The Secret AI Prompt Tool Silicon Valley Engineers Are Using

Welcome to the Next Wave podcast. I'm your host, Nathan Lands, and today we have a killer episode. Everyone's talking about vibe coding, but the reality is, for most things, vibe coding doesn't work right now. And even the guy who coined the term, Andrej Karpathy, he recently posted that he's now trying to provide more context to models because he's realized that's what you have to do to get good results back. The reality is, if you want to get the most out of AI coding, you've got to be showing the models exactly the context they need. Today I'm going to show you the secret weapon that all the top AI coders are using that you probably haven't even heard of yet. Today I've got the founder of Repo Prompt, Eric Provinger, on here, and he's going to show you exactly how it works and how you can use Repo Prompt to take your AI coding to the next level. So let's just jump right in. Yeah. I've been telling people about Repo Prompt for, like, the last, you know, probably six months or so. I've kind of felt like it's been almost like my, like, AI coding secret weapon.

Yeah.

Yeah. I really wanted to bring you on after I saw that tweet from Andre Karpath the other day. He used to be at Tesla AI now he's like, one of the best educators about how LLMs work and things like that. He had this. This tweet saying, noticing myself adopting a certain rhythm in AI assisted coding, I code. I actually and professionally care about contrast to Vibe code. You know, he coined the term Vibe code, which everyone's been using. And he. And he. Then he basically goes on to talk about, like, stuffing everything relevant into context. All this. I was like, he doesn't know about Repo Prompt. I'm like, how did this, like, top AI educator in the world, top expert, everything. Totally has no idea about Repro prompt. I was like, okay, so I need to get Eric on the podcast. At least try to help with that.

Yeah, I appreciate that. Yeah. I mean, yeah, looking at that, that tweet, you see exactly like that. That flow that, like, got me started. Like, you. When you. When you start getting serious about coding with AI, like, you start thinking like, oh, well, how do I get the information to the AI model? And, like, the UX on all these other tools is just not cutting it for it. Like, you need a tool to just be able to quickly select, search for your files, like, find things.

There's probably some people Listening right now, like, oh, it's okay.

Cool.

Nathan's excited. What is it? So like, maybe, you know, if you could explain, like, try to simplify it. And I think we should then just jump into a demo and we can kind of just.

Sure thing. Sure thing. Yeah, it'll probably be easier if I just show the screen there. I mean, the first thing you're going to do when you're going to open up Repo prompt is pick a folder. So this here is like the entire folder. But generally when you're working in a repo like this, you want to think through, like, what are the files that are going through and if you're using a coding agent, like with cursor or whatever, what the first thing they're going to do when you ask a question is, okay, well, let me go find what the user's trying to do. Let me search for files and pick those out. And with repo prompt, that's like on you as a user to do that work, there are tools to help with it. So, yeah, like I'll just show here with a website. This is my other project. This is the website for Repo Prompt site. I was just iterating on some pricing and just lowered the annual price there. And yeah, so help me update all the docs pages. So if I do that and then I just do Gemini Flash quickly to show what that looks like. So it will actually search for files using an LLM based on the prompt that you've typed out. So like, you know, a big part of using repo prompt is that you have to know, you know what it is that you're trying to select here. Right. And you know what? I noticed a lot of users, they were just putting everything in. They would just say like, okay, just select all and that would be it. And you'd be like, okay, well, I.

Was doing it at the first.

Yeah, yeah, I mean, I mean, yeah, I mean that's the easy thing to do. You're like, okay, well there's the code base. Perfect. But you know, there's plenty of tools that can just zip up your code base and that's easy. But like the power of Repo prompt is you can be selective. You don't have to select everything. So I can just hit replace here and then. Okay, well what did that do? Okay, well that actually found all these files here that are related to my query. Put them in order of like priority, of like importance based on what the LLM judgment is. And of course, if you use Gemini Flash, you're Not going to get the best results. Compared to like using you know, like a bigger model like Gemini 2.5 Pro, it'll be a lot more thorough. But you know, like that's the thing. So it'll pick those out. It'll use something called codemaps to help with that. You can see that the actual token file selection queries is just 6k tokens. If I were to switch over to the repo prompt code base, which is a Swift code base, you can see that the code base is a lot larger. So the map is 85,000 tokens. So I'll explain a little bit what is a map really? What do you do with that and how does this help and what does it do working with a code base if you've spent some time programming in the past? I know a lot of folks are getting into just using AI and they're not super familiar with all the technicals there, but. Yeah, exactly. Well, there's a lot of files and you want to try and get an understanding of the files. So revamp has this codemap feature and what this will do is it will basically as you add files, it'll index them and extract what's called like it's a map, but it's like a high level extracted view. Kind of like an index.

