Check out the conversation on Apple, Spotify and YouTube.
Intro (0:00)
Aakash: As product managers, we spend a lot of time thinking about strategy, our users, our business impact. Now more and more, we’re starting to think about how we code, how we engineer things. We are getting more into coding, but the area that I feel a lot of PMs could grow more in is how to design well. How to use the latest and greatest AI innovations in order to build better slides and charts at work, in order to brainstorm and plan better designs before they make it to a designer or an executive review. These are really important skills and maybe my favorite voice online is Meng To. He has been publishing design guides for years on YouTube, on Twitter. I’ve been following him. You might have seen him on Greg Eisenberger, Peter Ying’s podcast. And he has been one of the most practical thought leaders on the topic of how to design with AI. And so I think we’re fortunate enough to have him on the podcast. And he’s going to break down for you as a product manager, what are the top use cases of these AI tools? And more importantly, beyond the surface guide of how to use them, how do I use them well? Meng, thank you so much for lending your expertise to the podcast and welcome.
Welcome (1:50)
Meng: Thank you. And this is going to be a deep dive into Codex. I’ve been using this since day one, but also before Codex came out just like four months ago, I’ve been building apps, using agents and doing AI for a long time. So I’m going to be sharing with you all my secrets. This is going to be a unique perspective. A lot of PMs are going to want to hear about this because it covers every single piece of the workflow from prompting, using what tools to use. If you only get out of this podcast what are the tools I’m supposed to use or what are some of the basic prompts, I think that’s enough value. But I’m going to do a deep dive. I’m going to show you all the demos, the images that I’m creating, the videos, and the HTMLs. It’s going to be a deep dive.
Aakash: That’s what we like here. Try to go the layer deeper than the surface.
The best tools for AI design right now (2:44)
Aakash: So what are the best tools for AI design? You use them all right now. What are the ones you like the most?
Meng: Yeah. So I’m basically all in on Codex right now. And I know a lot of people love Claude Code. Some people still love Cursor. I think Cursor is amazing. I’m not saying that as a bad thing. I used to be a huge Cursor user before OpenAI came out with Codex, which basically changed the whole mindset around not just building, but also starting projects and starting chats around those projects. But also we have OpenClaw, which recently completely took over the world, became the number one GitHub project of all time. So this is taking the world by storm and every single one of these tools are influencing each other to create the next best thing. And right now I feel like at the epicenter of everything we have Anthropic and then we have OpenAI with Codex.
Aakash: Amazing. I feel like sometimes I am a Claude Code channel, so it’s a breath of fresh air to see Codex. Can you fire it up? Can you show us how you use Codex and how to set it up so that it is used in the best possible way?
The full toolkit: Atlas, Whisper Flow, Obsidian (4:05)
Meng: Sure. So obviously I’m going to be covering Codex, which for those who don’t know is basically ChatGPT but 10x, on steroids basically. So you can not just chat with an AI, but you can create slides, create HTMLs, create full websites, make your mobile apps, everything. And then one thing that I mentioned that you should definitely take out of this is you should use an AI browser. So the one that I’m using, again from OpenAI, is called Atlas. And basically what is really cool about this is for me context is king. Anything that you have that is context is super useful. So you’re browsing something, you want to ask a question, you have the ask button, and you can also let the AI agent take over your computer and just start using it. So at first I just want to go over the tools. I think it’s super important to give you an idea.
Also Whisper Flow, which I’m sure a lot of people use. Basically, instead of typing everything, you want to use your voice. And Whisper Flow is the fastest one. It has the best accuracy and you want to deal with dictionaries, words that are hard to use. So that’s very good. For PMs especially, you want to use something like Obsidian because when you use something like Codex or Claude Code, you’re creating a ton of documents, tons of documents locally. Especially on an OpenClaw kind of workflow where everything happens locally because you have a ton more context and the AI has a ton more power because it doesn’t have to rely on an internet connection or a private database somewhere, but everything on your computer is private. It’s super super powerful and they also have this feature where it’s kind of like a brain tree of the whole information, the whole context around every document that you have done so far, basically your whole life. So this is super important.
Why Meng builds his own tools (6:00)
Meng: Now what I use is my own tool. And I know a lot of people, including PMs, especially technical PMs, which I’ve always been a technical person. Even as a designer I’m like one of the first designers who started teaching designers to code. And that’s a real story. I started my career almost like that.
And basically I build every single tool that I use, including my own video editor. I build my own SaaS project where I build websites from scratch and templates, which I have a ton of videos about. I build my own sort of design brainstorming tool using HTML. And as I mentioned before, I build my own video editor as well. And the reason why I do this is very simple because every human has a very unique set of needs. And in order to solve that, the only way is, one, to experience those frustrations, and second, to go through the experience of solving those frustrations. And so you have a tool like Codex which gives you all of these superpowers, and you want to solve them. And therefore, every time that I have like a podcast, I have a prep document that I talk to my Codex and it creates all of this stuff. And so you can see it gives me the hooks, it gives me all the demos, all the talking points, the research, and then we’re going to get into creating charts, creating storyboards, and the prompts to use them. And one thing I love is also generating avatars because a lot of people are Loom users, for example. They go into meetings and they want to record themselves like training something instead of having to go to these meetings. They record it privately and then they give it to their team. So having an avatar is super important. And nowadays we have AI that does all of this stuff.
