Pod Insights How I AI A 3-step AI coding workflow for solo founders | Ryan Carson (5x founder) About this episode Ryan Carson is a five-time founder who has spent the past 20 years building, scaling, and selling startups. In this episode, he shares his playbook for using AI to build products, turning “vibe coding” into a structured and scalable approach that can replace full engineering teams.
What you’ll learn:
1. A simple three-file system that transforms chaotic AI coding into a structured, reliable process
2. How to create AI-generated PRDs and task lists that actually work
3. A step-by-step workflow using Cursor to build features systematically
4. Why slowing down to provide proper context is the secret to speeding up your AI development
5. How to use model context protocols (MCPs) to extend your AI’s capabilities beyond just coding
6. Why founders can now build entire companies with minimal engineering teams and how Ryan is doing it himself
—
Brought to you by:
ChatPRD https://www.chatprd.ai/howiai—An AI copilot for PMs and their teams
Notion https://notion.com/howiai—The best AI tools for work
—
Where to find Ryan Carson:
Website: https://www.ryancarson.com/about
LinkedIn: https://www.linkedin.com/in/ryancarson/
X: https://x.com/ryancarson
—
Where to find Claire Vo:
ChatPRD: https://www.chatprd.ai/
Website: https://clairevo.com/
LinkedIn: https://www.linkedin.com/in/clairevo/
X: https://x.com/clairevo
—
In this episode, we cover:
(00:00) Introduction and Ryan’s recent AI projects
(03:25) Demo: Creating a PRD with Cursor
(05:00) Ryan’s open source links: https://github.com/snarktank/ai-dev-tasks
(09:23) Repo Prompt: https://repoprompt.com/
(09:53) Quick recap and common mistakes to avoid
(11:00) Demo: Generating a task list from the PRD
(15:31) The importance of context when working with LLMs
(18:07) Demo: Working through tasks systematically using Cursor
(18:56) Change management
(20:00) How task lists save time for product managers
(21:50) Demo: Using MCPs for front-end testing
(24:50) Specific MCPs and what to use them for
(26:45) Demo: Using Repo Prompt to gain precise control over context
(31:23) Music’s role in Ryan’s development stack
(32:10) Lightning round and final thoughts
—
Tools referenced:
• ChatGPT: https://chat.openai.com/
• Claude: https://claude.ai/
• Gemini 2.5 Pro: https://deepmind.google/models/gemini/pro/. https://deepmind.google/models/gemini/pro/
• Repo Prompt: https://repoprompt.com/
• Taskmaster: https://github.com/nooqta/taskmaster
• Browserbase: https://browserbase.com/
• Stagehand: https://docs.stagehand.dev/integrations/mcp-server
—
Other references:
• PostgreSQL: https://www.postgresql.org/
• Prisma: https://www.prisma.io/
• SQL: https://www.sql.org/
• MCP: https://www.anthropic.com/news/model-context-protocol
• VS Code: https://code.visualstudio.com/
—
Production and marketing by https://penname.co/. For inquiries about sponsoring the podcast, email jordan@penname.co.
Transcript Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.
0:00 Speaker 00 I think the biggest mistake that I do that everyone does is they try to rush through the context where you just don't have the patience to tell the AI what it actually needs to know to solve your problem. And I think if we all just slow down a tiny bit and do these two steps, it speeds everything up. Nobody really knows how to do this stuff. The only way you're really going to figure it out is by getting in here and getting your hands dirty and see what works.
0:22 Claire Vo? That's a place where so many engineers and product managers get stuck in a loop like, who's going to take this PRD and actually break it down in the right steps? So even just this is such a time saver for people building products.
0:34 Speaker 00 Building this new startup, I literally feel like I'm able to do all of it. Am I able to do it as well as a dedicated product manager? No. Am I able to think as deeply as CTO? No, but I am able for sure to build this company.
0:49 Claire Vo? This is the way, people. I'm telling you, pay attention.
0:55 Claire Vo? Welcome to How I AI. I'm Claire, product leader and AI obsessive here on a mission to help you build better with these new tools. 2025 is definitely the year of the vibe coder, but you can't always vibe your way to a scalable execution strategy. In this episode, Ryan Carson, a five-time founder with 20 years' experience, shows us how he brings PRDs, task lists, and some advanced prompting techniques to Cursor to make sure he's not just vibing, he's building the right things. Let's get to it. Today's episode is brought to you by ChatPRD. I know that many of you are tuning into How I AI to learn practical ways you can apply AI and make it easier to build. That's exactly why I built ChatPRD.
