Transcription
Okay, Amir, you're a vibe coder I look up to, and I wanted to have you on the pod. And what are we going to learn today?
First of all, I'm very flattered. Thanks a lot. Um, today I'm going to do an extension of the conversation you recently had, actually, with James, aka the boring marketer. I think he did an excellent job showing how he's using Cursor to do a lot of vibe marketing and how he's using it to grow his business. I want to take a step further and show you how you can essentially use Cursor. Pretty much, it's the interface for work. You can use it for finance, analytics, design, and marketing. And I'm going to show you some workflows and tips on how you can essentially have a team of agents to do specific things for you, especially with MCPs.
I'm intrigued.
You've intrigued me.
You have my attention.
All right. That's great.
[Music]
What's really interesting is—I think I want to take a step back—is with a lot of LLMs, we previously didn't have access to a lot of third-party services or apps. You know, with ChatGPT, we briefly saw connections to like Zapier and custom integrations. Um, but what we didn't see was how you can seamlessly just grab data and interact with that data. And essentially, when Anthropic came out with MCPs—model context protocols—it was an interface for you to be able to interact with those apps. And it's pretty much synonymous with APIs. And the MCPs made it easy for us to be able to access this data, create data, change it, transform it, all through LLMs.
And I currently have a company called Humbolytics. It's an analytics tool that lets you run split testing and CRO. And, you know, I come from a background—I've previously started companies and grouped teams and scaled teams. And my co-founder and I this time around said, "What? You know, how far can we scale this company with AI and Agentic Workflows?" And, you know, let's try to push to the ceiling. And I think AI did a big part in, like, educating us, but also enabled us to do a lot of these workflows. Um, and then when MCPs came out, complete game changer.
Yep, complete game changer.
So, um, you know, with Cursor, we traditionally saw people just building apps with it in development, and I was like, "Okay, I'm spending a lot of time building in Cursor. What if I try to not leave Cursor at all? What if I do my finance, my marketing, my SEO, and my development all in one place?"
Is that possible?
Absolutely. Absolutely.
So I think what we can do actually is start off with finance. Okay. So, one of the biggest things for me is I would love to be able to just get a list of invoices and be able to get an update on my profit and loss, yeah, directly from my accounting app called Zero.
Yep.
So what you can do is you can actually just connect to the Zero MCP and say, "Connect to Zero MCP and list out my profit report for May 2025 or 2024, for example." And what it's going to do is actually going to connect to the Zero MCP and directly pull that information for you. Now you can also update your accounting software directly through here as well. So let's say, for example, you want to do a payroll report or you want to create a quote or an invoice, it'll do that automatically for you. So I can say, "Create a quote for one of my vendors," and what I did here is, you know, this is a financial report that I just recently created where I got a list of my bank accounts and my revenue accounts, my costs. This is all sample data, by the way. So anyone wants to, you know, hack my accounts and uh, other line items, and I can use this directly to actually access that information. So right now we have an MCP that got a list of my accounts, and it's going to say, "Okay, I can create a quote for you." So I can say, "Create a quote for UX services for ABC Furniture." And it's going to go ahead and do that and then give me a link to actually access that quote.
Someone—so I know someone's going to ask, like, "Well, why should I even do this in Cursor, like what, you know, why is this better than, you know, going on to Zero directly?" Let's say I don't want to click things—you know, too many things to click—and I think we're naturally moving towards voice and text to do actions, and that's essentially where we are. So what it comes down to is being able to actually do this directly in here and programmatically. I can spend a lot of time repeating the same actions, and they use a lot of no-code tools to do it. That, to me, is the most interesting piece is like the programmatic piece of this.
Exactly, right? Like, how much time are you going to save by being able to program that stuff?
Absolutely. I think even, you know, right now, like we just created multiple quotes right now.
Yeah.
Within one line of sentence that could have taken many steps. So if I click on this right here, you can actually see the quote that is now live directly through here. So I'll just log in.
That's crazy.