So again, this is my entire workflow and this is everything that we’re going to cover. And like I said, I just wanted to start with this because if you’re taking any value out of this podcast it’s all of these tools that I’m showing right now. Just Google them, downloading them is just enough to get a ton of value. And then I’m going to switch to Codex and start from scratch.
Aakash: Awesome. While you’re doing that, I wasn’t totally clear on two things. One was how Obsidian plays in, maybe I just missed the connection. And then the second was how an avatar helps you with Loom videos.
Inside Meng’s real Codex setup (8:43)
Meng: Okay, so I’ve never shown this to anyone outside of our team, but this is my Codex, my real Codex. It’s not a demo one. And these are all my projects. So for example, I have content. And this is a project, and all of these are local folders.
And I just want to answer your questions first before we get into the step-by-step of getting started with Codex. Obviously at this point I assume that you already downloaded Codex. And the idea is that when you talk to AI, you ask a lot of changes, or you want to write documents, or you want to fix issues, you want to fix and move pixels. And sometimes you want to plan YouTube videos, and sometimes you want to draft and prepare a podcast. And what’s happening here is that oftentimes it’s going to generate these documents. And these documents are like content.md. And this is how you consume the content. And this is great. Codex has everything. It’s like an all-in-one app. It can read your HTML, you can open your HTML, you can open your MD files, you can open the images, it’s going to go into full screen, you can open your code changes and all that stuff. However, you have to understand that all of this stuff also exists on your computer. And you’re not always going to go back to your Codex to find these documents.
How Obsidian fits the workflow (10:06)
Meng: So you need a way to organize all of these documents somewhere. And this is where Obsidian comes in. Because Codex, yes, it does a lot of this stuff, but Obsidian is going to allow you to organize your documents, your knowledge base, and the brain tree. And you can always reference the Obsidian content, which by the way are basically just folders. And again, this is the superpower that OpenClaw introduced to the world, is that everything has to be local. Everything has to exist on your computer first, and the rest is for all of these apps to digest that content. So for example, if it’s talking about a coding project, well, you need HTML, you need React, you need a browser. If it’s an MD document, you need something like Obsidian that organizes and reads all of this stuff with beautiful imagery and stuff like that. Or the app that I build for myself. For videos, for example, you need a separate app for organizing your videos because you’re not going to do that here and so on and so forth. So this is why we need these other apps.
Aakash: Got it. And how does the avatar help with the Loom videos?
The UGC shift and AI avatars (11:24)
Meng: Right. Okay. So this is really interesting. I kind of dropped that knowledge. I understand that it’s kind of like out of nowhere, but right now there’s a huge shift into UGC content. Have you heard about UGC content?
Aakash: User generated content.
Meng: Yes. So UGC, user generated content, where basically this is what everyone’s doing. This is the new marketing 2.0. Basically, people are tired of looking at companies and getting corporate bureaucracy messages from them. Instead, they want people like you and me who have real human experience in an era of AI that feels like they’re getting not only a ton of value, but also authenticity and a perspective that they can relate to. So having an AI avatar allows you to go from, today I have to do my hair, I have to put on a nice shirt, sometimes you don’t feel like it. So having an AI avatar, which by the way C-Dense is really good at, or HeyGen for example is a really good AI model for that. Basically those can generate a clone of yourself including your voice, and then you can just use that and type the stuff that you want your AI avatar to say while you do the screen sharing. Which by the way is what this app that I’m building is about, and you just put the two together and then you get your presentation because everything is about presentation. So you know UGC is a presentation. It’s just in a form of something you would post on Instagram or TikTok or Twitter. But if you’re talking about as a PM, you want to give the message to your team. So the best way to do that is you have to record yourself doing a screen recording, say stuff, and then showing your slides. It’s a lot more human.
Aakash: Nice. Amazing. Yeah, I think I’ve seen like some creators create like almost like photorealistic avatars. So I’m excited for that demo coming up.
Meng: Yes. And with Image 2.0, everyone starts with Image 2.0. I’m going to get more into the details of it, but for now let’s get into how to get started with Codex.
Getting started: plugins, skills, and computer use (13:42)
Meng: So the first thing you want to do is to get into the plugins and the skills. The difference between a plugin and a skill is just how deeply integrated it is. Usually a plugin is more like it has the whole backend stuff and the whole integration, more like a Figma plugin or Photoshop plugin where there’s a whole team behind it, whether it’s OpenAI or Slack or Chrome or Linear. So you want to definitely use this. And one of the most amazing things is computer use, and we are I would say in the middle of that journey, but eventually it’s going to get so much better. But right now it’s so good. It’s able to work on your computer. It can open Figma and start designing something on its own by using your own mouse. And that’s so good because you’ve heard of MCP, you’ve heard of CLI. For those who don’t know, CLI is when it goes to your command line and starts communicating with APIs. And MCP is when Figma, for example, introduces some sort of plugin and then it’s connecting to the AI agents and it does all of these tasks. But then you have computer use, which works on everything. It doesn’t require someone to build something. It’s going to just use your computer as a human would. So I would imagine eventually you’re going to be able to do computer use and you have an AI agent that records on your computer and creates videos and presentations using your AI avatar. So that’s just how powerful it is.
