Check out the conversation on Apple, Spotify and YouTube.
What We’re Covering Today (0:00)
Aakash: Where should PMs actually be using AI tools in their workflow?
Mike: I think exposing users to context limits for the models are BS on the back end. Most of them are handling this better anyway, and you can see that even in tools like Claude Code, there’s a much better way.
Mike Ball is leading product at David’s Bridal, which is undergoing a massive digital transformation. He has been testing out all the AI tools. We have put our heads together in this episode to create the AI native PM operating system.
Mike: We’ve been talking a lot about Claude, Claude and Claude. Why not try GPT? GPT with 4.0, I think the last two model releases, I felt like got more towards lazy and less useful, so I just don’t spend as much time in there. I just feel like Claude’s more reliable for me.
Aakash: If you’re stuck behind corporate structure, enterprise tools and your organization’s not really supporting AI, here’s how you can kind of think about working around it with your own tools and projects.
Mike: That’s a way better method than most people do.
[Sponsor break: Bundle announcement with Mobbin, Arise, Relay App, Dovetail, Linear, Magic Patterns, Deep Sky, Reforge Build, Descript, and Speechify]
What Makes an AI Native PM? (1:44)
Aakash: Mike, welcome to the podcast. What makes an AI native PM?
Mike: I think there was a term floating around for a while that I think was on the right track. I’ll say, think in prompts is kind of this fundamental thing that sounded silly at the time, but when you start doing this day to day and you’re actually integrating tools, I think it’s how you end up working.
What is the thing I need to get done? What is the best way to do that thing? And if you’re thinking about AI as an extension of yourself, you’re thinking like, what are the instructions, what are the steps? And if you’re reflective, you probably do that to yourself at some level anyway. You have your inner dialogue.
I think the biggest thing is the AI native PMs are actually working from what do I need to do to what are the steps that I need to get done, to what are the best tools to get me those things. And they’re pushing themselves beyond that mental block that’s like, “Oh, that’s too technical. Oh, that sounds too hard, or I’m not sure how to go about doing that.” They’re working their way up to that, even if the things may be outside their comfort level.
The Operating System Concept (2:43)
Aakash: So, can you show us what the best operating systems really look like?
Mike: Yeah, I mean, it should be a pretty familiar view for most people who are playing with AI right now. The first thing we worked on together is this map. And the more I thought about it, the more it was kind of obvious to me that it’s not necessarily a set of strict tools. That’s why you mentioned operating system.
You’re actually operating with a layer of abstraction from UI. Logging into 20 different tools for a bunch of different tasks throughout the day is a bunch of time wasting that people aren’t gonna have a lot of patience for.
What I see more and more people doing, especially a lot of the thought leaders who are teaching and educating others in our space right now, is you have your centralized tools like Claude Desktop or Cursor. And you’re connecting those to everything else. You’re either bringing things into it, like context, data, whatever else, or in some cases, you’re having it interact with the other tools that you maybe use, like a Jira or a GitHub.
Maybe I’m working on a PRD and I’m not sure if something got completed or if something actually shipped, or if that was in the backlog. Instead of opening a tab and checking Jira or looking through the PRs, I can actually say, “Hey, can you check and see if this issue on this project was completed? I think it was, but I’m not sure off the top of my head.” Or, “Hey, I know there was a PR waiting for this, did that ever get pushed? We’ve been waiting for a bug fix in prod or something.”
I don’t have to leave the tool that I’m in right now to get an answer for that, which is nice. I can just kind of stay in my flow state without changing the UI for anything, which is pretty cool.
There’s other things that I think happen outside of the OS. Maybe some design concepts or brainstorming. I’ll use tools for research, and we were talking earlier about not all the research you get is something you wanna bring into your actual project, so there’s maybe an intentional layer of abstraction.
But most of the work I’m doing is happening in this middle area, between either Cursor or Claude Desktop.
Demo: Cursor and MCP Connections (4:49)
Aakash: Awesome. Can you demo for us how this all gets put together?
Mike: Yeah, for sure. So, let’s start with a really simple, more technical task. If you’re not familiar with Cursor, there’s kind of three areas to this. There’s, this is collapsed right now, but this is the Cursor agent. You can see conversations I’ve had about how to troubleshoot or different things that were happening in the past. You can collapse that down.
This middle area, you have two things: your file that you’re editing or looking at, which you don’t actually even have to have up. Most of the time, I don’t. I have Cursor settings, which is telling me I have some MCP configuration errors. We’ll talk about this. And then down here is terminal, where I use to interact with Claude Code.
The left side is your file search. I don’t really actually use the IDE like most developers would for the files. I’m using Cursor to connect tools to either the agent that I showed on the right side or I’m strictly using the terminal inside Cursor, so that I can navigate between files and interact with Claude Code.
For example, this was a case where some of the content for all the wedding planning tasks and our thing was in a spreadsheet and it was driving me crazy, so I wanna start moving it into an actual content management system.
I can ask whenever I start a new session, for example, for the most recent changes. Sanity is the CMS. I have the MCP connected already.
Aakash: And for people who don’t know, CMS is like a content management system. It’s often like the back end for a blog or something like that.
Mike: Yep, and MCP is Model Context Protocol, which is letting you connect these apps and tools like Claude Code or Cursor to the other tools that you’re using, without having to do most of the coding and integration yourself.
So, it’s gonna check basically, it has an API key that proves that I can access the information. It has kind of its own map to the API so it can go and find the information that I’m asking for. It’s gonna think for a second as it does this and try to get me an answer.
