isn’t.

Every talk and every pitch I’ve given in the past year has used an AI-generated deck and an AI-written transcript. Throughout, it draws on my project memory and the audio of my client meetings. One listen before bed, one listen on my morning run — and by the time I’m on stage I know it well enough to speak along with the AI.

Today I want to settle three things in one go.

1. What AI deck tools on the market are actually missing

I used to be at Tencent, and I’ve been in advertising for over a decade — close to 20 years now. Setting aside all the smaller proposals, I’ve done more than 500 pitches. In the early days, two or three weeks per proposal was normal; by the time I was a senior planner deep into my Tencent years, it was about two or three days.

Doing this with Claude Code now takes 20 to 30 minutes — because it has everything you need.

I’ve tried the whole spread of AI deck tools, domestic and foreign — NotebookLM, every flavor of “AI PPT” service. So why do you come away feeling like you barely saved any time?

• It doesn’t know what you’ve done before — none of your language style, visual templates, or design sensibility
• It hands you a generic template — no distinction between a knowledge-share, a strategy review, and an investor pitch
• It stops at a passing grade — it doesn’t know the backstory of why you’re making this proposal
• It only delivers the deck — but a deck is PowerPoint, and the hard part is getting the point across
• It can’t rehearse with you — practicing in a mirror or recording yourself burns another half day

2. Three things that make AI truly know you

It isn’t about dragging hundreds of decks into it. It’s about three things —

1. Project memory — the memory Claude Code maintains inside each project folder. Add to it every working day and six months later you have a complete hands-on log. When I put this talk together, I didn’t have to tell it anything about what I did over the past three months. It just reads it.

2. Meeting audio — and the key word is audio, not the transcripts Tencent Meeting or Feishu export. Those are full of ASR errors. I drag the audio straight into Claude Code, give it the background briefing (who was in the room, why we met), and Gemini 2.5 Pro transcribes it with very high accuracy. Those meeting records become highly accurate context inside the project.

3. Past decks — but what it learns is the visual style, not the content. I give it my representative proposals, it converts them into images or Markdown so it can read them, and it works out my visual fingerprint and my pitch logic.

3. HTML rehearsal mode + voiceover in my own voice

What I have Claude Code prepare isn’t a .pptx. It’s an HTML web deck.

Here’s how I framed the request:

“Vertical paging, one deck slide per page. Presenter mode — the transcript can be shown or hidden at any time. Calculate speaking pace at 200 characters per minute and control the runtime. Then use MiniMax speech 2.8 HD to turn the transcript into spoken audio. Every time I flip a page, play that page’s voiceover.”

I’d already configured my MiniMax API key in Claude Code beforehand. I gave the request almost offhandedly — “just pick some broadcast-style voice” — and it ran the whole thing.

The final delivery:

• A 13-page HTML deck (generated in my PPVI design style)
• A 15-minute transcript (written in my language style)
• 13 MP3 audio clips (voiceover cloned from my voice)

One listen before bed, another on a run, and by the time I’m on stage I can deliver it fluently in my sleep.

4. That session at Aranya Jinshanling — two hours to get everyone articulating themselves

When I gave this talk, I hadn’t prepared it the day before. It came together on a whim the same day.

I was on vacation at Aranya Jinshanling with a few friends. Someone suggested everyone take a turn sharing what they’d done over the past three months (three months for me, maybe one month for them). From deciding to share to starting to share, we had less than two hours — plus a lunch break.

Among the friends there were big-company executives, the principal of an agency that used to handle a high-end sports brand, a director, and a producer. One woman there couldn’t use AI at all a month earlier — let alone write code; she wasn’t even that comfortable with a computer. When I’d shown her AI applications before, her reaction was always “whoa,” “the sky is falling,” “am I going to lose my job.”

But after a month of using Claude Code — inside those two hours — she laid out everything she’d done that month, completely and fluently.

Every single person used this method to put together something complete and coherent to say. Sitting there watching it, I was genuinely floored.

5. Why HTML and not .pptx

Plenty of people will say HTML doesn’t look as good as a deck.

But when you need fast delivery and you need the text to show the logic with real clarity — HTML is the richer medium. It can carry audio, tables, charts, dynamic interaction. It reads smoothly on a computer and on a phone.