But for now I’m using computer use a ton for testing. So for example, a user has a problem, you’re going to go through that flow and you’re going to use computer use for that. It’s going to click here, go to this website, log in, and then go through that flow. Oh, we have this bug. And then it fixes it. So it reaches those pages and then it fixes the bug. So computer use, super super important. Spreadsheets, presentation, Chrome, similarly to computer use. But I would say that nowadays Codex has its own browser and so the AI is able to do everything by itself. And Gmail is super important as well. So it’s going to be able to go to your Gmail and it’s going to be able to reply to your emails and it’s going to be able to do a lot of cold reach, which I know a lot of people do.
Trust and security with AI agents (16:21)
Meng: Now, one side note about this is that I want to talk about trust. Because at the beginning of the year we had OpenClaw and then OpenClaw took over the world by storm. Now the biggest problem with OpenClaw at the time, and I wouldn’t say big, but I did it for like a month until I got into Codex, but the biggest problem with OpenClaw was security. A lot of people were worried, including myself, that if it reads my email, someone can inject a prompt and then start hacking my computer and getting all my OpenAI keys and stuff like that. And so there were no guardrails at the time. But now it’s a lot better. And thanks to things like Codex, you have something that is within a trusted channel, which is Codex or OpenAI. And since they have a lot of accountability versus just an open source project, they can put all of these guardrails. It’s going to ask you for permissions and all that stuff. So which is why I trust installing a Gmail plugin. And nowadays I actually use it to send emails. For example, when there’s a failed payment, I would sometimes send an email saying, hey, is there something wrong? But for a lot of PMs it could be like, someone reported an issue on Linear or something and you receive an email, you can reply right away with an auto reply, but using AI instead, like an agent, which is a lot smarter. So this could be super useful. Obviously Google Calendar would be really good. I’m not a super calendar user at the moment. I’m also not a Linear user, by the way. And for me, I actually use Discord. Because I’m very customer facing, so I use Discord for everything, team chats and all that stuff. And AI can do pretty much everything.
Skills: what they are and how to build them (18:16)
Meng: Now we’re going to get into the skills. So skills is more like anyone can create a skill, including you. You should be creating your own skill. So it’s a lot more hands-on and it’s a variety of stuff and small stuff such as prompting and image generation, or what model to use for image generation or video generation. Or if you’re talking about design for example, what are the design MDs that you should always be considering, or what are the taste skills that you should always use whenever you design a landing page or a slide to make it 10x better looking. And then you have front end stuff. For example, I’m building a Mac app and also an iOS app, which is what DreamCut is on. So I use SwiftUI skills. Playwright is really good for taking screenshots. Netlify is really good for deploying your website because you’re going to create all of these HTMLs and React projects, but they need to be hosted somewhere. So something like Vercel or Netlify, you just drag and drop your HTML or your React project and it’s going to turn it into a URL and it’s going to be hosted on the internet and then you connect it to a custom domain name and then voila, you have your SaaS product or your landing page. So these are the skills that I have and I would highly recommend most of them, at least the ones that I just mentioned. Copywriting is super good as well, so it helps you write better for your landing page but also your documents.
How often should you update your skills? (20:01)
Aakash: How often are you updating your skills?
Meng: I update them when I notice that there’s a lack of understanding that my AI agents have. So for example, when I was building my SwiftUI apps, I was noticing, oh, it’s missing some knowledge about performance and stuff like that. So I would go on the web and I would look for a skill around performance, about Mac development or iOS development. Which by the way is super fun. If you’re like a PM or a founder, you want to build something, start building apps. There’s so many needs and there’s so many ways to create an app around something and you have so much power when you build an app for let’s say iOS or Android because you have access to all of these technologies that nobody is using on an app level. So that’s why I think it’s so fun. But obviously everyone knows how to build a website. So I would download the skills for building a website but also like the taste skill for front end development and design skills and CSS and Tailwind skills. Other skills that I would recommend if you’re building landing pages are libraries like JSAB for animations, or Lenis, or Logomotive, or Barber. There’s just so many of these. I know I’m dropping a lot of names and I cannot cover them all, but there are categories of skills, categories of things you should be using depending on what you’re building, what you’re generating. And then you have to study the AI models that are doing images, videos, just AI agentic stuff, coding. It’s a world of stuff and there’s just so much to learn, which I hope just by dropping the names you’re going to be doing some research on them and then integrating them into your workflow.
Projects: from VS Code and Figma to chat-based builds (24:05)
Aakash: Awesome.
Meng: Okay, so now we’re going to get into the projects. So I want you to think about Codex like ChatGPT but on steroids. That’s basically what it is. Everyone knows how to use ChatGPT. The first time that OpenAI introduced ChatGPT, everyone was like, wow, the future of UI is basically a chat. And I fully disagree with that, by the way. There’s a ton of little mutations, but there’s also a ton of stuff that we’re unlocking. And we’ve never seen more UI nowadays than we’ve ever seen in the past. But at the center of it all is the chat. So like I mentioned before, I was a huge Cursor user. I was also into the nitty-gritty of using Figma for design and using something like VS Code for coding. And we have a whole new generation of people, especially PMs and people who’ve never done coding before, who are getting more and more technical because of AI. And for us the technical people who have done this for like a decade, we are 10xing the workflow. Basically the stuff that we’ve always dreamed of building before, now we can do it. Like for example, never in a million years would I think of building my own video editor, yet here I am. So the PM who’s never done technical stuff is now doing technical stuff. The one who’s never done, like the teenagers who’ve never done an iOS app now suddenly gets into coding and uses Swift to build an app in a matter of days.