1:39 Claire Vo? ChatPRD is an AI co-pilot that helps you write great product docs, automate tedious coordination work, and get strategic coaching from an expert AI CPO. And it's loved by everyone, from the fastest-growing AI startups to large enterprises with hundreds of PMs. Whether you're trying to vibe code a prototype, teach a first-time PM the ropes, or scale efficiently in a large organization, ChatPRD helps you do better work fast. And we're integrated with the tools you love: V0 Dev, Google Drive, Slack, Linear, Confluence, and more, so you don't have to change your workflow to accelerate with AI. Try ChatPRD free at chatprd.ai/HowIAI. And let's make product fun again.
2:24 Claire Vo? Hey Ryan, it's nice to have you here.
2:26 Speaker 00 Thanks, it's exciting to be here. I've listened to every episode so far and I'm honored to be here myself. I can't wait.
2:32 Claire Vo? So I'm gonna start with an easy question, which is, what are the last three things you built with AI?
2:37 Speaker 00 I don't know if you call constantly using ChatGPT with your kids, building something in AI.
2:43 Speaker 00 I feel like I'm the constant AI coach in our family and I'm always delighted actually with what our kids are doing. And because of that, my amazing 14-year-old kiddo, Devin, he said, Dad, like, you know, I've been thinking about this game. And I said, well, like, let's build it. And so we're building a primitive little side-scroller and he's like the creative director, so.
3:04 Claire Vo? Oh, that's so fun.
3:05 Speaker 00 So that's sort of thing one. And thing two, I would say thing three is the startup I'm building, which is just this huge amount of work where I'm coding. But then there's just all sorts of little quick things I vibe code, you know, all the time. So all day, every day.
3:18 Claire Vo? All day, every day, ChatGPT and vibe coding. I'm gonna make us all t-shirts and hats.
3:26 Claire Vo? Okay, so one of the things that I think you do really well compared to other vibe coders is you bring some structure to the process. You're quite wise about how you use Cursor. And so I'd love for you to pull up your screen and show us how you get Cursor to follow a plan.
3:43 Speaker 00 Let's do it. So the reason why I do this is because I've been coding and coding and coding with AI. And you just learn as you do this that you really have to get good about context, what you're showing the AI, what you're asking the AI to do. And you end up cutting back all the things you're doing to a manageable amount of things that the AI can actually do. So the process I'm gonna show you, I'm in Cursor now, in case you're listening to the show, you can download it for free at cursor.com. It's basically a VS Code fork, right? So if you ever use VS Code, it's a great tool. So what I've done here is opened up a basic project that I vibe coded yesterday just to kind of show you what's going on.
4:27 Speaker 00 This is a stupid little CRM tool for a yacht club, because I thought that would be kind of funny. So there's that. So, all right, so say that you want to make a change in here and it's a bigger change than just a small, you know, a quick, hey, will you change this thing? Say it's larger. All right. So you would probably want to create a product requirement doc, right? Like PRD. And, and if you're watching the show and you don't know about Claire's chat PRD, absolutely check it out because it's amazing. But if you need a lighter lift, then here's what I do. So I've created three files over here in my cursor rules folder. And I'll walk you through what they do and how they work. I've also open sourced these. And so we can throw those, a link in the show notes for you all to grab that.
5:12 Speaker 00 So the first one is a simple instruction for the AI to create a product requirement docs. Again, Claire knows everything about this and lives this and breathes this. But to a lot of people that don't know about PRDs, right? It's, it's how you describe a feature that you're wanting to build. And so this rule explains the AI how to write a PRD for the user. So the way you use this is, is very simple, right? So I'm going to go over.
5:40 Claire Vo? And I want to pause really quickly because I love your initial prompt, which says, this is a PRD that's suitable for a junior developer to understand and implement this feature. That's such an interesting little call out.
5:53 Speaker 00 Yeah, well spotted because again, as you code and code and code with AIs, you, you start to realize that they're like a genius PhD student, right? But, but they, they can't seem to connect these really simple, obvious things that you and I know. And so saying junior developer is kind of a way to instruct the AI. Let's keep this at the, at a certain level.