Yeah. Because like this—like what you're talking about right now is just about like how do you—how do you turn SaaS—how do you turn SaaS into programmable?
Right. Exactly. You know, and I think that this might look like, "Oh, it's cute," but it's actually not just cute, right? It's—it's really about saving a lot of time, being more accurate, being more productive.
Exactly. And I, you know, I demoed this to like 300 accountants, and they couldn't believe it. They're like, "You just probably completely transformed our workflow when it comes to tax season where we have to, like, uh, do a list report of all the invoices and quotes that we have for our clients or get tax reports done." So, I just literally using one sentence created a complete quote for ABC Furniture on UX design consultation services. And what's amazing is that it automatically knew how I would approach the services and gave it—gave it a price pointing for it, which is pretty cool.
What's interesting is I have right now, uh, you know, a list of the dashboard of my business, and then I—what it can do is actually open it up and get a finance report that says, "Here's all the accounts you have. Here's the latest core value. Here's your tax rates. What you spend money on," and just, you know, you can just deploy this to a server like NiFi and just share with anyone within your team within seconds.
Crazy. Absolutely crazy.
Yeah. So, putting—you know, this is our accounting team. They—they run our financials here, and I really don't have to do much within Cursor. Anytime my co-founder bugs me like, "Hey, what's this line item I'm seeing here?" I'm like, "Just go check the report. It's—it's already up there, all through Cursor."
I love it.
So, going back, we covered our finance team. Sam Altman, the co-founder of OpenAI, just said that it is the era of the idea guy, and he is not wrong. I think that right now is an incredible time to be building a startup. And if you listen to this podcast, chances are you think so, too. Now, I think that you can look at trends to basically figure out, uh, what are the startup ideas you should be building. So, that's exactly why I built ideaser.com. Every single day, you're going to get a free startup idea in your inbox, and it's all backed by high-quality data trends. How we do it, people always ask. We use AI agents to go and search what are people looking for and what are they screaming for in terms of products that you should be building, and then we hand it on a, you know, silver platter for you to go check out. Um, we do have a few paid plans that, you know, take it to the next level, uh, give you more ideas, give you more AI agents, and more—almost like a ChatGPT for ideas with it. But you can start for free—ideabrowser.com. And if you're listening to this, I highly recommend it.
Um, I also have a UX team that monitors my landing page all the time, takes screenshots, and gives me feedback based on our analytics data and says, "You know what, you need to make some changes here, or you need to do this." And how that works is I use Playwright. Um, so I say, "Use Playwright MCP." And Playwright is really interesting because it's essentially an agent browser where an agent will open up the browser for you and perform specific actions. So you can use it to QA, to take screenshots, to navigate, or give you feedback on, like, errors that you're seeing. And there are other competitors to Playwright.
Right.
Yeah. There's, uh, FireCrawl and Browser-based as well.
Yeah. Yeah. So those are great tools as well. FireCrawl I use a lot, and I'll show an example for that as well where we use the scrape. Um, but Playwright I like because the MCP is so easy to install.
Yeah.
And it's free, which is, you know, fantastic.
Yeah.
So say, "Use Playwright MCP to go to homolytics.com and take a screenshot." So what I'll do is I'll just, you know, take this away, and I want to take this time to actually talk about models as well. You know, with Cursor, it really depends on, like, prompt engineering. Um, it plays a big part on, um, the outcome that you get, and models play also a more crucial part. You like two types of models. You have the thinking and reasoning models, like Claude 3, Gemini 2.5, and Cloud 4 Max mode.
Mhm.
Um, that are really good at complex problems. And then you have the agentic models, like Sonnet 4 or, um, you know, Opus—Opus—Opus 4, uh, and—and those are more like action-based models. So I—I've seen a tactic a lot of developers or just AI prompt engineers use is they have a very big task or complex problem, and they'll get a thinking model to actually come up with a plan to execute.
Mhm.
Then switch over to an agent model and then actually execute on it.
Makes sense.