Like an index of your code base.

Exactly, yeah. And this works for a bunch of languages, Swift being one of them. And if I were to select everything, kind of show you what that looks like. So these are the map files. So I'll open that up, I'll open VS code just to show you. So you've got all the file definitions and just what's private, what's public and you can scroll through and so you give this information to the AI model and it's able to be very selective. Well, it's able to have a lot more information in a lot less context. And so the context builder uses that data to help you find what the relevant files are based on your query. So it has like a kind of peek inside the files without having all of the details and it's able to kind of surface that relevant information for you so that you can use that in a prompt. So yeah, I've, I've gone and I've explained a little bit about like all this ways of picking files and stuff. Like what do you do with that? You know. Right. Like that's like the big question. Right.

I think it's also good to explain like why it's important because I, I think one thing like, one thing I love about Repo Prompt, so when I first started using it, I had been, like, using just, like, a custom script I had created to, like, take my code base and, like, and then, like, put, you know, the relevant, you know, context in there, which a lot of times I was just doing all of it. I was literally putting all into a single file and I'd. I'd copy and paste that into ChatGPT O1 Pro when it first came out. And I think I tweeted about this someone and someone told me, like, oh, you got to try Repo Prompt. So I tried Repo Prompt. The fact that I could, like, see how much context I was sharing with the model was amazing. And it seems like that's super relevant too, because, you know, at least from the benchmarks I've seen, you know, everyone's talking about how much context you can put into their LLM, and now you're getting like, 1 million. And I think llama was 4 or something crazy high like that.

I think it's 1 million. Oh, no, it goes to 10. I think llama.

Yeah, okay, I'll be 10. I think on the benchmarks, though, right? I think on the benchmarks for llama 4, as soon as you went over, like, 128k context, like, nowhere near the 10 million, like, the quality just, like, dropped like, Like a rock.

Yeah, well, you know, like, until Gemini 2.5 came out, like, pretty much all the models. And I would say, Even including, like, 01 Pro, you would really want to stay below 32k tokens in general. I just, I. I find over that you're just losing a lot of intelligence. So there's this concept of effective context, the effective context window. At what point does the intelligence stop being as relevant for that model? And for a lot of smaller models and local models, it's a lot lower and you probably want to stay around 8k tokens. But for bigger models, 32k is a good number. It's only now with Gemini that you're able to use the full package, the full context window. And O3 is quite good, too, on these benchmarks, too, for context utilization. But, yeah, so you're using this context, you've picked out your files, say you want to use as many as you want, 100k. Like, what do you do with that? And so I have a question here. I'm just going to paste it to O3 and you'll see, like, what is O3 getting out of this? So it's getting basically this file tree. So it's getting a directory structure.

Of this project. It's getting basically the high level code maps of the files that I haven't selected. So basically when it's set to complete, everything that I haven't selected gets kind of shipped in and then you have the files that I did select. And so then the context is able to go ahead and. All right, there you go. So O3 gave like a really nice response here with tables, loves tables on how to do that. And so this is like a great way to kind of just get this information into O3, get the most out of this model. And O3 is an expensive model if you're trying to use it a lot. Like, this is a great way to kind of get more value out of it, move fast and get good responses.

Hey, if you take a look at my web presence online, it's safe to say that I'm a bit AI obsessed. I even have a podcast all about AI that you're watching right now. I've gone down multiple rabbit holes with AI and done countless hours of research on the newest AI tools every single week. Well, I've done it again and I just dropped my list of my favorite AI tools. I've done all the research on what's been working for me, my favorite use cases and more. So if you want to steal my favorite tools and use them for yourself now, you can, you can get it at the link in the description below. Now back to the show.

When I demoed this to Matt like many months back, the thing like blew him away. He was like, it can do that. Was just showing him that like you're, you're putting all that context together and then you're giving it instructions. And now you get one button to copy that.

Yeah.

And then you go to your LLM and then you paste that and you're pasting so much information. And I think the average person doesn't like people who are just using ChatGPT or people who are coding with cursor. Like people are just testing it out. They don't realize that you can do that, that you can literally just copy and paste all of that context in there and that the LM gets that and it understands what to do.