6:18 Claire Vo? Yep.
6:19 Speaker 00 You know, we've got some process, we've got some clarifying questions. You'll see this in action. I'll actually run it.
6:24 Claire Vo? Great.
6:24 Speaker 00 So let's do that.
6:25 Claire Vo? Great.
6:25 Speaker 00 So I'm going to at include this file, which puts it in the context window. And then I'm going to give it a simple instruction, which I've pre-written, which is, I'd like to add a report that shows me all the boat names of members and how many emails they've been sent. This is a big deal if you're in a yacht, in a yacht club.
6:42 Claire Vo? You have a lot more fun use cases than I do for vibe coding.
6:48 Speaker 00 I try, you know, so let's see what happens. So I'm gonna go and hit go. We're in agent mode. For those of you who are astute, you'll notice I'm using Claude 3.7 Sonnet in max mode. I actually tend to use Gemini 2.5 Pro and it's funny, I didn't notice that I had selected that, but fine.
7:08 Claire Vo? I'm now a default 03 girl.
7:10 Speaker 00 Ooh, ooh.
7:11 Claire Vo? Yeah. And then when 03 gets stalled out, I go to 3.7.
7:14 Speaker 00 Right, right. It's funny, I'm pretty much default Gemini 2.5 Pro. I love the max mode. It is sort of expensive. I probably spend maybe 300, 400 bucks a month. But, you know.
7:26 Claire Vo? Worth it.
7:27 Speaker 00 It's okay. Okay, so what we've got here is the AI came back and gave us some clarifying questions on the PRD. I'm gonna answer a couple of these just to, to show you how this works. So if you're listening along, the AI has said, okay, great, I'll help you create a PRD. Have a couple of questions for you. The first one is, what is the problem this reporter is trying to solve? Well, we are, what problem is this? We're trying to give visibility into how many emails people are getting. All right, so that's thing one. Let's go, let's answer the second one. Who specifically will be using the report? All right, we'll say admins. Where should this report be accessible? You pick. And so there's a couple, I would say there's probably too many questions in here.
8:09 Speaker 00 But one of these things again, that I want everyone watching to kind of get used to is, you'll notice in this PRD rule file, I've done something specific here where I've said, I want these questions to be dot notation where it's 2.1 and 2.2. Because you end up, otherwise the AI would just give you a bunch of questions and it would, it would put more than one question in a bullet point. And it becomes hard to use. So you kind of get used to this specificity, right? So I'm just gonna say the rest, make your best judgment.
8:41 Claire Vo? I do the exact same prompt. I say, you pick whatever you think is best.
8:46 Speaker 00 Yeah, cuz I'm kind of lazy, right? So, alright, so let's do that. So it's gonna fire up and start creating this PRD. Now, what I've chosen is, is, you know, over here, I've got a task folder. PRD in there, and then we're gonna generate a list of tasks in a minute. So it says, okay, great. I'll now create it, and it's generating it. So that will happen in a minute. I'm gonna actually nip over to my browser, cuz I do wanna show you a couple other interesting things. So first of all, this is my amazing yacht club member app that we just five coded. Now, we're gonna pop this in the link, but this is the repo that has these prompts, and it will give those to you.
9:28 Speaker 00 But there's this really cool open source tool called Taskmaster. This is open source, it's completely free, and it's like a hyped up version of what I'm showing you. So it's, it runs in as a command line interface tool. It's great. It's actually was too much for me. I wanted less, less power, less control, or less power, more control. So this is a great alternate to that. So I'm gonna head back to Cursor.
9:54 Claire Vo? Just to kind of recap where we are so far. So you have a cursor rules files that gives rules and instructions for generating a PRD. You generate the PRD in agent mode using whatever your preferred model is. That puts that PRD markdown file in a tasks folder that you've put in, in your repo. And now we're gonna look at that PRD and show how you work through it to build something.
10:20 Speaker 00 Exactly.
10:21 Claire Vo? Okay.
10:22 Speaker 00 You're a good student. So here is our PRD. It's pretty straightforward if you've ever looked at a PRD. I'm gonna show you. So you've got functional requirements, on-goals, design considerations, etc., right? And let's back up. Like, what are we doing here? What we're doing is making it clear to the large language model what we want done. I think the biggest mistake that I do, that everyone does, is they try to rush through the context where you just don't have the patience to tell the AI what it actually needs to know to solve your problem. And I think if we all just slow down a tiny bit and do these two steps, it speeds everything up, right?