Yeah. Yeah. And it works really, really well. So you can kind of pass this—pass these models through. So going back to our UX team. So, I just had my UX agent right now taking a screenshot of—of our browser here, and, um, you know, displayed a—a screenshot, and what it did in here is, um, it has a—yeah, a screen capture of what our landing page looks like, and I actually prompted it—so I have a custom mode here that is specifically focused on design and growth—and I'll say, "Based on the screenshot, provide some examples on how you would improve this or give me some, you know, recommendations on what split tests to run." So I can say, "You know, uh, reference the screenshot and provide an outline of, you know, UX growth suggestions and A/B tests," and why do you think, you know, doing this within Cursor is better than, let's say, ChatGPT or something like that? It's a closable system, right? So put yourself in the shoes of a developer or designer that is building a landing page directly within Cursor using custom code. They now have access to MCPs that can run all these tasks in parallel for them without actually having to work with other people or having to switch tabs and go somewhere else. So, if I'm looking at my landing page, I want instant access to my website traffic data. I want, um, access to the—the screenshot of the web page and just get an agent to actually get those feedback for me and then use that feedback to make the changes to the code itself. So I can use this output right here to say, "Okay, based on this output, start making changes to the landing page," right?
You know, that's crazy.
Yeah. So what we have here is essentially a complete, comprehensive review of, like, what to change in the hero section—you know, "Add social proof, video testimonial above the fold, user funnel assessment," and some, you know, suggestions on what to run for split—split tests.
This is like good. This is good advice.
Yeah, it's—it's solid advice.
Solid advice. Yeah, it's solid advice. Yeah, like, you know, our current—our current homepage says "One-click analytics."
Yeah.
U and it says "AAA conversions in 60 seconds." I mean, that's pretty good. Like—
Yeah.
Like I'm—I'm intrigued, you know.
Yeah. You're like, "Say more."
Yeah. Yeah. I'm—I'm curious. So, and you know, and—um, it gives you some implementation notes, and you know, I can, for example, right now reference, um, another MCP—our own MCP. So we dog food here—you know, we—we have a Humbolytics MCP. We say, um, "Refer to Humbolytics MCP and get summary insights," and it'll connect to our, uh, analytics MCP and actually call all the key metrics and insights that we have, and I can use this as context to inform our design decisions. So I can say, "Use this now," you know, "revise this UX review based on the insights that you have right here." So like it's saying, "User journey insights," you know, "our traffic and engagement metrics, traffic sources," and what the opportunities are to actually grow it.
It's interesting. I saw like Brazil is your number one traffic source. It's like, you know, "Why—why," you know?
Yeah, that is a good question.
Yeah. Why—why is Brazil—I mean, surprisingly we have a lot of Brazilian customers, a lot of Brazilian, um, startups, high-growth startups actually that use—use our services, and, um, yeah, like it's interesting, but it's interesting, right? 'Cause you—you probably—you might have missed that, right? And this allows you to see it, and then you can ask further questions and kind of, you know, get more context—get more context.
Get more—and like, you know, the issue is—is like I think we're moving towards LLMs as the interface for everything.
Right.
Instead of me having to go to our app, what's the browser? LLMs are the new browser.
Exactly. Exactly. Yeah. So instead of having to open up our app to say like, "Oh, how many page views did I get today or in the last seven days," it's—I'm by speaking to it, I'm already seeing that. And I think over time, we're going to see generative UI where it's like, "Okay, cool. Now create a UI based on that," you know.
Yeah.
And I think we can—we can get a sense of that as well where it's like, "Okay, I want to show this to my team members and say, 'Hey, like here's all the key insights you need to know.'" So, you know, we have our page views here, our geographic distribution and traffic trends that I can just quickly share with our team.
Amazing.
Yeah. So, at this point, we've covered finance. We know—we know how much we're spending and how much money we're making. We know our growth, and we've also done a comprehensive UX analysis. Where—where do you want to go next? We have QA and sales and marketing. I mean, probably sales because, you know, it's important.