So I've just gone ahead and I've got Claude open. So Claude, you know, in contrast to ChatGPT, Claude is very good at following instructions. Like it's the best model at following instructions, I find. And so I've gone ahead and just asked me help change the dark mode color scheme. And I pasted that in to Claude like this with this XML copy button. And I think this is another thing that like repo prompt does quite well is that it's got tools to kind of send information into the LLM, but it's also got tools to go ahead and write an XML plan and it's going to create this theme selector and it's going to add these files and change files for me. And what's cool with this is that I can just go ahead and use CLAUDE with my subscription and then have it modify all these files. So it's basically creating all these files and it can search and replace parts of files too, so I don't have to re update and reupload the whole thing, have it output the complete code. So a lot of models struggle with, you know, people noticing like, oh, this model's really lazy, it's not giving me the whole code. But like this kind of circumvents that issue because it lets the AI just kind of get an escape hatch and just do what it needs to do here. You know, LLMs are beasts that seem to really like XML. So if I were to hit copy on this prompt here, I can open that up here. You'll see like, oh, first of all, all of the parts of the prompt are formatted with these XML tags. So there's the user instructions, you've got the file contents. If I scroll up, you know, they're all kind of separated out. So everything's structured in like an optimal way for the model. And you know, when you're outputting xml, the thing about XML is that it's like, it's, it's not like a nested structure. JSON is, you know, what people are used to looking at. But it's something that like models struggle to output reliably. If you're using like ChatGPT or you know, Claude on the web, like it's just not, it's just not going to like respect that formatting the way you want it to.

I think of it as like an information language. It's like that AIs can really simply understand. It's probably the best one, right?

And it's, it's easy for them to follow. So I have like examples in my prompt on how to output this kind of stuff.

Do you, do you check everything or.

I try to. If I, if I care about what I'm doing, like that's the, you know what Car Parthi's saying is like you gotta, you got you. When you're, when you care about what you're coding, it's very important to go back and forth. And you know, sometimes when I'm, when I'm coding like this, like, I'll iterate, like. So I pasted this question right with O3. And often what I'll do is I'll read through the answer and then I'll change my prompt and then paste again into a new chat and try and like, see where the result is different. Because basically I look at like, here's the output. Okay. I actually don't care maybe about this copy link button. Okay. Then I'll put, specifically put a mention in my prompt to say, like, let's, let's kind of just focus on this part of the question and kind of reorient it. And that's the nice thing with this, is that I can just hit copy as many times I want. If you're, if you're paying for a pro sub, like, there's no cost to trying things. There's no cost to, to hitting paste again. And except for that, you know the pace limit with, with O3 here. Yeah. So you just try again, you just paste again, let the model think again and try things. And I think that's like a really important way of working with these models is to experiment and try things and see how does changing the context, what files you have selected, your prompt. I use also these stored prompts that come built in the app. So there's the architect and engineer. And these kind of help focus the model. They give them roles. So if I'm working on something complicated, the architect prompt will focus the model on just the design and have it kind of not think about the code itself, whereas the engineer is just the code. Like, don't worry about the design, just kind of give me the code.

Maybe you should explain, like, when you say engineer prompt, it's literally you're just adding stuff that you copy and paste into the LLM, saying like, you're an expert engineer. And this is what I expect from you. I expect for you to give me xml. That's your job. Do it. And that's literally how the LMS work. And they're like, okay, I'll do it.

Absolutely, yeah. Giving them roles is crucial. Telling them who they are, what their job is, what their job description, what do I look for? Like giving them a performance review, evaluation, all that stuff. Like, I find like, the more detailed you are with your prompts, the more you can help. Like, they kind of color the responses in an interesting way. So just adding the engineer prompt, you See, like it spent more time thinking about it, but actually that's partly because it's trying to add call tools that are invalid, but just as a contrast to the old response. So here, this time it kind of said, okay, this is the file tailwind, here's the change, and this is the change that I'm going to do in a code block. And so like just having it kind of answer. So, you know, for the longest time before I had any of these XML features, I was just kind of using repo prompt and like getting these kinds of outputs and then just copying them back into my code base manually and kind of reviewing them. I was like really the antithesis of vibe coding, where everything's kind of automated. Yeah. So you know, like, I think, you know, the working on repo prompt, it's really like there's, there's building your context, that's like the biggest thing. Just picking what you want, you want to front load that work. And you know, in contrast to using agents, you're going to have those agents kind of run off, do a lot of work, call a bunch of tool calls. You see like O3 kind of thought for 15 seconds, thought through some tools to call. It didn't really make sense, it just kind of kept going and ended up doing this. And if you've used cursor a lot, you'll see often using O3, it'll call tools that will read this file, read that file, read this file. But if you just give it the files up front and you just kind of send it off to work with your prompts, right away you get a response and you're like, okay, well does this make sense to me? Am I able to use this instead.