11:01 Speaker 00 So we've got a PRD. It's not rocket science here. Now, next, what we want to do is generate tasks for this. So let's go and generate the task based off this PRD. So if you're listening on back and cursor, and I'm including the file that generates the task. It's called generate tasks. And I'm gonna say, please generate tasks. Tasks for them. I'm gonna tag the PRD.
11:27 Claire Vo? Got it. So this is another rule, probably similar to your generate PRD. One that explains what a task is and how to do it well. And then you're giving it the context of the PRD itself.
11:38 Speaker 00 Exactly. So let's have a quick look at the generate task list. So the goal is to guide an AI assistant in creating a detailed, step-by-step task list. This is what I want the list to look like. This is the process. You'll see it's gonna ask me a couple of questions. And it's gonna pump out tasks in this format as a markdown file.
11:58 Claire Vo? I have to ask, how did you create these rules? It's a lovely prompt. It's well structured. It's clear. How did you get good at writing these instructions prompts?
12:09 Speaker 00 The same way you did. Basically, I tried a couple of things. They didn't work. I got more specific. And then, of course, you know, I had a very intelligent LM write this for me. And then I edit it. You know, again, you sort of learn some of the tricks of the trade. Like, I want this task list to be in markdown. And I want there to be checkboxes. So we can check them off. I mean, silly stuff like that. So the biggest thing I want people to walk away from the show is, you know what? Nobody really knows how to do this stuff. The only way you're really gonna figure it out is by getting in here and getting your hands dirty and see what works. And then stick with a model that you consistently like. Like, I'm getting to know Gemini 2.5 Pro really well.
12:53 Speaker 00 Like, what it's good at, what it's not good at. So let's see what it's good at. All right. So I just said, please generate tasks for this PRD. I tagged the file. We're using a reasoning model, so we're seeing the thinking tokens kind of whiz by.
13:09 Claire Vo? And all this thinking might cost you a little bit more, but you get a little bit more visibility, you learn more, and it looks like it's doing a good job.
13:16 Speaker 00 Amen. Yeah, I feel like for the extra five cents, like, this is absolutely worth it. So in this instruction, the instruction is, give me some basic tasks and then ask me if they're okay, and then tell me to go to proceed. So you'll see, it says, ready to generate the subtasks, respond with go to proceed.
13:38 Claire Vo? You know, as I'm looking at this side by side, what I really like about your generate tasks instructions is you've given it a very explicit process. It's you get this file, you do step one, you get this next file. And it's not quite an agent, but it really brings in this like agentic thinking of where are the decision points, where are the user interaction points, but in sort of more this linear, step-by-step chat mode.
14:05 Speaker 00 Exactly. And I have been hounding the Cursor team nonstop. Like, why don't you just build this into the core developer experience of the app? I don't understand why this isn't just the way you use Cursor. And they keep saying, we're doing it. So, you know.
14:18 Claire Vo? And for those listening, that delightful "brrr" is when Cursor is done generating.
14:24 Speaker 00 Isn't that the best? I love that.
14:25 Claire Vo? I love it. This episode is brought to you by Notion. Notion is now your do-everything AI tool for work. With new AI meeting notes, enterprise search, and research mode, everyone on your team gets a note taker, researcher, doc drafter, brainstormer. Your new AI team is here, right where your team already works. I've been a longtime Notion user and have been using the new Notion AI features for the last few weeks. I can't imagine working without them. AI meeting notes are a game changer. The summaries are accurate and extracting action items is super useful. For stand-ups, team meetings, one-on-ones, customer interviews,
15:06 Claire Vo? and yes, podcast prep, Notion's AI meeting notes are now an essential part of my team's workflow. The fastest-growing companies like OpenAI, Ramp, Fersell, and Cursor all use Notion to get more done. Try all of Notion's new AI features for free by signing up with your work email at notion.com/HowIAI.
15:32 Speaker 00 So we're going to accept those changes. We're going to hop over here and let's just kind of see what it did, right? So.
15:37 Claire Vo? Wow.