If you want more polished visuals, NotebookLM has an API on the enterprise tier. Claude Code does the content and structure first, then calls the NotebookLM API to convert to .pptx — the best workflow going.

Wrapping up

This is what I mean by the “invisible AI brain trust” —

AI takes over a large chunk of the offstage preparation, and all you have to keep is your presence onstage.

Next episode I’ll do one on how to get AI to teach you to find models and get API keys — that material is genuinely worth money.

One last question: what are you presenting soon? A quarterly review? An investor pitch? An industry conference? Tell me in the comments, and next episode I’ll take the scenario the most people care about and demo it in more depth.

Source: EP0009_audio.mp3 · ASR model gemini-2.5-pro · full text of the roughly 14-minute original recording

[00:00] Come on, friends, good to see you again. Now that we have AI, making decks has become an easy thing — but giving the talk is still not easy. Over the past year, every talk I gave, every pitch I made, the deck was generated for me by AI, and it even wrote the full script. The whole process pulls from my project memory, plus the audio files of the online meetings I had with the client early on.

[00:17] A little while back I did a talk, and AI quickly prepped a 13-page deck based on my project material, wrote the full script, and turned that script into a podcast-style audio version. And the night before the talk, plus the next morning while I was out running, I listened to it a few times over, and I was completely comfortable with the content — I just walked up and gave it. That experience is honestly hard to beat.

[00:38] So in this episode we’re going to nail down three things in one go. First, what those AI deck tools on the market are actually missing — why they can’t get anywhere near as good as what I do with Claude Code. Second, how to get AI to genuinely understand you, like an assistant who’s been with you for years, handling this stuff for you. And third, how to turn a deck script into a voiceover-style audio you can listen to over and over.

[01:01] These three things I basically do entirely in Claude Code. In cases where the visual presentation matters a lot, I’ll also use NotebookLM to prettify the deck. I used to be at Tencent, and I spent something like a dozen-plus years — close to 20 years — in the ad industry. Over that stretch I put out close to 500 decks, maybe more, counting all the proposals big and small, and that’s not even counting revisions — just pitches, roughly 500-plus times.

[01:24] On those proposals, early on, one proposal taking two or three weeks was completely normal, because you have to go back and forth with the client. On the fast end, by the time I was pretty far into my Tencent years, as a senior strategist, a proposal could be fully done in about two or three days. But now, doing this with AI only takes 20 minutes, maybe 30, because it already has everything you need.

[01:45] Maybe all you need to do is bring one core creative idea to this particular proposal, or one core point of view, and it’ll get the thing done for us in my own language style. So, no more preamble — go ahead and like and save this, so it’s easy to find when you want to come back to it. Let’s actually do it live. Since I already generated that deck at the time, I’m going to reconstruct it now and show you roughly how I did it. Voice input method, as usual.

[02:15] We need to state the goal of this talk. The audience for this one is a group of Claude Code users. They’ve been using it for about a month, so they’re past the beginner stage. And through this session I want to cover the projects I’ve built with Claude Code over the last three months, and how I manage Claude Code’s memory, how I manage Claude Code’s projects, and some of my thinking on how to get Claude Code to work with me better on these projects.

[02:41] Including some of the rules I’ve defined for Claude Code, and some of the hooks files. These count as deep usage techniques, and I want to share them in a linear logic, roughly 15 to 20 minutes for this audience to listen to. So now I want you to prep me an HTML — HTML format, yes, a web page — and this page should scroll up and down, with every page styled like a deck slide.

[03:06] At the same time this page needs a presenter mode, so I can show or hide the script whenever I want. And please write that script in the language style of my previous talks. Note that my speaking pace is about 200 characters per minute, so watch the word count and don’t run over the talk’s time. And once all that’s done, I need you to use the MiniMax 2.8 HD audio model, just pick some broadcaster-sounding voice, and turn this script into a voiceover audio.

[03:37] When I get to each page, play the audio for that page. Okay, if you have anything you need clarified you can ask me, and if not you can start working. And once the work is done, check the document yourself, and only call me in to check it after there are no problems.

[03:52] Alright everybody, we can see it’s handed the requirements over to Claude Code. So let’s let it get going, and let’s take a look too. Of course I also know what you want to say — there are tons of AI deck tools out there right now, overseas there’s NotebookLM, right, and domestically there are the AI PPT ones, all of that. And I’ve tried every one of them. So why is it that when I used those other tools, I felt like I hadn’t saved any time? Think about why that is.