Of letting it a little bit more work? At least right now, but it's. Yeah, I think you get a lot better results.

Yeah, yeah, Just front loading that context, being able to think through and iterate on that. And that's the whole philosophy around it. It's just like thinking through, like making this easy. The context builder helps you find that context. You know, eventually I'm going to add MCP support so you can query documentation, find things related to your query as well, and just spend time as an engineer sitting through what do I want the LLM to know and then what do I want it to do? And then make that flow as quick and as painless as possible. And like that's kind of everything. And I think, you know, going forward and you know, as you get serious coding with AI, like that's, that's what the human's job is in this loop is. Engineer's job is figuring out the context. I think that's, that's the new software engineering job.

I'm just, like I said before, I'm so surprised a lot of people haven't talked about this because, like, for me, like right now, cursor is good for like something very simple, like, okay, change some buttons or change some links or change whatever, you know. But anything complicated repo prompt, I got like, way, way better results. So I'm curious, like, you know, have you ever thought about like this being used for things outside of coding? And do you think it would be useful for anything outside of coding? Cause I kind of think of like a. In a corporate context of they have all this information they're trying to use. Let's say it's for, you know, and there's different divisions of a company. Like, let's say one's like a marketing division. It feels like there's probably like, you know, instead of like just trying to feed all of the company's information to LLM, probably they should have like the relevant, like, okay, for this kind of content, what are my relevant marketing guidelines, the brand guidelines, other information that's useful to the LLM to tell it, like writing style and things like that.

Yeah, I mean, I've gotten academics reach out to me telling me they're using it for their work. There's folks in different fields for sure. Like, in general, like, you know, if you're working with plain text files, you know, repo prompt can service those use cases for sure. Like it's, it's all set up to kind of read any kind of file and then apply edits to any kind of file too. Like, I don't differentiate, you know, distinguish or anything. Like, if it's, if I can read it, then I'll. I'll apply edits for you. And yeah, like, I think a whole bunch of work is around just like gathering context and kind of iterating on stuff. Like even, you know, in doing legal work. I do think though, there is a flow that is still missing from this app. It's just that kind of collaborative nature. I think there's still some work that needs to be done to make this a more collaborative tool, make this a tool that syncs a little bit better with different things. For now, developers use Git and that's that kind of collaboration bedrock.

Yeah, I mean, yeah, that's something I think too is like, yeah, repo prompts super useful, but you have to Be a little bit more advanced, like, than the average vibe coder or the average person using an LLM. And I'm kind of curious, like, why did you not go the VC route? Or like, like. And where's Repo prompt at right now? Like, what is your kind of, you know, where is it now and what's your plan for it?

Yeah, I mean, you know, I. I've had a lot of folks, you know, kind of, you know, bring that up to me, and they're kind of thinking through, like, you know, why not VC or whatever. And I think, like, I mean, it's not something that the door is closed on forever. It's just. I think right now it's. It's working. I'm able to build and, you know, I'm able to. To kind of listen to my users and, you know, pay attention to what they need. And I think, you know, it's just not super clear to me, like, where. Where this. Where this all goes, you know, like, this is an app that is, like, super useful, and it's. It's like, helping me, and I'm able to build it. But, like, is it. Is it something that necessarily makes sense to, like, have, like, you know, $100 million invested into it to grow a huge team, to, like, build maybe? I don't know.

So right now you're still working at Unity, is that right? So you're, like, working at Unity and then building this at night? Well, like, yeah, family.

So I was in paternity leave for a huge part of the last year, and that's kind of like, a lot of the time I was able to build this out on. So I. So, you know, raising a child, you know, that's. That's a lot of work, too. So, yeah, a lot of it right now is still just, like, you know, nights and weekends, spending a lot of time on it.