So yeah, this is where we are. And we started all of this with chat. Nowadays I haven’t written a single line of code in the past six months and I’ve heard a lot of engineers say that as well. So where are we right now if nobody’s writing code? We’re moving from VS Code and Figma into the chat box. Basically Cursor is also like VS Code but with AI agents. And even Cursor they’re moving away from the sort of GitHub-centric, files and folders centric into the chat-based project. And the same way with Claude Code as well, they’re moving away from Claude Code in the terminal, which is highly technical, into something like Claude Cowork and Claude Design. So you’re seeing this shift into a more project-based chat. And this is where we are right now with Codex. And Codex has started this movement.
Building your project folder system (27:03)
Meng: So with all of this, you have the projects. And a project is basically just a folder. So now if you want to start a new project, you’re going to click here. You can either start from scratch or use an existing folder. Now the folder can be anywhere on your computer. And keeping in mind that your AI nowadays can have full permission on anything on your computer. And in fact, I have this set to full access 24/7 because I trust it because it has guardrails. So what you want to do is you’re going to start this project and you’re going to go to your downloads folder. Everyone knows what their downloads folder is. And then from the downloads folder, you’re going to create a folder called Projects. And then in the Projects folder, you’re going to create a folder for every single project that you’re doing right now. And let me give you some examples. So these are the apps that I’m building. But I also have Content, which is for all the MD files and all the text, the preparations, the video scripts, if you’re a YouTube channel for example, or if you’re making presentations, you would create these markdown files. And then we also have a folder that you can put all your customer support stuff or the issue stuff, like anything that the customer might have and you want to follow up. You also want to have something for your company. If you have a company or you have accounting, you want to create a folder for that. So you’re going to put all of your invoices, all of your registry and documents and receipts, but also all your company files, all the context of the team that you work with and so on and so forth. And then you might have a folder for Skills that you’re going to put all your skills into so that you can use them for the next stuff. If you have kids, you might have a folder for School where you’re going to put all of the kids’ assignments and their homeworks so that the AI has context. Because you don’t want to give AI too much context, otherwise it’s going to cost too much in tokens. That’s why you want to have something like this.
Plan mode: how to start every build (29:52)
Meng: So I’m not going to get into the coding aspect, even though I think this is a hugely beneficial thing to do. You create a new folder, let’s say you want to create an iOS app, then you start a new chat and you can use Whisper Flow and you can start with the context. Usually you can ask questions first. The first thing you want to do is plan mode. So you’re going to start with, I want to plan this. And you always want to start with I want to plan this. And the reason why is because you don’t want to get into the coding yet, because you want to approve every single thing that you want to build first. And then you’re going to be like, I want to build an app that scans a QR code and sends automatically the content of that QR code to my email, for example. So then you have an app idea, and you want to plan it first.
Permission tiers in Codex (30:57)
Meng: So let’s talk a little bit about the permissions. So you have the default permission. I think most people are going to start with this. Now this is when you’re a new user. You don’t trust the AI yet. Your level of trust is like zero to 100. At the beginning it’s more like zero. And then eventually you learn to trust. So default permission, and again this depends on you, whether you value speed and the token cost and all that stuff. You can start with standard and you can start with medium. And this depends on the level of difficulty that you think this project is going to take. So for example, if you don’t have a ton of tokens, you have like a $20 a month plan and you just want to experiment, you want to go with medium or low. But me, I’m on a $200 plan because I use this every day. And one of the key reasons that I’m actually using OpenAI is because they’re far more generous in terms of tokens. I’m sure some of you have already experienced it, oh, you know, I’m already at my limits after five prompts on Claude Design, for example. Well, that’s the reason, because their models are super costly. I don’t think they have like a low, medium, high, and extra high setting. And these are related to the token cost. So like I said, if you have a $20 a month plan, you might want to use low. If you have $200 and you value accuracy over everything else and you do a fleet of agents, I like to do 20 agents running at once, then you go with extra high because you can afford to wait on them.
Running multiple agents at once (32:40)
Meng: So this is a high difficulty one, so I’m going to set it to full access because I already trust it. And you can see everything is happening locally. So I’m going to start with that. And then it’s going to start thinking and it’s going to come up with a plan. And while I’m doing this, I can also go to Content, which is for example I want to create a slide for the stats that I have, like new users, how many new users I have, what are the revenues that I’m doing every day for my projects, DreamCut, Aura, New Form. And I want you to look into my downloads folder and in the download folder there’s a projects folder and you’re going to find those projects folders right there. I want you to read those folders and I want you to look for the revenue stats and the new user stats, the conversion rate, etc. And then I want you to create a slide.