All right. All right. Okay. Cool. Let's do that. So, how do we get to creating email templates and actually automating our sales outbound process?
I'm going to walk you through this. I use my favorite—favorite—favorite MCP, which is FireCrawl. And, um, for, you know, for the audience that might not know what FireCrawl is, um, it's a great tool that we use essentially to have agents run in the browser to scrape content. And what's interesting is you can actually create, uh, context files. And these context files are really relevant when it comes to actually, um, being able to inform your LLM about what it is you're talking about. So we have a Humbolytics LLM.ext text file that—that is very comprehensive around everything—product is how much it costs, what it is, who our ICP is, and you can use, um, FireCrawl to actually generate that file for you, uh, to scrape humbolytics.com, for example, and it'll scrape all the content that's on that website to create that LLM text file for you, and you use this as context for any task you want to get done—any—any questions so far?
I mean—not so far.
Not so far. Okay. Clear. Crystal clear so far.
Crystal clear. Perfect. So, we use this context file, and, um, we have an MCP that we run to actually get our, um, keyword data as well. And we use this as an input between what we've scraped and what we have in the keyword data to start creating both sales and marketing content. Okay. Um, so what I can do, for example, is say, you know, "Based on the keyword data, based on the LLM file that, you know, James covered last time in the—in the pod as well, um, create an article based on this XYZ topic." So I can essentially run an MCP—say, "Use Perplexity MCP to research top, you know, split testing tools and refer to Humbolytics context file and the keyword data to generate an article."
Mhm.
And what this does is you have an agent that will go use Perplexity MCP to actually run a deep search on a topic that you want and then use the keyword data that you've already gathered to create a very comprehensive article with keyword—like keyword—good keyword density and so on. So you can even use, like, content templates that I've created myself—like "Best product in year," "best competitor," "how to do you know a thing"—and it'll create an article that's quite accurate.
Mhm.
Typically when you ask LLMs to write an article, you'll see sometimes it hallucinates. Yeah. When you're using MCPs like Perplexity and FireCrawl, you can use as a way to cross-reference because it's scraping relevant information that's up to date.
Mhm.
I noticed both you and James used Perplexity MCP for deep research. Is that like the—what people should be using?
I think so. Yeah. Yeah. I think, you know, across the board, you have FireCrawl, Perplexity, and Exa.
Exa—yeah.
Uh, Exa is great when it comes to accessing data sets.
Yeah, FireCrawl is great for scraping data.
Yeah. But I think Perplexity—their prompt and their approach to search is probably the best. And for Exa data sets—like what would be an example of a data set that you might want?
Um, find me a list of, um, like five companies in the construction software space.
Yeah. You know, it'll go by category, industry, the type of companies, revenue, number of employees, and from a—like how much do these services cost? Do you know, like, off top of your head?
I just—I just put my credit card. I'm like, "I'll let the vibes take care of it."
I wish I knew. I'm like the ROI. Like, are we talking like tens of dollars, hundreds, thousands?
I mean, so FireCrawl specifically, um, I think I pay like $9 or $19, but Exa when it comes—they have a pay-as-you-go service. And then, you know, we're talking like $5, $25, right? Which is—compared to the time you put into it—the opportunity cost is insane.
Yeah. Okay. Yeah. Cool. Yeah. I, you know, I'm—I'm a vibe coder and I'm also a vibe spender. Like, just—just take the card.
Yeah. Yeah. Um, so an example of this, you know, if we had, you know, run the Perplexity search, um, I created an article on the top tools. Um, it has the right keyword density. So what this will look like is something like an article like this. So it has, um, you know, it's—it's prompted in a way to give me suggested alternative titles, meta descriptions, keywords, content statistics, and article summaries. And how this works is by a very customized mode within Humb—within, um—within Cursor, you can actually create custom modes. Have you—have you used those before?
No, I haven't.