[Demo shows Cursor connecting to Sanity CMS via MCP and retrieving recent changes]
So, it’s looking, it found 10 total documents, it’s checking within Sanity itself, and it’s showing me changes to the actual task that we have at our planning thing. I could go and ask it more questions about the task structure or things that we’re working on, or I could go ahead and ask it to make changes, like if I wanted to add something to it, like a task for car rentals for the honeymoon.
Instead of me having to go in Sanity and spin up—it runs the studio locally, it’s pretty technical to get into—and then going in and adding new tasks manually, it’s just gonna do that for me. So, creating wedding planning test titled “rent a car for your honeymoon.” It’s creating that document, which means creating the content, and then that goes into the tool. I don’t have to go to the tool.
So, this is a not super exciting task, but if you’re editing different things and managing them, you can interact with a lot of these tools that have MCPs or even APIs without having to leave your home base. That’s conceptually how that part works.
The Power of MCP (9:17)
Aakash: Very cool. So, this is the power of MCP connecting to whatever other tools you have and using the operating system as the home base to connect to things.
Mike: Yup, and that’s a really simple version of it. I’ve done this with Supabase for structuring database schema or even doing migrations on the fly. I’ve done it with hosting a little app. If you end up vibe coding something you want to launch live, Render has a nice MCP. I didn’t have to know anything about DevOps to actually get it up and hosted and get it running, which is pretty nice.
Aakash: Very cool. So, if we go back here, that was kind of connecting to technical tools via MCP and bringing that into Cursor. You can do the same thing in Claude Desktop if you want. What I just did in the IDE we could connect the Sanity MCP to Claude Desktop. I don’t have it on mine right now because it would be duplicative, but you can, if you’re more comfortable with this interface.
You can do the same kind of thing. You can go into Claude settings, go to connectors, and then you can connect all these different tools they support out of the box. And then you can also go and configure your own if you want to as well, which just is a simple text file.
[Sponsor break: Linear – highlighting how OpenAI, Perplexity, and Vercel use it, agents functionality, and free year offer]
Custom Instructions and Project Setup (12:14)
Aakash: And I noticed you had some custom instructions in your Claude. What is your favorite custom instructions?
Mike: At the project level, mostly what I do—so we have a bunch of different initiatives, so I try to frame the context of the project, rather than everything in the entire application. It’s all related in this case, like this is my own account, so it has different projects in here.
This is the wedding planning app specifically. Media Network is advertising and partnerships that we have across ecom and everything else. So, having specifically, this is the app we’re talking about, this is the app we’re working on, or this is the ecosystem we’re talking about, and this is the specific lens we’re working on, or David’s Bridal—I don’t have one in there, but that’s more of overall architecture, tooling, internal things and initiatives.
I try to set that for the specific ones that need to be differentiated. And then hope that it can maintain, but I also use the memory MCP so that it has those relationships that it picks up over time to the different ideas, cause sometimes I’ll have to say, “Hey, I need you to actually look across project knowledge,” or “I’ll need you to pull information from these two different projects to do what I need you to do right now too.”
Aakash: Oh, cool, very cool.
Mike: And they’re rolling out memory features with Claude, which I think helps with that too, but having the memory MCP was the thing that’s let me do that for the past 6 to 9 months or however long I’ve been using it now.
Claude vs. ChatGPT (13:39)
Aakash: And some people might be asking, we’ve been talking a lot about Claude, Claude, Claude, why not ChatGPT?
Mike: Anthropic ruled. I mean, they created the MCP framework and have supported it best. Claude Desktop had it out of the gate, and I think supports it well. They have their own team that rolls out supported MCPs and things like that too.
OpenAI has MCP support, I think via API and maybe they’re rolling it out in the chat app now. I was pretty happy paying for both or at least using some GPT with 4.0. I think the last two model releases, I felt like got more towards lazy and less useful, so I just don’t spend as much time in there.
I feel like they started rolling out connectors in the consumer app. If you go in via UI you can do it, so you could do all this stuff there if you want to. I just moved away from that because I think the type of work that I do is much deeper and requires a little bit more connecting and tooling. I just feel like Claude is more reliable for me.
Aakash: Yeah, I think Claude is also just a better writer and PMs are doing a lot of writing, so generally, you’re gonna prefer Claude. But if you have ChatGPT at your work, they’re starting to roll this out so you can do a similar sort of stack with ChatGPT at the center.
Mike: I have a post somewhere about a deep dive on MCP on my newsletter, and there’s a handful of desktop apps and open-source projects that you can run that kind of act as a gateway, so they’ll let you connect MCPs and then they’ll connect to whichever models, so you can jump between them.
I have one called, I think it’s Fire, but it starts with a 5 that I’ve been playing with that is pretty cool, but it’s like a project to get set up and going. But if you don’t have the tools you want or they don’t have the capabilities you want, you can actually kind of hotwire MCPs in a weird way, which is interesting.
Design Concepts and Figma Make (15:21)
Aakash: Cool. What’s next?
Mike: Let’s look at design concepts cause you and I were doing this earlier and I thought this is a simple but pretty interesting use case when we’re trying to conceptualize this.
I opened up this, which is similar to what we just were showing on the other screen in Canva. And the bad thing about this is it’s flat. So, it’s just an image, we can’t edit it or tweak it without a lot of prompting and back and forth.
So, what I did, and I do this with some product design work and development too, I just dropped it into Figma. I’ll zoom in on this so we can see it a little bit better. I just did a blank file for this, but what I’m gonna do is if I right click this frame, you get this “send to” option. And you can go to Figma Make.