So again, this is why I’m using Whisper Flow, because when I’m using Whisper Flow, I can give a ton of context and it also formats everything that I say in a super nice format. And you can see here while I’m doing this, it’s already done. So now I have my build plan for my iOS app. So now I just need to read through this and kind of make sense of it. If I have any question about it, I will ask a question. It’s super important to ask the question. And when I’m done, I would say, okay, let’s do it.
So architecture, we’re going to need the QR scanner technology, we’re going to need a backend. So the AI is really smart and 5.5 is considered the gold standard nowadays. So you have two of the best AI right now, which is one, Opus 4.7 and two, GPT-5.5 at extra high. And so it’s super good. Give it a try. Honestly, give it a try. I know you’re going to build your dream app just from listening to this. It’s giving me the plan. It gives me the MVP features. It gives me the format, the scan, the content. It’s going to give me the URL, I receive the email, safety features, and so on and so forth. So this is great. And then I can start building. Or if I have a new question, for example, I can say, can we expand on the email? And can we actually have an image, like maybe a screenshot of the QR code? So let’s say the QR code goes to a website and I want to use a technology that takes a screenshot of that website and sends it as part of the email. So obviously you want to start with a question, otherwise the AI might think that it’s time to get into the coding. And this is what it looks like to build an app. Let’s say you’re a PM or someone who’s never done coding before. This is how I would start.
And then for Content, super important, it’s exploring all the files, a superpower. It understands everything that I’m talking about. It understands that I have a downloads folder, it has a projects folder, I have DreamCut, New Form, and it even understands all of the names that I have that are not normal names, right, they’re specific names. And the same with generating your avatar or generating a slide. For example, I can say, can you create a slides presentation in a Keynote format? So it can even create a specific format.
Aakash: So even for a slide, planning is worthwhile.
Meng: Yes, exactly. And this is cool with Whisper Flow as well. You have a dictionary now. So it understood that your name was written this way even though it’s not a conventional name in terms of the way it’s written. So it was able to figure it out. Why? Because when I use Whisper Flow, I correct it and then the Whisper Flow understood that I corrected your name and then it remembered afterwards. And that’s just the power of AI. It understands context and it’s more proactive rather than always waiting for you to ask for something. So here I’m asking for a very specific thing. And because I’m asking for a Keynote file, the question is, is AI able to create a Keynote file? Yes or no? And you can ask that. But most importantly, if the AI is not able to do that, then you can always say, are you able to create a Keynote file? If not, what if you use computer use to open Keynote and then go step by step into creating that presentation in Keynote? And then while it’s working, you have also the ability to queue new commands.
So here it gave me the plan because that’s what I asked for. But also I can steer it and I can queue new tasks that I want to give. And you can also sort of steer the conversation. So for example, I can say something like, what if we generate the image first as a slide in GPT Image 2.0 first? Let’s create five of them. Let’s use this taste skill and let me make sure that it looks good before we actually build the actual slides with layers. So it understands the context of the text. And if I didn’t correct it, it would just say GPT 2.0, which doesn’t make any sense. It could be like an AI model. So understanding these keywords is really key to knowing what to build and how to say it right so that the AI doesn’t make a mistake. On top of the fact that you’re using the planning mode and you’re also asking a question. So those are the three factors before you actually get into building and generating the images.
Screenshots as the highest leverage context input (42:03)
Meng: The last bit of knowledge that I want to give you that I do like 90% of the time, especially when you get into the nitty-gritty when you’re building a website, is you want to take screenshots of everything. Because the AI, an image is worth a thousand words. You can use Whisper Flow to talk and give a ton of context because the more context the better, and you don’t want to type all of that context because you get lazy as a human. You’re limited by your speed and your thinking is way faster than your typing. So the second best thing is using your voice. You can talk way faster than you can type. And the third best, even better than that, is taking a screenshot. Because taking a screenshot, you select for example the browser, and then you have this new shortcut which is like Command Command, and then it takes the screenshot automatically of that browser that you had focused. And this is just a new feature that was launched last week.
Codex Mobile: work on the go (42:59)
Meng: And another new feature that was launched last week is the mobile. This is incredible. When I go outside or I go commuting or I take a taxi or I’m waiting for my food to be served at the restaurant, I can open my Codex which is connected to my computer and I can go to each of these tasks that are being worked on and just keep working on my project by just chatting. So the mobile, the Codex mobile, sits in your ChatGPT app. All you need is your ChatGPT app and there’s a little button that says Codex. You click on it, you connect it to your Codex on your computer, and then you have access to all of these projects. And then you can follow up with all of these chats and everything that you do and ask is being worked on your computer and not on your phone. And that’s the magic.
Aakash: Does your computer need to be on for that?
Meng: Yes. So there are a few options that you have to do in your settings. I’m not going to go into my settings just because I don’t want to share the sensitive stuff, but you’re going to go through a workflow where you’re giving permissions to, for example, keep my Codex on even though my screen is locked and work on these things. So as you can see, I’m working on slides and it’s taking. It understood that I have a folder for content and in my folder for content I have a prep file around your podcast and we have stats, we have these beautiful slides. I’m also using the taste skill that gives me better design skills and better fonts and typography that looks like it was done by a senior designer.