You know, I feel like everything with AI right now is like, who knows what's going to happen? Like, yeah, in a year, everything could be different. In five years, who the. Who the hell knows, right? And in that kind of situation, I think there's a lot of risk, like, right now, because AI is such a big wave. That's why we call the show the Next Wave, right? It's such a large wave of transformation happening that you are going to see the largest investments ever, I think, in history, as well as the largest acquisitions ever. And I think these have yet to come. Like, people think, oh, it's already. These have yet to come. We're, like, in the early part of this transition. And so in that situation I think you definitely should consider it because there's going to be so much money thrown at the competitors and I think people like Cursor, Windsurf, maybe Windsurf's getting bought by OpenAI possibly for a crazy amount of money. They'll probably try to integrate Windsurf. They'll try to probably integrate repo prompt type features into their products at some point I would assume the best two routes for you in my opinion would be either to go really big and go the VC route or to go more like hey, who knows what's going to happen with it. I just want to get my name out there and I can leverage my name for something else in the future and open source it. That's my kind of thought on strategically what I would do is either go really big or open source it and make it free and just put it out there and say and get some reputation benefit from it.

Well there is a free tier. It's not open source, but there is a free tier. The thing about open source actually is something I've thought about a lot and the big issue with it right now, especially as people are building AI tools is that like it's never been easier to fork a project and kind of go off and just build it as a competitor. Like if you look that's what I'm saying.

Like, like I think that you're like giving up on like the monetary side of it. You're literally the reputation like hoping to get a reputational benefit from it versus monetary.

But even that reputational side, like, like if you, if you looked at Klein like Klein's a big tool, you know that came around actually started around a similar time as me working on repo Prompt. Um, and if you're not familiar the folks listening on the call, a client is a, is a, is a AI agent that sits in VS code and, and, and it. It's pretty cool. But the thing that is not so cool about it is that it eats your tokens for lunch. Like that thing will churn through your, your, your wallet like faster than any other tool that exists. Just because it goes off and reads files stuffs the context as big as possible. You know if, if like Cursor is trying to like be conservative with, with what it serves the model to save on their inference costs like Klein's the complete opposite and will maximize your costs as much as possible.

Sounds like an open source project.

Yeah, yeah, but it's really powerful. It's very powerful. It Does a lot of good. So a lot of people really enjoy using it because it does good results for certain things. But yeah, that cost is very high and. But the thing that I was trying to bring up with this is that like, so Klein was actually forked a few months ago by another team of developers and it's called Roux. Roux is the alternative. And a lot of folks, you know, if you look at open router in some stats, like Roux is actually surpassing Klein. And so that fork is now overtaking the original. And that's the kind of space that we're in where different teams will kind of take your code, take it in their direction, and then all of a sudden they've overtaken you and you kind of lose track of where things are going there. So it's a crazy space. And it's never been easier to open pull requests. With AI, you don't need to understand the code. You're like, oh, I have this open source project, I'm just going to fork it and add my features and kind of go. And it's a tricky thing, but like, you know, having a free version and kind of trying to ship and, and try and grow a community of users who are passionate will like, can talk back to you and you know, I mean, that's kind of the route I've taken right now. And it's, it's kind of been working so far. I was in beta for a long time, but yeah, you know, leaving beta, I, you know, it's, it's like, you know, still new, figuring out where to.

Go next and it's Mac only right now. Is that correct?

Yeah, that's, that's true. It is. Only Mac is Mac only. And I think a part of that is, is that I started off, you know, just kind of trying to think about like, you know, how do I build this in a good way. Like, I started with Electron actually building this out and the problem is like, I immediately ran into issues trying to build for different platforms and like I spent a bunch of time debugging, just getting SVG icon rendering, you know, all these little things that are just like rabbit holes and you're like, okay, well you're so abstracted from the base of like what's happening and you spend a lot of time just solving build issues that it's like, I'm just going to go ahead and do build native and just run with it and have better performance doing so. You know, if you open an IDE like VS code, you open up like a huge Repo, what actually happens is that it'll load the file tree and it will just kind of lazy load everything. Like not everything needs to load because if you're opening an IDE as a coder, traditionally you only have a couple files open at a time. Maybe you have a dozen, but like, that's at the most. Like you're not going to be processing 50,000 files at the same time. But an AI model can, you know, if you give it to Gemini, like, and Gemini will want all those files. It'll want as much as you can give it because they can read all of it. And so like, you need a tool that is built different, that is kind of organized in a way where it's kind of thinking first through that performance of mass data processing that you need to kind of do. So, yeah, it's a whole different way working. That's why, that's why it's native. Because, like, I want that performance. Processing all these files, there's all this concurrency happening where you're like in parallel editing these files, like processing them and doing all this stuff. Like, you know, it's very hard to do if you're just using JavaScript or TypeScript.