I’m sure you can just go to Figma Make and upload that too, but the cool thing about doing it that way is you can do this with any of the designs that you’re working on with your team anyway, so it’s gonna send that. I think it picks up, in this case there’s not very many layers, but if you’re sending a design file that has layers, it’s also gonna pick that up and make it better.
Prompt: “Help me make this an interactive visual that I can edit and rearrange.”
I don’t really trust Figma Make to code anything useful. I haven’t seen it functionally finish anything to the level that I’m like, “Oh yeah, I can move this into GitHub and work from here” so far. But I think it has a unique advantage in that if you’re working on your product files that are structured in your design system, it’s easy to move them in here, and you can visualize maybe a specific flow for a user, a variation or a new feature, and see how you might integrate that, and then bring that back to the designer.
Cause they’re probably working on other things anyway and be like, “Hey, conceptually, functionally, this is what I was going for, what do you think?” And when you bring it back, I’ll show you, you just copy it and it’s fully layered and everything, so they can take the different pieces that they want from that and just add it into the design, create the new model, create the different screen or whatever, and bring it back into the same file.
So, it’s a big time saver for designers.
Aakash: So if you’re thinking about which AI prototyping tool is most designer friendly, Figma Make is up there, is that the AI prototyping tool you generally recommend?
Mike: I don’t even use it for prototyping. I think I use it mostly for design variation. If I’m, if in my head, and maybe you’ll relate to this as a PM, if I’m like, “Oh, we really need to see this edge case, this error handling, this state for this functionally,” and I also am asking the designer or the design team to start working on the next thing because we’re trying to move pretty quickly, but I wanna have it for the developers.
I’ll actually take the design we have and throw it into here. And then make it, usually it’s interactive at some level. So, what you can do in here, which is nice, is you can edit these, and then you can change the state of it.
[Demo continues showing how to copy design from Figma Make back into Figma with all layers intact]
The nice thing about it is you get different UI elements, you get the states that they’re in or whatever. So it’s just kind of a nice shortcut to get the visual state that you want without having to sit down and be like, “Here’s 20 screens I need you to work on for this week that are all edge cases,” the functionality we already agreed on kind of thing.
Aakash: Yup. And what do you recommend for AI prototyping if not Figma Make?
Google AI Studio for Prototyping (20:14)
Mike: Personally, I think the last 3 months, I’ve been blown away by AI Studio. And what I mean by this, this is Google’s DeepMind developer-focused or AI developer-focused team. AI Studio is like a play space for developers. And they’ve done a good job of also not alienating non-developers. They’re really encouraging people to get in here.
But this is where you can test the newest models, you can get an API key really quickly and for free to start playing around with some of these things. So, if you’re like, “Oh man, they just made a huge announcement, I wanna go play with this thing,” this is kind of the ideal place to go.
On a personal note, I also feel like in terms of how they manage chat context and memory in the actual experience, this is infinitely better than Gemini. I’ll go in there and ask for an image, and then if I ask for a variation of the image, it’ll just give me the same thing again. It doesn’t know that I’m asking reference the first thing and give me this edit to it.
There’s a lot of little contextual things like that in the Gemini app that I get frustrated with, that don’t exist here. I think the developer team just built a better experience for developers than the Gemini app did for consumers.
If you’re in AI Studio and you got to build, you can one-shot. They have a bunch of pre-selected functionality or example apps. My team was asking me to prompt and create a bunch of—we had some dresses that were on sale and come up with Halloween outfits, for example. Instead of me manually doing that and trying to understand all the different dress subtleties and stuff, which I’m not great at, I was like, “Can we just spin up this app real quick?”
It wasn’t a consumer facing app in this case. It was just like, can we create a costume from this? We just use this to generate some marketing images that we launched. I did this in literally 10 minutes because somebody was messaging me in Slack like, “Hey, this is my prompt. Here’s an image I’m trying to go for, but I keep running into X issue.” I’m like, “Let me just put it in here quick,” and it was usable in 10 minutes.
Aakash: Got it. So, Google AI Studio for prototyping. How does that fit into the overall operating system?
Mike: What I tend to do is go into build and work on some of these rough concepts. Here’s just a general notebook type app I was playing around with. It’s infinite canvas, but it looks like a notebook.
What you can do once you get to a point where like, “OK, I want to play with this more,” I’ll either push it to GitHub or I’ll deploy it to Cloud Run or both. I’ll download the zip, push it to GitHub, and then I’ll switch to Cursor and just open it up and start working inside Cursor to edit from then on, get the local setup and start running it.
That way, I have more of my normal workflow. I check my browser and see the app running locally, I can experience it, and then I go back in here and give it feedback or start describing the changes that I wanna make to it or whatever integrations that I wanna do.
It’s just that back and forth, more like a true developer workflow at that point. But it has to be at a certain level of quality before I pull it into my main operating system spaces.
Knowledge and Progress: Confluence and Figma Integration (23:55)
Aakash: Makes sense. What’s next in the operating system?
Mike: Let’s see, I think one of my favorites has been progress and knowledge. We’ll combine the two in this case.
I’m just gonna go to a new chat inside one of my projects. I’m just gonna ask what do my Confluence docs say about the vision board, which is just a feature that we launched a while ago in the MVP.
So, this is gonna use the Atlassian MCP which uses that new Rovo AI search, and it’s gonna go through and check to see specifically. I’m hoping it’s smart enough to know I’m in the planner, and so it should look in that space and Confluence and find my doc. It’s going to check for my requirements documents. And then it’s gonna give me some kind of summary for this.