15:38 Speaker 00 I know. It's fun, isn't it? So what you've got up here is the relevant files. Now, this is a trick I learned from a friend on X where my thinking here is that this should help the LLM just remember what files are we really focusing on? Even though I will specifically tag these in the context, I think it's helpful. One of the things I want to briefly touch in, I probably anthropomorphize LLMs probably too much. But, but because they're trained on human output on the web, you know, my belief is we need to give LLMs like the right context and be as helpful as we can so they can actually solve our problems, right?
16:16 Claire Vo? I, I completely agree. I'm also very polite to, to the LLMs. It's how I get people to, people to do work. It's how I'm going to get the agents to do work.
16:25 Speaker 00 Right. So why wouldn't you, you know, be, treat an agent like you would treat a human? That's why.
16:30 Claire Vo? Yeah, I agree.
16:31 Speaker 00 All right. So then we've got, you know, pretty detailed list of tasks, right? We've got one, two, three, four with subtasks. We've even got sub-subtasks here. And then we're going to start iterating on this. So here's how it works. I've got another rule called task list. I'm much probably need to rename because that doesn't really make sense, but this is the instructions for iterating through these tasks. So I'll kind of walk you through that. So this is task list management. These are guidelines for managing task lists in markdown files to track progress. The task completion, we have a, we want to do one subtask at a time. Like this is really important that the AI doesn't start trying to do all the tasks. When you finish a subtask, immediately mark it as complete.
17:16 Speaker 00 And then you'll see I say stop after each subtask and wait for the user's go ahead. And there's just a little bit more clear clarity here.
17:26 Claire Vo? If people see me shaking my head, it's because I am realizing now I'm using cursor like such an amateur that's
17:34 Speaker 00 I doubt that.
17:34 Claire Vo? This is so good. I'm just floating through the ether saying,
17:39 Claire Vo? oh, three, take me away. And maybe I'm, I'm overconfident in my product management skills. And so I'm putting this all into the chat. But this is, this is the way people, I'm telling you, pay attention.
17:51 Speaker 00 I appreciate that. I mean, I think the answer is yes to both those. I mean, sometimes you do just need to roll and see what happens. But I'm learning over and over that if I don't follow this process, I end up down some rabbit hole and I have to revert. Right? So let's, let's go ahead and continue here. What I'm doing now is I'm tagging the task list rule, which tells the AI how to act. And then I'm going to say, let's start. And then I'm going to tag the, the tasks.
18:18 Claire Vo? The task list.
18:20 Speaker 00 Which is.
18:21 Claire Vo? So we got a PRD, we got a task list, and now we have a set of rules that knows how to work through a task list and actually get work done.
18:29 Speaker 00 It does. Exactly. So, all right, so let's start. It's going to think about it and it's looking through this big list of tasks right now. And so what it's saying is, okay, let's start the first subtask. Define Prisma scheme email campaign. I'll start by reading the existing Prisma campaign. Okay, existing blah, blah, blah, you know, so it's just thinking it through. And boom, it has checked off 1.1.
18:56 Claire Vo? Ah, with a delightful noise. I have a question. Are you, are you hooking it all into Git? How are you doing sort of the, the change management here?
19:06 Speaker 00 Yep. So what I tend to do is I will commit after I, after I finish either one of the bullet, one of the, one of the parent tasks. If I, if I feel like the app is in a state that it's workable.
19:19 Claire Vo? Yep.
19:19 Speaker 00 Now I will commit at that point.
19:21 Claire Vo? Got it.
19:21 Speaker 00 If I don't, I won't commit until all of these tasks are done.
19:26 Claire Vo? Oh, wow.
19:26 Speaker 00 Um, which, you know, is probably, you know, half a day of work. It, you, you know, you kind of get used to like, oh, all right. If I had to revert now.
19:37 Claire Vo? Yeah.
19:38 Speaker 00 How bad would it be?
19:39 Claire Vo? Yep.
19:39 Speaker 00 And, and I try to think about, well, what would I need to undo? So it said, uh, all right, shall I proceed with subtask 1.2? Say, uh, yes. Sometimes I'm really lazy and I just say, why?
19:55 Speaker 00 As in the letter Y. So it's just going to keep working. Now, we don't need to go through this whole thing.
20:00 Claire Vo? Yeah.