Yeah, I assume so. Like, I thought about that. Like when I use repo prompt, it seems like you've done a really great job of building it. It works really well. And it is all just you, like right now.

Yeah, it is just me. Yeah, I've been working on it a lot.

Yeah. That's crazy. That's crazy.

Yeah, it's come a long way. I iterated a lot on it, you know, but like, you know, the first version was super jank. It's just, you know, I. But that's the power of dogfooding too. Like, if you're not feeling like folks listening. Dogfooding is when you like kind of use your own product to iterate on it and build with it and you kind of like make it. Make it a habit of making sure that you're a number one user of your app, you know, your own product, to make sure that you see all the stuff that sucks about it. And for the longest time, like, you know, it's really sucked and just that struggle and that pain of using it and forcing yourself to feel that pain, like, that's what makes it good. That's. That's where you're able to kind of feel. Feel those things that the other the user is using in the Apple field. And. And that's when you end up with something that, that is great in the end.

So where do you think re prompt is going? Like, long term, it's weird, you know.

Like in December, like OpenAI announces 03 and they're like, oh, it beats all the arcade GI tests. And you're like, well, is this AGI? Like, what is this? Like? And then, and then it shifts and it's like, okay. I mean, like, it's a better model. It lies to you. It's. It's not like the Messiah, you know, so, so it's hard to say, like I, I don't, I don't know, like, where we go. Like, I have ideas on like, you know, using LLM so much. I have ideas on like, where the future is one year from now. I think, like, you know, I'll have to adapt this product and keep iterating on it to kind of stay relevant. So it's going to keep changing. But like, I think that the flow of kind of, I'm kind of pushing towards, of that context building, I think that remains relevant for a while longer. And, and what improves is like the layers of automation around that work. So I think like, long term, I still think that is kind of the vibe that I want to go towards though, of course, like, you'll have to, you know, have an agent that kind of like does more research for you on your behalf. That will kind of run off a little bit, being better at doing parallel work. You know, some folks using CLAUDE code, like, they, they have this thing called Git work trees that they use where basically they, they split the repo into like parallel universes and they let like Claud code kind of try things on these parallel universes and iterate and go off in the distance. And that kind of workflow is going to be important, very expensive workflow, but it's one that I think is one of those ideas that you'll kind of want to explore and go down that rabbit hole. So I think just like integrating mcp, just embracing that universality of all of these different tools. So for folks listening, if they're not sure what is, MCP is another acronym.

We got lots of AI.

Yeah, yeah. I mean, so the idea there is basically, traditionally, if you use like Claude or OpenAI, they have tools and those tools, you know, one of them could be like search the web or one of them could be like read the fil on your thing or look up documentation or these kinds of things. And there's this protocol MCP that creates an abstraction layer so that any client app can implement this protocol and then Users can bring their own tools. So if a user comes in and says, oh, I want to use. And there's this new one that's really cool, it's called Context seven. Where basically they've gone ahead, built a server that fetches the latest documentation for whatever programming language you're using and will kind of pull that in as context. So you can say, okay, great, Fetch the latest Angular docs or whatever docs you care about, and then you can bring that in. So that kind of work where you're like doing that, that, that context retrieval, that's super important. Or like Stripe has one too, where basically all the docs for their tool is set up and you know, you just plug in the Stripe mcp and then all of a sudden, if you're trying to vibe code your way through integrating Stripe, like, that's super easy. The work is kind of handled. You can plug in your API keys onto it so it can even talk to the backend for you. That whole work is kind of automated. So it's really like, it's all about like having tools for, for folks using these models to kind of automate connecting to different services in this like, universe of all these different, you know, I.

Just kind of think of it most. I mean, it's different than xml, but for me, the way I think of it, because I'm like, I'm a person who's like, yeah, more on the business side, slightly technical code a bit. You know, I think of it as almost more just how XML is like the information language that I can understand. The MCP is like the same thing with any, any service you want to use or tool. It's a way that you can for the AI to know how to work with those things very clearly.