And then what I’ll do actually, I grab a Figma link at the same time. It should use the Figma MCP. And I’m gonna ask it to compare.
Aakash: This is some seriously connected workflows, so I haven’t been working in this connected of a way, and now I’m really seeing the benefit of the operating system concept.
Mike: Yeah, I think the trade-off—you were watching the screen as it was thinking and finding this information and pulling that in for me, cause this is something I’ve already written. It already exists. I just didn’t want to open Confluence and try to find that page and copy and paste the specifics.
But then what I’m gonna do on my other screen while I was loading is go back to my MVP designs. And I’m gonna grab just the link. So, the Figma MCP is kind of weird. It uses a lot of context and usually maxes out Claude, which is a complaint I have.
But if you give it a specific URL to a frame, so if you go into Figma and just kind of what I did earlier, but instead of saying “send to Make,” you just grab a specific share URL. You can go to share and copy URL or “copy as” they changed it now. I don’t know why they did that, but copy link to selection.
Now I’m gonna ask if I’m missing anything or if we missed anything from the MVP design comparison. And this is probably stuff I’ve already done. I’ve already gone through and I’m like, “Yeah, we’re not gonna do this, we are gonna do this.” Whatever.
But it’s gonna go to that specific frame, and I think the way the Figma MCP typically works is it’ll grab a screenshot of it instead of trying to pull in all that information. And if you saw it, ask me to allow it. So depending on what MCPs you set up, you can also set them so that they ask for permission every time they do something, or they don’t ask for permission to find content or read, but they do ask for permission before they edit anything or push any changes, which is how I have it set up for the most part.
Aakash: Nice. So you’re using the Figma MCP it sounds like very regularly.
Mike: It depends on—if you think about product life cycle—if we’re working on something new or we’re working on our version 2.0 set, there’s different waves where I’ll get in Figma more. I spend a lot of time in Figma still too, working with design on getting those things done. And then when I’m working on refining my requirements before we get to the dev team or when we’re working on first pie in the sky designs, I’ll definitely be pulling that into Claude or Claude Desktop probably more often.
Aakash: Got it. And that’s what I think is nice about the operating system approach is your tool stack can be very composable based on what you need.
Mike: And so, the whole pricing strategy thing from a SaaS standpoint is tricky, because it might be, I’m gonna pay for what I use. If more people move in this direction, that’s kind of like a developer API usage-based billing for a lot of these services. And it might be very fluid over time. It might be a specific use case where I need to use—I have a marketing type app that we’re spinning up that uses Supabase, and it uses Google Cloud and it uses something else, or it might be something internally that has a totally different stack, and I might need to be able to fluidly jump between them.
[Demo shows gap analysis results between Confluence docs and Figma designs]
So here it’s given me a pretty good gap analysis. The layout, the icons, everything. Wedding style doesn’t display, which is true, we actually don’t display wedding style from the onboarding process on the vision board at all. The documents say show bridesmaid only, which is true. I think we had wedding party versus we shifted to bridesmaid afterwards, so it’s catching pretty specific things that are little discrepancies in design.
It’s actually a huge lifesaver. Normally you’re pulling up the PRD, you’re looking at the design, or you just missed this altogether and second-guessing yourself. In this case, I’m still gonna second-guess myself and double-check it, but at least I’ll have some context and idea of where to point my time.
And if people are like, “You can’t trust it.” I think I’ve done both. We’ve been around long enough to have done both manually. I trust this as much as I trust myself going through a massive list and comparing every single module and things side by side. Something might get missed or something might not get missed, and it’s just nice to have a second brain to rely on to gut check yourself.
Aakash: Yes, that’s my favorite thing about AI is it’s like my second brain. Sometimes I wouldn’t have even asked a person for it, but I don’t feel as bad asking AI.
[Sponsor break: Linear]
Research and Context Gathering (30:51)
Aakash: Yeah, so we talked a little bit about prototyping, we have kind of knowledge reference pulling from Confluence and then gut checking it against Figma. We had an example where we took a generated image and created a prototype and pulled it into Figma so you could play around with different states. We haven’t talked about research yet, which we probably should have done first, but we can do that next.
Mike: Does that sound good?
Aakash: What do you use research for? Is this setting up your project context or what are the right points in the life cycle to be using this?
Mike: Yeah, so when I say research, I think ultimately what I tend to mean is context gathering. There’s a couple of different ways I usually go about this, and I would say I pair—I used to use Perplexity a lot more, because I think the good brain exercise is what questions actually matter to answer when you’re working on something new.
Everybody knows Perplexity kind of works: you ask a question, you get some answers. The use case that I think I had up here right now is this one that I used to test Reforge Build when it was in beta mode. It was an app that maps out paranormal sightings. My kids are into this podcast and stuff like this, but functionally, it’s interesting because it was a community app.
So people need to be able to post sightings. There needs to be a governance or a workflow, like a queue where you can review or flag if it seems like it’s spam or something like that. You need to be able to plot these things out on a map, and then there was a time dimension of how do they play out over time and being able to filter that.
Functionally from an app standpoint, it sounds really silly, but there’s a lot of really cool interactions that are there. If you’re testing a vibe coding tool, this kind of is a good way to stress test it.
Yeah, so in this case, I was like, well, I don’t—I know what my kids tell me about this stuff, but I went through and I was like, this is all I gave it.
Manus for Deep Research (32:49)
Mike: Manus is fantastic at this, by the way. I don’t feel like with agent mode on GPT or any of the other tools like even research in Claude, it tends to max out pretty quickly and burn through a lot of your usage.