[04:18] Let’s go through them one at a time. First, most AI tools have no idea what you’ve done in the past, and no idea what your language style is, and no idea what template design style you use — it’s purely handing you a generic template to work with. But if the content of this particular talk is a share, or a strategy report, or a pitch to investors, then a different style of deck actually needs a different presentation logic, a different structure, a different visual approach, right?

[04:46] A report to investors is definitely going to be extremely simple, extremely at-a-glance — a dozen-odd pages might do it. There are enormous differences here. But these generic AI deck tools can’t grasp that. And they don’t know your purpose in making this proposal either, and they don’t know what came before or after — they hit a 60 out of 100 and call it delivered, right? That’s the first point. The second point is that AI deck tools, the traditional tools, just finish the deck for you and that’s the end of it.

[05:07] But the genuinely hard part is — since PPT stands for PowerPoint, right, it’s about strengthening the point. So the core of it is how you get that point across when you’re speaking. And that’s where the script really matters. How do you prepare a script that’s genuinely influential, genuinely full of energy, genuinely in your own voice — that’s something the deck tools on the market today are completely incapable of.

[05:29] And then the last one is that they can’t rehearse with me. Before, I’d practice in front of a mirror or record myself, right, which took a lot of time. But with a tool like Claude Code backing me up, I can even do a version of the script in my own cloned voice and have it play back in my ears on loop. Sometimes when I go pitch, it’s like sleepwalking — I can just deliver the whole proposal completely fluently. That’s also why so many people use deck tools and end up feeling like they haven’t saved much time.

[06:00] Alright, pain points covered. Let’s see where it’s gotten to over here. OK, it’s asking which narrative structure to use. Let’s go with the timeline journey — that gives it a linear logic, so the audience can follow us more easily. Length — I already mentioned it, 15 minutes, 15 pages. Good. Design style. Pay attention to the design style here, because it’s actually giving me three styles to choose from. The first is Cruna business glassmorphism, which is a design style of my own I preset earlier.

[06:22] And PPVI is also a style of mine. PPVI is a fingerprint, you could call it, of my own distinct design style. It actually specifies which font goes with which heading level, and for its light and dark palettes, the usual ways to emphasize content and lay things out. So here we just need to choose our existing PPVI approach, and then we can submit and let it carry on.

[06:40] So back to what we were talking about — what’s the difference between this new approach and the old one. You can see, look, these are all the proposals I’ve done going back to maybe 2020. What I did was take the especially representative ones out of that pile and hand them to CC — just throw the files at it. It can convert them to images and read them, or convert them to MD files and read them, and it can pick up the thread of my pitching logic, my framework. And that’s a very important piece of context.

[07:10] The second piece is that every day, on every project, I’m talking to AI. For instance, on some of my recent pitches, I’ll open a dedicated folder, and in that folder I’ll put all the recordings of the online meetings with the client — note, the recordings, not the text transcripts the meeting tool exports. Because I know everybody holds meetings on Tencent Meeting or Feishu, and after the meeting it hands you a full stenographic transcript. That transcript is actually really inaccurate — its error rate can hit 20%, even 30%.

[07:43] So why do I want the audio file instead? Because in the process of transcribing audio, we can give it a prompt, we can give it the background. Meaning, once I tell it who was in this meeting and why they met, my large model — I use Gemini 2.5 Pro — can produce meeting notes at basically 99-point-something percent accuracy. And once that’s done, I drop it into the project folder, and it becomes an extremely accurate piece of context.

[08:09] So look, when I was making this share, I didn’t actually need to tell it what I’d done over the past three months. It just goes back through the MD files I keep in memory, my notes, right, even my past conversation logs, and takes what it needs — the pieces that let it lay out my methodology, the core methodology. It just pulls the memories it needs, on demand. And it understands me really well. Really well.

[08:30] So, once we’ve fed all that into the deck — I mentioned it’ll learn my color style. Let me just show you. And the most important part? You can see it’s working right now — but when I originally made this document, it wasn’t even the day before the talk. I did it the same day, because that talk came together at the last minute. I’d gone with a few friends to Aranya at Jinshanling — we were there on vacation.