Yeah. And funny enough, you mentioned xml, because that's actually one of the things that I do a lot with Reborn is like parsing xml. And I've shown with that and I think one strength there that I have that like a lot of other tools are kind of ignoring. So traditionally when you're, when you're working with these language models, a developer and you say like, and you see, and you can see this, if you use ChatGPT, you'd be like, hey, like search the web. It's going to use the search tool and you'll see it say you called tool search and it'll go through. But what happens when it's doing that is that basically it calls that tool, it stops, waits for the result, and then continues. And when it continues, it's Basically like if you think about it like in this weird way, like how tool calling is. It's funny like I think about like you have this, the robot is kind of being reboot as a new session with that new context. Because basically every tool call is a new query. So you're giving back the old information, but you're not necessarily talking to that same instance. It's like, it's like a different AI instance that is answering your question from the new checkpoint. So like that's like a weird thing. So you know, as you're, as you're making all of these tool calls, if you, if you use cursor, you know, It'll make like 100 tool calls in. I think they have a limit of 25 actually. But by the end of it, you know, you've gone through 25 different instances of these models and then you get a result at the end and you're like, well, you know, it's like weird like what actually happened? You know, there's some data loss, like weird stuff. You know, we don't know how.

It does seem like that could create like reliability issues, right? Because like, you know, the LLMs, like sometimes they give you amazing results and other times it's like what is this? And so every time you're doing a new tool, it sounds like you're almost recreating that the chance of it going wrong in a way.

Exactly. Yeah. You're aggregating this issues. You don't even know where that info. It could be different servers that are actually processing all these different tool calls. And maybe it's weird, sometimes you'll have like, oh, that server has some chip issue on its memory and that actually causes some weird issues where Claude is actually really dumb today. But on the other one it's a lot smarter because their chip, the memory chip is working fine. You know, you don't know. So that kind of thing. So the way that I've kind of gone about this is that like the way I call tools is you, you have your XML and the AI will just answer in one instance and, and, and it'll just give you the whole thing. And then I parse it and you can call a bunch of tools in there. It can be like, hey, like I want to call this, this, do this and this and then I just parse that and then bulk call the tools and then get the results and then we go another instance with the results and you can kind of go back and forth like that. So like not have to wait on each single One, you're actually just bulk sending them out, getting that data. It's a lot more efficient. You're able to process say like 25 queries, you know, get read 25 files, bring them all in, you know, let's work from there and see how it goes. And so that kind of thinking. So I think there's a lot to kind of play with in terms of, you know, how you're even getting this data back and forth from the LLMs, because at the end of the day it's all text, text and images, maybe some video in some cases, but like really text just for your coding. Like that's, that's the thing that you're working with and you can do a lot with tax, manipulating it and playing with it to kind of get good route.

What do you think? Like I've heard, you know, YC and others, I think Gary Tan said that, I can't remember if it was 80%, but I think he said like 80% of the code for the startups going through YC right now is AI generated. That number could be wrong. Like where, like do you think in three years from now, like, like in three years from now, do we still have like normal engineers who don't use AI at all? Is that a real thing?

Well, first of all, like I think saying a percent like that of how much of it is AI generated, it's a bit misleading because like could be higher. I can basically like, no, but like I can go ahead and like every line of code, I could basically like type it like in pseudocode to the AI model and like have it paste it back in as like a fleshed out JavaScript function and say 100% of my code's written by AI, you know, and you know, if you're, it really depends on how your workflow is, what your pipeline looks like. I do think fundamentally the job of an engineer has changed. It's already done, it's already completely different. You can't work the same way. But it depends at what point in the stack you're working on. I work with some folks who do some really low level graphics work and I talked to someone about how they can't really use AI models because the AI model just hallucinates everything. It's just not trained on anything that they work on, so it's just useless for them. But then if you work on someone who's a Web developer, well, 100% of the code, 98% of the training code is web code and web framework code. And so it's like, okay, well, yeah, 100% of that work can be done by AI. It's really easy. But I think as we move forward more and more, you're going to want to have AI models thinking through hard problems for you because it just happens much faster as they get better at math, at solving connectivity. Architecture. Architecture is something that these O3 and O1 Pro and hopefully O3 Pro just excel at. They're very good at finding good ways of organizing your code and helping you plan how you connect things together. The job is fully changed, I think from today on. If you're not using these tools, you're not learning how they work. I think that's an issue because I don't think a traditional engineer who spends his whole career just typing code up, that doesn't exist anymore. But what does exist is someone who understands code and who can read it and who understands what questions to ask. If you're prompting about the code, if you understand the connections, that's where you're going to get the best results. And that's why a tool like Repo Prompt is so helpful, because you're able to do that and feed the right context in. But if you're just saying make the button blue or move it over here, I mean, that works to some extent. As long as your instructions are simple enough and you know what you want, you can get there. But at a certain point you fall off and that's when it stops working. And maybe that point where you fall off gets further and further as the models improve. But I don't think that in the next 10 years we get to a point where that point stops existing.