So, this is nice, it’ll just run independently and let you know when it’s done, and it gives you its entire trace of everything it did. All these different sources. And then, when I say context gathering or research, what I also love about this is I can ask for specific deliverables and it’s also gonna show me.
So, it did a sample set CSV, a combined CSV with different sites. It gave a data sources report and a markdown file, so I know where that data came from. It gave a quick start guide to be able to use it, and then it’s giving me a more human summary, but I can access all these files that it gave me.
So, even if I really just wanted one of these deliverables, I can access every piece of them independently and use them elsewhere, which I really like about this tool. So maybe it gave me a specific set that I thought was more useful than that.
What I think I actually did in this case, when I was doing this is I pulled this information in, and then I gave all of it, like the raw information from that to Claude. And I asked it to give me, so I could test these tools, product requirements, user research and personas for it, and then a technical strategy or approach for it.
So, I have these 3 different artifacts now in this joke project and I can pull these down, and that’s how I typically would go into AI Studio like we talked about before. And then give it those files of the requirements to build something new.
I did that for Reforge and it did a great job with the design. Or you can take those files and then work on actual requirements and refine them and then even build them. If you wanted to go straight to Cursor, you could, or go into Figma and try to work on some design concepts for it or something from that Make screen. It just depends on how you want to go about doing it.
This is specific for me to test vibe coding apps.
Aakash: That’s where I would—Reforge Build fall in your top three AI prototyping apps then, in terms of where it got to from, “Hey, is this functional? Does it match my requirements?” And is it actually from a product taste standpoint, does it match some of the details I put in here about UX?
Mike: I think it did a good job. I think Studio tends to be still where I go to get it to a functional point fast, easy, reliably, and then I can take that code and run with it. Or I can host it there.
And then I would say, I use Replit for a long time, probably early on before they dealt with a lot of issues, and I just got really frustrated with that agent lying to me about like, “Oh, I fixed this,” or, “Hey, I implemented this new thing,” and I spent more of my time being like, “Why did you implement a new thing? We had that working yesterday.”
And then you go back into the code and you’re like, we were using Python and now we’re using Python and there’s a node implementation that are fighting each other and the whole thing’s broken or something. So I think you can use those tools to get a concept up and running, but I feel like they’re designed to make you feel like you’re progressing when maybe you aren’t.
And so, if you’re more technical or more capable, maybe it’s better to jump into something that lets you transition from concept to code a little bit better.
Aakash: Yeah, so Manus over ChatGPT agent mode, Reforge Build and Google AI Studio over something like a Replit.
Mike: Yeah, and this is even stuff for date nights. You know, here’s what we’re into, here’s the weather, here’s where we live. I want something within an hour drive out, 1 hour drive back, something like that, so we’re not gone too long. You can give it all these things and it’ll just run and bring back all the research and different options and things like that too.
So, the big reasons I jumped to this are, A, it’s just good at running independently and being thorough, and B, being able to get multiple assets. This is a lot of different pieces to prove that it did the thinking. Like here’s research on the OpenAI SDK that they rolled out to build an app inside ChatGPT. And I said, “Make sure you review this documentation for the technical approach or whatever.” And it does. It has the details, but you can trace it back, which I really like.
Manus vs. Claude Research Mode (37:05)
Aakash: Very cool. When do you use Manus versus Claude regular?
Mike: I still feel like my frustration with Claude is chat length limits, and then usage limits, even though for work, I’m on the max plan. 200 bucks a month or whatever, and I don’t really hit those as much anymore. But if you actually turn on research mode, I feel like it just runs and runs and it doesn’t do a great job of showing its work or its thinking, like Manus does.
So, if I have something quick I’m working on in here, I don’t really use research in here very much, to be honest with you, because I feel like it just burns through tokens.
And the other reason I do that is because if Manus gets something wrong and I don’t agree with it or it finds information that I don’t like, I can pick and choose. Here’s prep for this podcast, and I gave it the notion content and was like, “Go through blog posts or here’s what I’m thinking about this, how should we position it or how are we thinking about it.”
I had some decent ideas in here. This is just my notes, but I can pick and choose what I actually bring into my core operating system, so that it doesn’t give it the wrong idea or it doesn’t start—cause the bad thing about a lot of the LLMs is they’ll pick and choose what’s in the memory to really anchor themselves to and then they become common assumptions or common beliefs. Everything it responds to is based on a common belief that you gave it.
So, if you’re not careful about what you feed it, I feel like you end up with the equivalent of a conspiracy theorist LLM partner who is running with random ideas that maybe aren’t as important as some of the core information you gave it.
The Composable Operating System (39:10)
Aakash: Cool. What else should we understand from the operating system perspective?
Mike: I think conceptually, if you hear the word composable in technical architecture a lot or Martech, where you can kind of—it’s kind of like modularity with furniture or whatever—you can kind of pick and choose how you want it laid out or how it’s configured to meet your needs at that time.
These are all different use cases and things that you deal with throughout your given week or month or whatever. I think there’s also things that come up that are—you know, I have a specific project or specific asks that has unique requirements.
I was doing something in the domain space tied to this next round of applications for new TLD strings or top-level domains, which would be like a .link, for example, or .com or whatever. And when you apply for these, there’s a huge cost to apply for them, and there’s a really big list of evaluation criteria that goes into it.
Is there a patent or a trademark already out on the string that you’re applying for? Is there cultural sensitivity or a significant geographic region that might have stake to that first? Are there any language conflicts where in one language it sounds fine, and a different one is offensive?