[09:00] And we were also going to swap notes on how everyone’s one-person-company projects were going, or some lessons learned. It was spur of the moment — everyone said, let’s each do a short talk on what we’ve learned these past three months, right? And at that point, with that suggestion on the table, from deciding to do the talks to actually starting them there was less than two hours, maybe plus a lunch break. Every single person used this method to put together a complete presentation. It was genuinely stunning.

[09:25] You have to understand who was there — a senior exec at a big tech company, then someone who used to run an ad agency for a high-end sports brand, and also a director and a producer. Women who two months earlier had absolutely no idea how to use AI — forget writing code, they could barely work a computer. And in that situation, within two hours, they’d taken what they’d done over the past few months, over the past month — theirs was a month, mine was maybe three months.

[09:50] Whatever they’d done over that stretch of time, they turned it into a complete, completely fluent presentation. It was so — you can’t imagine how stunned I was watching it. Because a month earlier they were still the kind of people who, when I showed them some AI application, would go, whoa, the sky is falling, oh so that’s how it works, am I going to lose my job — that was where they all started. And after using Claude Code for a month, just one month, they could produce content like this.

[10:16] Let’s — let’s wait a second. While we wait, let me share something. I actually covered this yesterday too — why use HTML for the presentation now instead of a PPT, right? Sure, a PPT’s visual presentation can be prettier, right, and what NotebookLM produces is definitely great. But when we need fast delivery, and we need the text content itself to lay out the logic really clearly, HTML is a richer-media way to do it.

[10:41] Because as it just mentioned, there’s the audio part, and there can also be tables, or charts, or dynamic interactive pieces. And with HTML, it reads really smoothly on any computer, even on a phone. And if we take it a step further, we can make that HTML editable, so it’s not like what NotebookLM generates, where the content is complicated to edit, right? It really is a big trend.

[11:08] Let’s talk a bit more about MiniMax, right. You can all see that I casually tossed off a line about using the MiniMax speech 2.8 HD model to generate the audio, right. And even though I said it that casually here, I’d actually already configured that model’s API key in there ahead of time.

[11:28] As for how to get AI to teach us to find the right model and get hold of that API key, I’ll do more videos on that later, so make sure you follow my WeChat Channels account. Honestly, this kind of substance — you could sell it. So, alright, let’s wait a moment and see what it produces.

[11:50] Okay, it’s finished now, so let’s look at the page. This page is a web version of a deck, and we go into rehearsal mode, and we can flip through — it’s a deck. And this deck also adds what’s called rehearsal mode, which gives you, for this page, what you need to do, and what the script is, all right here. And now let’s listen to the automatic playback.

[12:13] Hi everyone, I’m Zhang Pinpin. Today I want to share how, over these three months with Claude Code, I turned it from an ordinary AI assistant into a partner that genuinely works reliably. This method has three layers of configuration, a seven-step process, and a set of values. When we flip the page — let me start with an observation. The biggest pain in using AI isn’t that it isn’t capable enough; its capability keeps climbing. What actually breaks you is that on the same task, it gives you one version today and a different version tomorrow. The first thing I did was write AI a value priority list, because AI often runs into — so when we flip the page, it goes straight into narrating this page’s content.

[12:45] You can also export the whole audio as one file and listen to that. And there’ll be people who say, your page doesn’t look great, right. This web page really can’t look as good as a PPT. And at times like that, we can bring in NotebookLM. For enterprise there’s actually an API you can call directly. Once you call that API you can turn a page like this into a PPT file.

[13:06] A while back I was lucky enough to sit in with Mulan AI — the company everyone knows, the really great one doing video-canvas workflow agents. I’m good friends with the founder, and when Lu Shan was doing his pitch, I used this same method to help him lay out this deck. Let me show you. For this deck, Claude Code did the structuring first, and NotebookLM finished it into a very, very polished pitch deck.

[13:33] And then including when we’re doing AI comic dramas, we analyze some classic characters, and you can do it this way too. For example, this deck analyzes the male lead in Neon Genesis Evangelion, Shinji Ikari — an analysis of his personality. It’s in a completely usable state.

[13:52] So that’s it for today’s share. If you want to see more, remember to follow me, follow me — okay, see you next episode, bye-bye.