20:01 Speaker 00 But I just wanted to show all of your amazing listeners and audience. This is a pretty easy process to follow. And I've built huge features with this. I mean, you know, 10,000 lines of code reliably and almost never had trouble. I still feel like this human in a loop part is really important where after each task, you are kind of checking what's happening. I've noticed that it often does introduce some small problem or there's a linter, you know, error, and then you gotta go fix it. So.
20:35 Claire Vo? And, you know, this is great to get the actual engineering work done, but if I just take a step back for product managers out there that don't know where to get started with Cursor, even if you just did the PRD task list part, I'm looking at this task list right now and it's got, and basically like epics and tasks in it. And that's a place where so many engineers and product managers get stuck in a loop. Like who's going to take this PRD and actually break it down in the right steps that are going to work in our code base. So even just this is such a time saver for people building products.
21:10 Speaker 00 I think so. And now you can over engineer this process, you know, this is literally a markdown file.
21:17 Claire Vo? Yeah.
21:17 Speaker 00 It's somewhat hand cranked. You know, I thought, oh, maybe I'll use, you know, Asana's MCP, you know, server and create Asana tasks.
21:25 Claire Vo? Yeah.
21:26 Speaker 00 And I was like, no, like, it's actually easier for me just to see a markdown file and know what's happening. And I can even add tasks to it. So my encouragement to everybody is just start small, start simple and get good at that and get comfortable at that. And then you can, you know, graduate from there. So.
21:43 Claire Vo? And 30 chime noises later, we are going to have a report about your yacht club.
21:49 Speaker 00 That's right.
21:49 Claire Vo? Emails. I mean, speaking of MCPs, are you using MCPs at all in your Cursor experience? How is that fitting into your workflow?
21:56 Speaker 00 I am. So I'm gonna show you a couple of those now. I wasn't using many at the beginning, and then the first MCP I started using was for Postgres. Because it's really useful to ask the AI, hey, can you go see if I, if this data is actually in the database? So I started that, and then I've gone down the rabbit hole. So I'm gonna show you a couple MCP servers that I think are really useful. So let me go and share my screen. Okay, so what you are seeing now is on the left, I've got browser base. It's the actual backend. I've got a free account, so I just want to try this out and see how it works. I'm gonna show you the fun that you can do with an MCP here. So on the right, I've got Cursor, and I'm in my Cursor settings, and I go to MCP.
22:41 Speaker 00 And you'll see down here, I've got a browser-based MCP.
22:45 Claire Vo? Yep.
22:46 Speaker 00 Which I've set up. I've got Stagehand as well, which is really fun. So let me show you how this works. So we're gonna go back into, sort of, new chats. Turn on agent mode, and I'm gonna say, navigate to ChatBRD. Yeah, and take a screen grab. Now, let's see if the MCP gods cooperate.
23:14 Speaker 00 Because I did this this morning, and I was like, wow. This is super cool. Okay, so what we should see over here in a minute. Let me refresh this to see. It did. Oh, my God, this is so cool.
23:29 Speaker 00 Okay, so in the, in the cloud.
23:32 Claire Vo? Ooh.
23:34 Speaker 00 So what I'm doing is controlling a headless browser in the cloud from Cursor.
23:41 Claire Vo? The future is now.
23:43 Speaker 00 It's in, with bonkers. Well, it's like, okay, what use is this? Like, that's kind of cool. I mean, that's actually do something kind of fun. So I'll say, navigate to Pricing. Okay, so we should actually see the cursor move over here in this headless browser in the cloud.
24:04 Claire Vo? You're just showing off at this point.
24:06 Speaker 00 I, well, I just learned this this morning, and, and I was thinking, oh, my God, this is so neat. Let's see if it, if it actually works.
24:14 Speaker 00 Yep, there it goes.
24:15 Claire Vo? Whoa.
24:17 Speaker 00 So again, well, besides kind of a parlor trick, like what actual use is this? I think it's going to unlock a huge amount of front end testing for me.
24:27 Claire Vo? Yeah.
24:27 Speaker 00 Right? So as you know, like, it's really kluge right now, trying to to squash bugs on the front end. When you're like taking a screen grab and you're pasting into cursor and you're like, no, I mean over there.
24:40 Claire Vo? Yep.