So there’s this long list, and for that, it was like, you know, I wouldn’t say that the API for the patent database is something I’m going to use again. But at the time, I can go get that, I can connect that via Claude Code, for example, to the application that I’m working on or the process, the agent that I’m trying to build, and that’s a one-time thing.
Now, that wasn’t even an MCP in that case, I just had to go get the API key to be able to use it. But it’s the same process either way. With the MCP you configure it and you put in your API key. With this one, you give the API key to your coding agent or Claude Code, and it handles that part of it for you.
So, I’d say as a big takeaway, there’s lots of different things you get asked to do in any given day or week. And the right set of tools is not constrained to the ones that you can log into and use. There’s lots of different solutions out there for different things that are sometimes free. Very generous free tier plans on APIs or things that don’t cost money at all, that can make you more effective or make the solution that you’re working on more valuable.
Aakash: That makes sense. Yeah, 100%. I think this is an incredibly powerful way to work.
Communications and Email in the OS (41:36)
Aakash: If we go back to the nano banana image, I think there was also something you had put in there around communications and email. How do we handle that side of it in the operating system?
Mike: Yeah, I think there’s MCPs for Gmail, there’s MCPs for Slack, there’s MCPs for calendar connector. I actually think even in Claude, if we go to the browser settings, maybe there’s native connectors just for this.
Yeah, so they do this via Claude right now. This is my other account. If I wanna connect my calendar or something and ask it what I have coming up this week, or if I wanna connect my Gmail and ask it to look up when did we schedule the podcast interview with Aakash, it can do all those things for you. You just have to give it access to be able to do that.
We can even do this right now.
Aakash: I’ve never used this use case, and I consider myself a Claude super user. I’m having so many epiphanies in this episode.
Mike: Well, I have a hard time thinking of a lot of the automation like n8n or even Zapier sometimes. What do I do often enough that I feel like there’s a direct use case that I’d wanna configure something or I’d wanna be able to do something?
[Demo shows connecting Google Drive to Claude and searching for documents]
So, for this one, for example, the connector will let you add manually, and you can pull it in, but I think the connector should also search my Google Drive for me unless that needs to be configured separately as an MCP.
Aakash: Wow. So Drive is so hard to search. Even though Google is a search company, I don’t think its search is that good, so being able to use Claude is a huge win.
Mike: Yeah, so this is the most recent one. This is what I talked about with the super guys about planning that vacation and when the memory MCP ended up being really useful to me. It was like, “Oh, your 5 year old likes crystals. There’s a place you can go mine for crystals here. It’s 30 bucks for a bucket, and it’s in between your stay in Lake Tahoe, so you don’t have to drive 4 hours straight. You can break it up and get some exercise.”
But yeah, you can get all this context, you can view the full document or pull it in or whatever. So, it’s pretty easy to use that. They’re MCPs, they’re calling them connectors cause I think MCP is probably intimidating for less technical people, but they finally made the update. At first, it was just on desktop and you had to configure it there. Now they’re running it in browser and Claude, which is nice.
Aakash: So let’s say you weren’t on the Google stack, you’re stuck in Microsoft. Could you do this there too?
Mike: Let’s see, I feel like they do have an Outlook connector.
[Searches through available connectors]
Yeah, so they’ve got Notion, Figma. Let’s see, Microsoft. MS Learn Docs. No Outlook as far as I can see, but I don’t know if that’s a common one for Cursor.
Microsoft is a little bit behind on getting this stuff out the door. You got Russell, PostHog, Netlify. All these specific ones, you can connect to ClickUp.
Chrome Dev Tools is sweet, actually, especially now that the Cursor agent has browser use, so it can go in there, it can pull DevTools report, and it can actually test new functionality that you’re working on in your local instance. Not that I totally trust it blindly for testing anything, but it’s a nice extra step instead of like, go to the browser, copy the console error, paste it back in.
But yeah, this is a long list of all sorts of different tools. Firecrawl is a developer tool, but I’ve used it to pull data, shape data and shape content for different tools and stuff too, even pulling our color swatches down, honestly. It’s just easier.
Amplitude, if you have product data in there, you can pull that in. Cause sometimes even if I’m in a coding space, I’ll think in that coding space. What does the data say? And you just shift Claude Code to plan mode instead or whatever the case is.
Licenses and Tool Stack (47:04)
Aakash: Awesome. So, this is the operating system for AI native PMs. One thing that sticks out to me is this is a lot of AI tool licenses, a lot of IT requests, a lot of money to be spending. Can you just walk through what are the different licenses that teams should be getting access to their teams, or product leaders should be getting access for their PMs here?
Mike: Yeah, I think the big ones, or what level of access you give to—and depending on how technical your PMs are—maybe Jira is enough, or if you’re in Linear or whatever the case is.
But I think read access for AI tools versus write, if you’re conservative. I think at a personal level, some of these are free. And some of them are being done separately like using the Gemini app. I use that through my personal email, which is on Workspace, and I play around with that. Or AI Studio is free with any Google account.
And as long as your team’s aware of like, “Hey, let’s not take a bunch of IP or private information or product specific things that are internal and bring it in there,” you can still conceptually get a lot done outside. If you’re just trying to come up with ideas and then bring them into your tools like Figma for example, or set up your own sample app that’s an example or use case for what you’re trying to do before you spin it up internally or have a developer start working on it, or even start building up your own private repo for your own internal apps kind of thing.