Now, that being said, of course it’s going to make mistakes. So for example, sometimes it likes to cram a lot of information, a lot of lines, and it’s not smart enough to know that it’s not able to cram all of this stuff in a way that makes sense. So for example, you can see it’s trying to put so many features in a space that doesn’t allow it to, and that’s where it struggles. So over time you kind of understand how to deal with these things, these pitfalls. And then you overcome it. And this is where you as a human become better at making these micro decisions in a split second and then you say, this is wrong, fix this. And you take a screenshot for the context. It does this beautiful animation, gives me the context, gives the screenshot of the image that you’re talking about. You go to a website, you go to a slide, you do that shortcut and it gives you the context and so on and so forth.
So the beauty of all of this is that all of this is happening all at once. When usually traditionally, if you think about ten years ago using Figma or using VS Code, you’re limited by one task. You do one thing at a time. You do one project at a time. You have one cursor. What you need to understand nowadays is that you don’t have to do that anymore. You have an army of agents. So when I say an agent, this is an agent, this is an agent, this is an agent, and they’re all working simultaneously. And on top of that, each of those agents can generate multiple designs at once or multiple documents at once. And then you are kind of cloning yourself, if you will. And the only task left is you as a product manager who is managing this fleet of agents to create your dream projects.
The taste skill for design quality (47:32)
Aakash: For that taste skill, is that one of the skills that OpenAI provides or did you create it? What does that look like in order to generate slides of this quality?
Meng: Sure. So I’m going to ask you a question because I don’t want to switch between the browser and this window. So the taste skill, I want you to go find on GitHub. There’s a taste skill. You can also give me the list on the taste skill from a website. Maybe go on Google and also go on GitHub, find the taste skill. So I’m going to do it super fast. And there you go. Taste skill. It’s from the design taste front end. You have different versions. And then you have redesign skill, soft skill, output skill, brutalist skill and so on. There’s also one for image gen which is for image generation. So the one that I’m using right now is GPT and it links to the GitHub project and it gives me all the skills. It’s basically just a file.
Aakash: So you just picked up on this known file. You didn’t have to modify it much. And we already get a pretty good slide output. Sounds like you’d give that slide output a couple of iterations. Would you then ask it, once it’s done the Imagen one, to create a PPTX? How do we close the loop on this presentation?
HTML vs Figma vs Keynote: how to close the loop (49:56)
Meng: Sure. So again, it depends on your workflow and what tool you want and what is the last format that you want. So for example, I can say something like, I really like the first slide. Can you do that in HTML? And I want you to use the best image to HTML generation vision possible. And I know you can do it because I know GPT-5.5 is one of the best. So this is one way to go about it. I really like the fourth slide and I want you to recreate it in Figma. Maybe we can use computer use to open Figma and then try to recreate it. So that’s another way. And can you use the Figma MCP to create the slide, let’s say slide number four. So again, these are for experimentation but this is how I would close the loop. You have multiple options. I would definitely go with the HTML route just because the AI has really really good code generation nowadays. If you’re asking it to recreate something in Figma, it’s going to take way longer and you might need to know Figma and you might need to install the MCP. So if you want to stay in Codex, which I think a lot of people would want to do that just because you don’t want to have ten or twenty subscriptions all at once, like a Figma subscription and all of these, so you want to keep sort of the blast radius as low as possible. So it depends on your workflow, it depends what is the app that you use, but this is how I would do things.
Aakash: Okay. So Imagen is almost like a first draft, which you can give it feedback, then you can export it using any of these four different prompts, of which sounds like HTML is going to be the fastest.
Meng: HTML is going to be the most controlled, the fastest, and it’s all going to happen in Codex.
Aakash: Probably the least tokens then too.
Meng: Absolutely. So you have to understand also you can take that and turn that into a video. So for example, you can say something like, oh, I have a new idea. I want to turn those four slides into a video. So maybe you can use something like HyperFrames. And I believe there’s a new one. Maybe you can look into those. Let’s plan this first. And basically the goal, and this is good that you give the goal also, is that I want these four slides into a video that shows the slide one by one with maybe a fade transition, and each slide has a five second duration. And maybe in that video we can do a transition of each bullet point, like one second each. So that’s kind of how I would do it. But there’s just so many ways. And this is the beauty of where we are today with agents, is that you’re not limited to one tool. Before, my whole life revolved around using Figma. And of course now I could say the same about Codex, but in this case we’re actually using agents, and the agent is not using only Figma. And the agent has access to the whole world of tools to create the goal that you want it to do. And I think that’s really beautiful.
Aakash: 100%. I think that people who are stuck in ChatGPT haven’t realized that this can connect to all of your existing apps if you’re using it in Codex.
Meng: Absolutely. Everything you can think of, anything that fits your workflow can be done in Codex. And you can now use computer use. You can also do these screenshots that I highly recommend. And you can also do that on the go by using Codex Mobile.
Building your own AI digital twin (55:07)
Aakash: I think you had like the avatars we talked a little bit about. If you want to talk about that, we can quickly go over it, it’s up to you.
Meng: Sure. So the avatar is more like a personal thing. Just because I kind of teased everyone about the avatar thing and you had a question about it, let’s talk about it. So nowadays UGC content is a big deal and I think that one of the things that you have to create the most is a digital twin of yourself. So in order to create a digital twin, what you would do is that you would take the best headshots of yourself and you would drag and drop into this. But in my case, since it’s kind of personal, I don’t want to start dragging and dropping my portraits. I’m just going to ask.