24:41 Speaker 00 So now we have this this ability to start automating some of this behavior. You know, right inside cursor. So that was kind of fun trick. I thought I would show you. Now, if you're watching this, you're like, what? It really what we're doing here is I'll go back to the M C P part of cursor. What we're doing is is basically giving cursor the ability to interact with other apps. Right? And so we're saying, okay, browser base is pretty cool. You can have this browser in the cloud. You can do things with. I want to be able to tell the AI what to do. I don't want to have to know how to call the API's and do all that stuff. So that is what I thought was a lot of fun.
25:22 Claire Vo? And can you walk us through just a couple of these cases, the ones we're seeing here? So browser base lets you basically browse the web and and do a couple of things through through the text window and cursor. What are some of these other ones that you find useful?
25:36 Speaker 00 So Postgres is probably the one I use the most. So this for the startup I'm building, I'm using Postgres on for the database in Vercel. And there's just a lot of times where I just want to be able to tell the AI, you know, is this value in this row in the database? Like, I don't want to have to actually write SQL to do that.
25:55 Claire Vo? Yeah.
25:55 Speaker 00 And so you can go right into the chat window and just say, hey, you know, use the Postgres tool and tell me if this is in the database or not, which is really cool. Prisma is I'm using for this play project, which is fun. Same with SQL Lite. So the one I use every day is Postgres.
26:14 Claire Vo? And one of the things that I love about AI, especially in the sort of dev tools stack, is it just reduces toil. And one of the pieces I feel like is toil for engineers is how many tabs you have to have open to orchestrate working across your task list, right? You have to have your project management software open so you know, like, what task I'm working on and what's next. You have to have your browser open so you can do some work, your query in your database. And all this puts it in a single interface that you can seamlessly switch through in natural language.
26:46 Speaker 00 Amen. So I do want to show you one tiny, quick other tool, and I know we're almost out of time.
26:50 Claire Vo? No, let's do it.
26:51 Speaker 00 So you're probably hearing over and over, everyone's listening or watching me preaching about context.
26:58 Speaker 00 It is just so much more important than I think we understand. And again, if we anthropomorphize, it's like, well, how would you expect anybody to do anything unless you give them the right book or the right piece of paper? Right? And so I've started to use a tool called Repo Prompt. Again, I don't have any financial reason to say this. I don't own any of this company. But it's this really great tool for Mac. And I just want to show you how to use it really quick. The question is, why? Like, why would I just use Cursor? Well, the thing about Cursor is there's all this magic happening in the background with the context, where you don't know for sure what is in the context unless you tag it, right? Which is fine, but they sort of magic that away. And sometimes you really, really want to control the context.
27:41 Speaker 00 So what I've done is I've opened Repo Prompt. I've got the site on the left. And on the right, I've got the UI. And I've opened up the HowAI project, which is that simple yacht email app. Okay, so say that I want to throw a lot of these into a prompt to do something. So if I select the whole repo, let's go over to Compose. Down here, you can see how many tokens that is. All right, so that's 395,000 tokens. Okay, that's way, that's way too many. So let's go ahead and reduce that. So, all right, I know I want some stuff in the app folder. All right, why is it so big?
28:23 Speaker 00 Now we're at, you know, 324,000, still too big. So I think it's probably generated, so you get rid of that. Okay, good, now we're at 12,000 tokens. So the point is, if you know what context you want, Repo Prompt is a really powerful way to do that. So I'm going to go ahead and select Components, Lib. Let's open up the schema for Prisma, you know, and scripts. All right, now what? So I go up here and I put in a prompt. How can I improve the maintainability of this code? And then you can do some other neat things like, I want to include a stored prompt, and I'm going to use the Architect version. Now, what is that? Well, that is a prompt that Repo Prompt has written already
29:08 Speaker 00 with a bunch of kind of power moves in it, right? It should act like an Architect versus a Dev. So now I go down to Copy, and I'm going to say, I want to include the saved prompts, I want to include the files, and I want to include the user instructions. And I'm going to click Copy. Now what? Well, I'll show you. So let's go over to O3. Now, I could go over to Cursor, but I don't know exactly what Cursor is going to do to my context. Instead, I'm going to paste it in here. Now, what is this? What you've got, let's start at the bottom. It's basically putting everything in XML tags, right? So you've got user instructions, so that was the prompt I put in.