So I think that’s how I tend to navigate it. 90% of some of this thought and planning work doesn’t actually need a ton of private context or private data or company information. But if you’re actually working out of this day to day, I think your core tools—so, if you have Jira, you have Confluence, or you have Notion and you have Linear or whatever the case is.
Easy first step for PMs to start working this way is just connect it to where the actual knowledge they care about lives. What are my requirements, how can I get a pulse on dev progress, little things like that.
Aakash: OK, so to summarize the licenses you should be getting for your team, sounds like we can do the $20 a month Manus plan, probably the $100 or $200 a month Claude plan, potentially the free version of Cursor, is that right?
Mike: Yeah, yeah, I think you can, if especially if you’re paying higher end for Claude, I don’t think you actually need the Cursor agent if you’re not doing actual dev work. And even then you can use Claude Code for most of it.
And I’ll say for the Claude paid plan, the way that we’re handling this internally even, not just in product, it’s like, “Why don’t you start and show me that you’re using it?” That you’re getting some value out of it on a $20 a month plan or whatever the case is. And then if you show me you’re doing it, and it’s helping you and you’re leveraging it in the way that’s moving things forward, happy to bump it up.
And then if you start hitting limits on that, you’re actually using it, we can bump it up again. And that way, you’re not just giving everybody a blanket $200 a month plan that can add up pretty quickly if they’re not really using it, but you have them building habits that are using it productively and then $200 a month is nothing, compared to how many hours of work it’s saving them at that point.
Aakash: Nice. And then they might need access to the pro plan of Google AI Studio and nano banana Pro if they’re doing a lot of generations, so potentially $20 a month. So, the minimum viable AI native PM stack, what I’m hearing, $20 a month for Claude, $20 a month for Manus, $20 a month for the Gemini package. That’s basically it, right?
Mike: Yeah, and I would even argue at some level, if you’re working with IT, you can—everything in AI Studio, and that’s part of the reason I like it too, if you’re just connected to the cloud account and you can say use this API key. So your IT team could go in there and spin up a project and give an API key that specifically has access for usage-based billing and give you permission.
And then any of the stuff you do in there can just be billed to that, and then it is usage-based, and you’re getting wholesale rates essentially based on usage instead of paying another software subscription. It’s like, if you want Gemini in the app, you can.
But at this point I almost prefer like, oh I have an API key used specifically for testing. And Nano Banana 25 is like 4 cents an image so I could go crazy someday for a week or whatever and spend 20 bucks and then I might not use it again for 3 weeks and then I’m not on the hook for another subs—or 3 months and I’m not paying a monthly subscription for nothing. I think that’s probably the way a lot of this will move.
You just have to have the connector to your core tools where your teams are working to be able to move that knowledge in and out.
Making the Case for AI Tools (51:35)
Aakash: Sweet. So let’s say you’re a PM and you’re stuck where you don’t have access to Cursor, you don’t have access to Claude or Manus. How do you go make that case to product leadership and to IT to get access?
Mike: Yeah, I mean, I’m pretty bullish at this point. I think everybody in leadership, executive level—I mean, I work at a—so David’s is a 75 year old wedding dress retailer, OK? And so, we’ve got a lot of old beliefs and processes and things. We’ve got a lot of old tools and everything like that.
But when I talk to leadership and I say, “Hey, this is what my team has been able to get done at this velocity, and this is—our team is one quarter of the size of most other teams that are working on this right now. Here’s how we’re doing it.” Or if I say, “Here’s what this investment means, and ultimately what the unlock is for us to be able to get done this year,” you know, you have these different pain points, these different problems.
And if you can show and build the trust of shipping—I don’t know anything about shipping and logistics and things like that. But I took some of the data and reporting and things like that. I’m like, “Here’s 3 different ideas that might be causing us excessive overhead on some of this.” And they’re like, “Oh yeah, we hadn’t thought about problem B.” Like, OK, that took me 10 minutes, so imagine what a team of me could do, if they’re just asking smart questions.
They tend not to push back on that. If they do, this is gonna sound a little harsh, but I would be fairly concerned about any organization pushing back on people trying to use AI to be more productive and more impactful.
Aakash: Yeah, sadly, a lot of PMs I talked to are in that world. They’re living, you know, whether it’s automotive or finance or healthcare, where they seem to be living 5 years in the past. But I think what you said there of maybe you create a personal use case or you show the personal productivity or you show how it can get you more velocity, and maybe you chip away one tool at a time. Maybe you don’t hit the whole operating system right now.
I think in particular Manus might be one that you could potentially wait on, but you really can’t wait on Claude. So you start with, OK, I need Claude, and then you can move your way up. And if you’re stuck on a particular stack, like a lot of people I talked to, it’s just that they have access to ChatGPT.
Then I think what you do is you say, OK, let me try to do the best I can with ChatGPT, and then let me show them with my personal work how Claude can upgrade that and then bring that to them. Does that sound right?
Mike: Yeah, yeah, I think it’s like, show that you’re getting value from the technology first. Get buy-in on the fact that there’s an opportunity there to move faster, increase your impact overall, to solve hard problems that maybe you weren’t able to solve or you have different constraints around or whatever.
I think organizational buy-in on the value of the technology is step one. Cause a lot of people are stuck in, “Yeah, it seems like a cool toy, but how are we gonna use it on the business level? We have lots of other things that are more important right now.” Like there’s always lots of things.
If you show, “Oh, it accelerates us,” “Oh, it helps us tackle these hard things,” or “Oh it’s unlocking talent”—you have people sitting there who are capable of more and you’re just giving them menial jobs or a lot of overhead work, and they’re able to work around that, that there’s immediate value.