Yeah. Yeah. So, custom modes, um, they're essentially specialized agents. You can assign the model, you can assign a shortcut key if you wanted to, and then also access the tools. But what's really interesting is the prompt. So, I have this prompt right here. I'll show you actually, um, where it's—it specializes in generating comprehensive product requirement documentation. So, "You're an AI specializing in creating PRDs. Here's how you should approach it. Um, here's the key features you should keep in mind. Here's how you should map out the directory"—how do I copy this?
Um, you can—or do I need to do it myself?
You can do it yourself. Definitely, you can—you can use an agent to prompt it to create a prompt for you, which is possible. But I also share these resources online, and I have an access to this prompt as well.
Okay. Thank you.
Yeah. Yeah. Yeah. Yeah. But it's also good because it gives you, um, like Mermaid diagrams, which is key.
Yeah.
For a lot of agents to be able to follow the data flow. Um, and you know there's other great tools.
Explain what that is for people.
Yeah, for sure. Um, so
Memory diagrams are essentially visual diagrams or flows, for like, in most cases used to visualize architecture, right? Or data flows, and I think it's now, in my opinion, a big factor for providing context to agents to actually understand the data flow. So I'll say, like, you know, I can show you an example where I create an example mermaid diagram for a job board showcasing the flow between an employer posting a job and an admin reviewing it. So what this does essentially is it helps you map out what that flow will look like.
And I think the key part around vibe coding in general is you're the architect; you're guiding it. Clawed code is great in terms of you prompting it, and it goes and does it for you. But with Cursor, I think the differentiation is you're still the driver in the seat and you have a co-pilot. It's going to do it for you.
Yeah.
Yeah. Or you're the co-pilot actually, and the driver is doing it for you. So this is what a mermaid diagram looks like. So it can actually map out exactly what the flow will look like if you were to build an app. So I wanted to have a job board, um, where an employer can go in, access a dashboard, create a job posting, fill the information, and then, as a user or as an admin, I want to access this.
So this is quite complex.
Yeah.
Right. So imagine like if I was a product manager at a startup or a company, I'd probably have to do a lot of this before, LMS, with hand. Yeah, like manually doing this. So now you can just get a mermaid diagram directly in here, and it has all, all the mapped out for you.
Yeah, I used to do those by hand. Like I was a product manager, so
Yeah.
Yeah.
Yeah. It's not fun. It sucked. It's like I, I do not like doing this at all. Like it's, it's not fun at all. Um, and I think that's where, where it gets interesting to like talk about Cursor. Even if you're not building a product, but if you're, if you're a PM or you're a researcher, Cursor can be the interface for you to actually just do a lot of this work and use, use as a process to hand over to whatever it is that you're doing. So, as a PM, you can create PRDs. You can do these flows and have it focused on your work.
Yeah. You don't have to be building a business 100%. But ideally, you should be obviously building a business. But yeah, that's a separate conversation.
Yeah.
Yeah.
Yeah. But um, but the net of it here is like you can, you, you know, I created these folders to show that like you can essentially have all these different functions within both your business and the work that you're doing to show what's possible, right? Even if you're a QA agent, for example, you can create, um, you know, QAs. It's, it's the least like interesting on the surface, but it's so important, right? Very important, so so important. And I think what's interesting is like if you have a test plan.
Yeah. You can use something like Browser Base or Playwright to say, "Look at this test plan."
Yeah. And start acting on, like, start doing it. So I can actually get, you know, Playwright to follow this test plan to say, "Go do this, go do that."
Crazy.
Yeah, that's crazy. We have, you know, for our app right now, we have a lot of QA done with authentic browsers, and it works, and it works, and it completely works. So like I can, for example, go in here and say, you know, like, "Go to humbolytics.com using Playwright, uh, MCP, and click on login," for example, and you know, hopefully it works. In this, you know, you know, they say there's a saying, "Do a demo, lose a sale." I'm not trying to sell you on anything, but in most cases, um, we'll see, cuz sometimes you have models that break and they don't use MCPs properly, but um, even if it doesn't work, I'm happy because it shows that things break sometimes.