Okay, I want you to go into the content folder and I want you to look for the best pictures of me. And then from those best pictures, I want you to create an avatar, a digital twin of myself, using GPT Image 2.0. And I want different angles, different lighting, in a studio setting, in an outside setting, me using my camera phone, recording myself, or me in my studio, or as a YouTuber or as a content creator, or me at the office working at my desk. So I want you to generate ten images of me using GPT Image 2.0.
So this is how I would do it. But normally you would just drag and drop the image right here, just the same way. And then what I would do next, once you have your avatar, I would say something like, okay, so I have these screen recordings that I did using either Screen Studio or DreamCut. And then I want you to take those screen recordings and put my avatar and I want you to create a video script using a Markdown file. And I want you to say those things using my digital twin to say those things that I have in my video script and next to the video presentations that I did or the image presentations that I did. And then I want you to use HyperFrames to do all of that stuff into a one minute video. And I want the 16×9 content for YouTube. I want the 4×3 content with captions for X. And I want the 9×16 content with my avatar at the bottom and then the screen recording at the top with a hook. That’s going to be 9×16. Finalized at the end with a call to action that says something like, drop a comment, I’m going to send you the whole workflow.
That’s how I would do it. And by the way, I put a lot of knowledge that I’ve been doing a lot recently. So this prompt is going to be so good for you guys. But yeah, this is how I would do it. And obviously it’s not guaranteed that it’s going to work. There’s a lot of stuff. But if I have to imagine myself in six months time, in a year’s time, I would imagine that AI will be able to do this perfectly. Which comes back to the point where creating a digital twin is going to be amazing.
I do want to drop a little FYI there. There’s going to be a pushback on AI video generation and image generation where you don’t want to deceive consumers. So doing a digital twin is fine as long as you’re not creating yourself using a fake product or doing stuff that are not real. But I would imagine that creating presentations is fine and you as a human have to validate the content to make sure that you’re not deceiving anyone. But there will be a huge pushback because I’m starting to see a lot of these video generation using C-Dense that are creating like fake products. And a lot of people are going to complain to the FCC and there’s going to be a lot of bad stuff going on. But there’s also the good side of it, which is what I’m showing you guys right now, which is, okay, I want to create a video presentation that I’m going to send to my team, and these are the talking points, this is my AI avatar because I don’t want to do my hair, maybe it’s nighttime, the lighting is horrible. So this is the best way to do it.
HeyGen vs Seedance: the right video tool for the job (59:49)
Aakash: Should people be using Seedance for that video or is ChatGPT image fine?
Meng: So ChatGPT doesn’t have video generation. They already discontinued their video generation stuff. I would probably use Google Veo or Google Omni. But the number one thing that I would use if you’re focusing on presentation, I would use HeyGen. And HeyGen has a plethora of tools for lip syncing, for creating your avatar, creating your digital twin, specifically for the need of presentations. Now, Seedance is the best AI model for everything, including filmmaking. It has the best understanding of, let’s say you want to ask for a cat who does a vacuum in Tokyo. Well, Seedance is definitely going to do the best job at that.
How to investigate the design space of a product feature (1:00:44)
Aakash: Okay, amazing. Last point I just wanted to cover off in Codex is how do you generate divergent ideas for app design? Like we were talking about that QR email scanner. What’s your workflow or the recommended workflow for PMs when they want to really investigate the design space of a product feature?
Meng: Sure. So the first thing I would do, I think prompts are very useful to a lot of people. Okay, so I want you to open the app. I want you to open Xcode and I want you to run the app. And then let me know if there are any permission issues. Let me know if I have to create an Apple account or developer account. Let me know if I have to log in. But at the end, the goal should be to open the app and to run the app on either my phone or on the iOS simulator so I can test the app. And then once you run the app, I want you to take a screenshot of every single part of the flow. So to give some examples, we have the sign in, we have the payment flow, we have the editor flow, and so on.
And at every step of the way you can ask for Codex to take a screenshot of the app and to make you validate things and to ask you if the pixels are right or the designs are right. Now, it’s kind of hard for me to show you without getting into the nitty-gritty of all of these projects that I’m working on, which can be a bit sensitive, but in this case, if the app was already built, so for example, PowerPoint presentation. See, I’m constantly surprised by AI. It’s able to open the PowerPoint presentation and it created the PowerPoint presentation. So if you’re already familiar with PowerPoint or Keynote, you already have a basic slide that looks better than what most PMs can do in terms of design. I’m not saying it’s better than the senior designer, but it’s still better. And you can always upgrade this with the taste skill.
Summary of PM use cases (1:05:00)
Aakash: Amazing, guys. So we just walked through some of the most important PM use cases for you in design. Building app designs, using real screenshots to iterate and improve on those app designs, building slides and charts with your real data, and as you saw, minimally directing where to even pull that data from. It has full permissions because you toggled on for permissions. And this is how you really increase your leverage as a product manager.