So, from there, you can say, “OK, from this tool, here’s what I can do, here’s what I’ve been able to do, and here’s where I’m hitting a wall. Here’s what I could do if I had X tool instead of Y tool.”
And I think at least if you have that case built up underneath, it’s easy to justify. And it’s not too different from a feature roadmap or a product roadmap or a specific feature you’re trying to pitch. Did the users see value in the first thing that you shipped? What issues are they running into that we need to help them navigate? What are our goals that we’re not hitting that we need to move past or what is the unlock for this?
So, that’s muscle memory that all PMs have. I think the pitch is ultimately the same. It’s just me getting things done instead of users having this experience or us hitting this metric.
Using AI Throughout the PM Lifecycle (55:33)
Aakash: Amazing. And then finally, before we go, can you just walk through the life cycle of PM and when they should be using these various tools?
Mike: Yeah, I mean, I’m in the weird, extreme early adopters and testing and playing with this stuff. My default is you should be using them the entire life cycle.
I think there’s a huge value or there’s a ton of value in your upfront research and validation. And then that’s everything from what we showed earlier about pulling a bunch of data in. What about this consumer segment versus this one? What about this use case? Is it as relevant to our target audience as it might be to this other audience that’s kind of tangential?
Getting actual numbers and finding what is or isn’t available, what competitors are already in the space, who’s targeting. Based on the competitors that are here and their recent marketing and the recent releases, who is on a trajectory to go after the same thing that we’re targeting? Who’s gonna be neck and neck with us?
Or even, “OK, I’ve done my research, I’ve talked to customers. Here’s scripts from interviews, here’s our whole justification. Tear this thing apart. Red team this thing. Tell me what I’m missing, be a skeptic. Tell me what you think is wrong about it. Give me every possible angle so that I’m prepared to defend it and I actually believe it at the end of the day.”
And I think upfront, everything, and then throughout the process, it’s every interaction. Am I checking my design to make sure it meets requirements like we talked about? Am I using it to shape the actual tickets that I’m putting into Jira or Linear or whatever, to make sure they’re very clear and very specific? Did I reference the GitHub repo for the codebase to actually make sure that the ticket wasn’t conflicting with how the thing is currently built, if it’s an add-on feature or something?
There’s a lot of little things you can do that were not expected before, but are very easy to do now that add a lot of clarity and value for the entire team. So, I think life cycle-wise, even if you’re branching into prototyping and things like that, I don’t see a place where AI actually hurts unless you don’t know what you’re doing with it.
Biggest Mistakes PMs Make (57:39)
Aakash: Amazing. And along this life cycle, what are the biggest mistakes PMs make with AI tools?
Mike: Garbage in, garbage out still resonates with me. I think overprompting is one that I think backfires often. Really deep, really structured prompts that are meant to be reused and systematic like a copy paste without understanding the subtleties of every time, like the things that you’re doing differently, unless it really is just a repetitive task.
Like generating requirements, I don’t feel like I need a prompt structure to do that, cause it depends on what I’m starting with or where I’m at or what stage of thinking I’m in.
I think the data you pull, for example, I think laziness with AI—you put a prompt in tool A like Manus, you downloaded the files and you uploaded them without reading them in, say a Gemini and you prototype something out—you’re probably gonna get something that’s maybe along the lines of what you’re thinking but not very intentional.
Where you’re kind of proving that that level of product management isn’t valuable anymore. Taste is incredibly important. I think intuition is incredibly important. Being able to be like, being able to look at something and say I did this for 1 reason is 10x more valuable than like I put this dog together.
So being able to defend it and being really intentional about the end state of it, I think is—some people tend to let go of that too early on, especially more people that are junior that I’ve worked with tend to be like, “Hey, I put it in here like you said. I use AI. I put this doc together.” And then I read it and it doesn’t make any sense.
Like, “Yeah, you probably did use it, but none of this represents the information that we have or the interviews that we sat in. Did you use any of those things or not?” And they just don’t check it or they don’t question it. You have to be kind of skeptical with all of it, I think.
Closing Thoughts (59:22)
Aakash: Amazing. This has been a master class. Truly one of the episodes that I’ve had the most epiphanies on in terms of the gaps I have in how I’m using AI, all of the connectors that you talked about, this concept of an operating system is an incredibly powerful concept.
So if you are a product leader, start to think about, am I giving my team the access to the right tools? Have I taught them about this operating system concept? Should I go through a training on it?
And then if you are a PM, are you using enough connectors? Are you building an operating system? Do you have Claude skills that are helping you? Do you have Claude projects that have the context, and then are you connecting it to all these tools? This is how you actually get the most out of AI.
Nowadays, most of you guys are being evaluated in your performance reviews on how you use AI. This episode has just given you a bunch of keys to get a better rating on that.
Mike, thank you so, so much for dropping the sauce.
Mike: No problem. Thanks for having me.
Aakash: All right, bye everyone. Take care.
I hope you enjoyed that episode. If you could take a moment to double check that you have followed on Apple and Spotify podcasts, subscribed on YouTube, left a rating or review on Apple or Spotify, and commented on YouTube, all these things will help the algorithm distribute the show to more and more people.
As we distribute the show to more people, we can grow the show, improve the quality of the content and the production to get you better insights to stay ahead in your career.
Finally, do check out my bundle at bundle.aakashg.com to get access to 9 AI products for an entire year for free. This includes Dovetail, Mobbin, Linear, Reforge Build, Descript, and many other amazing tools that will help you as an AI product manager or builder, succeed.
I’ll see you in the next episode.