Absolutely. And that's totally normal. It's completely normal. And I think part of it is chances are the model is not the right model you're using, or your prompt's not right, or just something broke with the Cursor, and you just need to try it again. So, um, as this is loading, oh, there we go. So, it should open my browser right now and navigate to the page. Oh, there we go. Okay. Okay. So, it says, "Oh, the page is loaded successfully. There's a login."
Crazy. And it should click it. There you go. Oh my god. So I can say, you know, "Fill it out, uh, fill it in with an example email." Now, if I wanted to, I can refer to that QA plan that we had and say, "Hey, go through the step by step." Right? So, I just put a, it just automatically typed in a user example. This is magic, man. And it, you know, you know what's crazy? It passed the CAPTCHA. CAPTCHAs don't work anymore, right, in this world?
I don't think so. I'm not too sure. I'm not, you know, I think we just, we just showed that AGI is possible.
Yeah, it's true. So, so you can actually have complete automated QA testing in here.
Incredible. As you're building your app, have an agent spin up. I mean, that's why I wanted to have you on, on the pod is just to show like I, I had heard that you showed James the Boring Marketer how you use Cursor.
Yeah. And after, after you did that, he has never been the same. And I'm having that same visceral experience right now where I'm looking at how you're running your business on top of Cursor, and it makes me feel like, like I use, actually, Zero, for example, for, for accounting. Our, our company uses it, and I hate logging in, and I'm constantly asking our finance person XYZ. Um, and all these little tasks that could be programmable. Um, and I think the reason why I don't use Cursor for a lot of these things is a bit, as a non-technical person, it's a bit overwhelming.
Absolutely. So, like what do you say to, to someone who looks at this IDE and is like a bit overwhelmed?
Yeah. You know, cuz James was, James was that person 3 weeks ago.
Yeah. And now he's like fluid, right?
Yeah. N is, is out the window. He's a big N guy. I'm like, yeah, we need to, we need to check the market share of N because I think we had an impact here. But no, absolutely. I think it's, it's totally natural to feel that way.
Yeah. Um, I think I would actually have a very contrary intake to that, to say, "Open up Cursor and try to build your own app," cuz once you get over that gap of knowing how to deal with the interface, how to use the terminal on those command lines, you can then go ahead and do all these other tasks in a breeze, right? So if you get the hard part done, you can actually do all this. But it's totally natural to feel that way. And if some people don't want to build their code in there, it's okay because the intended use case of Cursor was, is not for these kind of tasks, right? I think because I have more of a development, developer mindset first is like, okay, how do I just keep it all in here? I'm also very like OCD and like to have things very clean and simple. Like a lot of my friends make fun of me for how simple my sidebar is and like, or your dock too.
I notice that.
Yeah.
Yeah. I got nothing in here.
Yeah. I'm like, I just need to keep things very simple here. So, if I can keep things in one place, I'm honestly a little bit of that rooted from that. I was like, "Okay, cool. I just want to see how far I can take this in Cursor."
Yeah. Well, it's cool cuz I, I, I guess I've never really seen anyone use Cursor like this. So, is there anything else you want to, you want to show, or, um, I think we've covered it, right? We, we essentially, uh, covered marketing, finance, design, analytics. Where I really end this off is like for anyone that wants to start off is just open up Cursor and work with just dropping a file in there and interacting with that file and then install an MCP. See how far you can go with MCP. Then work backwards to say, "What can I achieve? What is a task or outcome? I want to," you know, in this case, just get a list of all the invoices that I have to open to pay for. I'll include where to find you and, and get all your great content online in the show notes and description. I'll also include a link to Humbolytics in the, in the description for people who want to check that out. Thank you for being so generous with your sauce. We appreciate it. This is an audience that appreciates, yeah, your, like, this type of sauce. So, thank you so much for sharing.
Oh, I'm glad. Thanks so much. Thanks for having me.
All right.