29:49 Speaker 00 Then the meta prompt is telling it how to execute the prompt and the files that it's including. Let's go up here, so File Contents. This is kind of the key. So each file is included, and it's demarked specifically like this. So it says the file with the actual path. And so it, basically, it's very specific. It's saying to the AI, this is exactly the context you need. It's super clear. And then you can execute some sort of prompt on that. I do this for like, heavy lifting stuff, where I used to do this all the time with O1 Pro. Where I would go into repo prompt, I would select exactly the right context and then I would go into O1 Pro and say,
30:34 Speaker 00 "Think super hard about this. I've given you exactly the right context and you get amazing answers because of that."
30:40 Claire Vo? Yeah, and these, these new models have such big context windows, but I'm not copying and pasting 12,000 tokens into the chat window.
30:49 Speaker 00 Right.
30:50 Claire Vo? But this tool just does it for you. Really interesting. And then you're getting less of a black box here.
30:56 Speaker 00 Less of a black box. Yeah. And, you know, and I have a feeling that this stuff will probably go away. I think, you know, the context windows are going to get bigger. Tools like Cursor are going to get better at managing context, you know, but right now you have to do all this stuff. Like you, you can't just wave your wand and, and hope the LM is going to write all the right stuff for you.
31:17 Claire Vo? That, that's what I've been doing. So now I have a better process to, to follow here. Okay. Anything else really important in your stack you need to tell us about?
31:27 Speaker 00 The most important thing is this. So let me share my screen. So this is TS Knows live set from New York City.
31:37 Speaker 00 This is how I actually code. I just turn on some EDM and, and late at night, you know, code after everyone goes to bed, listening to amazing EDM. And that is an important part of my stack.
31:48 Claire Vo? So if I could ask for an additional cursor feature, it would be AI-generated streaming EDM matched to the generation pace of your tokens that ends with a drop instead of that cursor?
32:02 Speaker 00 I would pay extra money for that.
32:04 Claire Vo? You heard it here first. Get, get, forget all this task management context window stuff.
32:10 Claire Vo? Okay. We'll spend three minutes on lightning round questions. One, you're a builder, you're a founder. It's clear how this technology is changing building part of things. How is it in your mind changing the company and the founder side of things?
32:25 Speaker 00 Wow. It's a complete rewrite. So I've been fortunate to start three companies and, and see them acquired and, and one of them, you know, we had about 110 employees and I had a CTO and a VP of Edge and, you know, product managers and, and, you know, building this new startup, I literally feel like I'm able to do all of it. Now, am I able to do it as well as a dedicated product manager? No. Am I able to think as deeply as a CTO? No, but I am able for sure to build this company. By myself. And I mean, you're doing this with chat P R D. I mean, it's it's bonkers to me that it's actually possible. So I just can't wait for the future.
33:08 Claire Vo? I could not agree more. Okay. And then, of course, you showed us how you're a very thoughtful and organized product manager and manager of your AI. But you know, it sometimes doesn't listen. So what do you do? What's your what's your tactic for really getting it back on track?
33:28 Speaker 00 So I'm just too nice. I just say, please think harder about this. Like, I know you can do this. You know, think again about it. And I'm just not a mean person. So as much as I want to say, you know, god damn it, you know, I don't.
33:47 Claire Vo? I have a hypothesis that this is this is actually a parenting thing because I do the same thing. I say, I believe you can do this. I believe in you.
33:56 Speaker 00 I believe you can.
33:57 Claire Vo? I believe you can. Well, that's amazing. Ryan, where can we find you? And how can we be helpful?
34:02 Speaker 00 I am on X all the time. So X dot com forward slash Ryan Carson. And if you want to know a little bit more, just head to Ryan Carson dot com. And that is me.
34:11 Claire Vo? Great. Well, thank you so much.
34:13 Speaker 00 It's been a blast. Thank you.
34:16 Claire Vo? Thanks so much for watching. If you enjoyed this show, please like and subscribe here on YouTube, or even better, leave us a comment with your thoughts. You can also find this podcast on Apple Podcasts, Spotify, or your favorite podcast app. Please consider leaving us a rating and review, which will help others find the show. You can see all our episodes and learn more about the show at how I A I pod dot com. See you next time.
Previous How custom GPTs can make you a better manager | Hilary Gridley (Head of Core Product at Whoop) Next The exact AI playbook (using MCPs, custom GPTs, Granola) that saved ElevenLabs $100k+ and helps them ship daily | Luke Harries (Head of Growth) Page updated November 3, 2026 · transcript and summary are machine-generated