Why technical PMs are keeping their jobs (1:05:31)
Aakash: Another question I had for you, and we were talking about this pre-call, was around this topic of PMs potentially shifting out of product management. Right now last week we saw 10% of Meta get laid off. A couple of weeks before we saw similar numbers at Oracle. The market is starting to get flooded with PMs. And a lot of these people, they’re being forced out of a job. They’re looking at the world and they’re saying, whoa, maybe I should be building an AI app of my own. What are the things that a PM needs to know if they’re going to go from PM to founder to really succeed?
Meng: Right. So I just want to clarify something. When we say that PMs are getting fired, I think it’s important to clarify that it’s the non-technical PMs that are getting fired. The technical PMs are actually staying at those companies. And this gives us a very important message, which is that as a PM, you need to get more technical. And when we say technical, it doesn’t mean that you’re going to write the line of code. Like I mentioned before, I haven’t written a single line of code in the past six months. Technical is this. This is what I mean by technical. You’re no longer the bureaucracy player, the politics player in the company. Nobody needs that anymore because AI can kind of juggle around that. Your job is basically to understand all of the jargon. So what is the best AI model? When someone who’s more technical than you, who’s working on those projects, when they talk about GPT-2 Image 2.0, you know what it is. So those PMs who have these superpowers and who have all of this knowledge base and who have all of this workflow, they will be the ones who will be staying. And they will be the ones that will be transitioning to a new role. And for those who got fired, by the way I’m sorry, but you will probably transition into a new role that is more technical and then your job will be safe.
But beyond that, let’s say you transition to that new role, you also have a window, for me, to a far better role than any of those working at a big company. Because these big companies, they don’t have feelings for you. When they need you, they hire you. When they don’t need you, they fire you. But what if you’re your own boss? So I think for now, this is what I do. And I manage a team of six people. And basically I have built these products, they’re making well over a million a year, and I can afford a team, I can afford to pay my tokens, which is amazing because I love doing what I’m doing and I want to keep doing that.
So the end goal, the end game, is probably starting your own company. Because AI, you don’t need to be scared of starting a company anymore. Because AI can take care of all the paperwork for you, all the accounting for you, all the boring stuff, the janitor stuff, the stuff that you were always scared of doing, and even the marketing stuff it can do a lot of that for you. So at the end of the day, the last 10%, or the last 8%, is you orchestrating everything. You’re the one who does the quality assurance. You’re the one who makes sure that the plan that you give to AI makes sense. You’re the one who has the moat. You know best because any business rule is that you have to be the one who knows this topic the best, otherwise you don’t have any leverage. So when you have this knowledge, whether it’s around coffee, design, PM, lawyer stuff, you should utilize that and then let AI be your amplifier, your magnifier, so that you can quickly make these micro decisions and you can deploy a fleet of agents. I think that’s the end game.
But I think for now, you are probably, let’s say, either you got fired and you have to transition to a more technical role, or you’re already technical and you didn’t get fired, which is amazing. And then eventually you will get fired because AI will replace everyone at some point, or we will find new ways to do our job in which AI cannot fire us. But it’s a cat and mouse chase. But eventually, you solve that formula. And there’s by the way a formula for everything. There’s a recipe for everything. This is what I kind of learn nowadays. For everything there’s a formula. Whether it’s a YouTube channel, there’s a formula to go viral. Whether it’s an exposé post, there’s a formula to go viral. Instagram, TikTok. And you can look at my Instagram, you can look at my LinkedIn and all that stuff. I have viral posts on all those things. And my point is that there is a formula for all of those things.
The 5-star to 11-star framework (1:12:02)
Meng: And there’s also a formula which is for the end game to start a business. And that formula is you being the best at your job and you deploying your fleet of agents and you utilizing that last 8% of human taste to orchestrate all the tasks and make sure that the quality is always higher than the baseline. But you have to understand that baseline is getting higher because of AI, but also the ceiling of quality is also getting higher. So your job is always to get to if this is a five star, which is by the way from a podcast I listened to from the CEO of Airbnb, you want to get to 11 star. But at the minimum it should be five star. And then you have to imagine what is six star, seven star, eight star, nine star, ten, and eleven. And then your job is to always be above the five star, six star, or seven star. And to think about the future, which is the eleven star.
Aakash: Wow, what a persuasive argument for tripling down on your product management skillset, product strategy, vision, delivering a six, seven star, eleven star eventually experience. That is pretty much the future for all of our roles. Meng, you have broken it down for us in an amazing way. If people want to learn more from you or connect with you online, where should they go?
Where to find Meng To (1:12:48)
Meng: Sure. You can go to X, MengTo. You can also LinkedIn, Instagram. I’m starting my TikTok channel, so it’s also MengTo I believe. You can also find the products that I build. These are the stuff that I vibe coded and I’m using AI to create all of these things. You can check for yourself the level of quality that it is. And keeping in mind that each of these products have way over 10,000 prompts each. So do not give up on the first prompt. But yeah, Aura, New Form, DreamCut. I’m going to keep building products. I’m going to keep learning these formulas, and I hope you will too.
Aakash: All right, we’re going to link all his socials, Aura, New Form, and DreamCut down below. Subscribe to his YouTube channel. Follow him on all platforms. Meng, thank you very much.
Meng: Thank you so much for listening and thank you for inviting me to this podcast. It’s an honor.
Aakash: My pleasure. Bye everyone.