One thing that we didn't talk about, that I was kind of curious to talk about was like, what do you do at Unity? Like, and like a crazy backstory, like. So I used to be involved in the game industry and I used to see David Helgeson, the creator of, you know, Unity around parties all the time. He's super nice, super nice guy. Like I had some kind of like image in my mind of him. It's like this eccentric, you know, European billionaire who had like the popped up cuffs and he'd always like, you know, he'd go around with loud techno music with the roof down. And it's crazy how big Unity is now.

Yeah, so what I do there is, I've been doing kind of XR research and XR engineering. So I work on a toolkit called the XR Interaction Toolkit. And basically what that is for is it's A framework for developers to build interactions with xr. So if you're putting on an Oculus Quest or you're a Hololens or, you know, like, whatever, you know, Apple Vision Pro, you want to basically interact with objects in your. In your scene, in your world. You know, like in ar, if you're walking up and you want to pick up a virtual cube, like, how do you. How do you process that interaction of you grabbing the cube and picking it up and looking at it? So that's like. I've done a lot of research on that, that interaction of, like, playing input hand. Like, I written specs that are adopted for, like, the industry in terms of hand interaction. So, like, you know, just tracking your hands, how do you. How do you grasp something? What should you be doing? If you want to poke a button that's, like, not there, like, what does that look? So that kind of stuff.

That's.

That's what I do there.

That's amazing. That's like. That's a really complicated engineering work. I'm. I'm constantly. This. Actually, one of the reason I even, like, you know, helped start this podcast was I'm constantly thinking about where AI is going and wanting to stay ahead and also think about what does it mean for me and my family, like, with your child, you know, have you thought about that yet? Like, oh, of course. What do you think they should learn? And, like, and. And how?

I have no idea. Okay. Okay. I have no idea.

This is everyone, right? Like. Like, what do you even teach your children? Like, is it. Is it important to learn to code? Is it. Do we teach them logic, morals? Probably all of this and more, and being flexible and super fluid and.

Yeah, yeah. You know, but it is funny on that topic. Like, I. I look at engineers coming out and learning to code with AI around, and I think they're at a disadvantage. You know, it's. It's unfortunate that, like, you know, if you're starting to code today and you have AI to lean on, you just don't have that struggle. You just don't have the pain that, like, I had to go through when I started to code. When engineers who've been in the field for so long that had to struggle and not get the dopamine hit of a fixed problem right away to study it and understand how it works, that just doesn't exist anymore because the AI just solves it for you. And I think that's true in the code, but it's going to be more and more true in every field. And So I think.

If you want to really get into something, I think there's going to be a need for people to have the restraint to put aside these tools to struggle a little bit. I think there's a ton of value in using them to learn and grow. But there's also like that restraint that you need to form to kind of have the struggle because that's where the learning is. And it's really tricky. And I don't know how you solve that now because it's too easy not to struggle now, which is a big problem.

Yeah, I've heard Jonathan Blow, I don't know if you know of him, the game designer, he talks about exactly what you're saying, that it's in the future. Like, yeah, sure, AI could get amazing at coding in the future, but it's also going to create an issue where, just like you said, people are not going to learn to properly code. He was already complaining about before AI, the code was shitty. And then now with AI, it's like, oh, okay, now we're kind of screwed, I guess, because like, if we're in a situation where like no one knows what's going on and like, yeah, you're entirely dependent on the AI for everything.

Yeah, but that's the thing. I think maybe that's the middle part where we're at this point where it's like the AI is just not quite good enough to kind of solve all the problems. And you still have problems to solve and you still have people that need to kind of work with the machines to kind of figure out how to go. Maybe at some point in the future all of it is moot. And I know some folks think that, and maybe it doesn't matter, but like, I think, you know, there's going to be some discomfort in the middle where, where you know, the machines are not quite good enough to solve every problem. We lean on them as if they are and then, you know, we're kind of atrophying a lot of skills that we've heard. I've heard folks, you know, I haven't driven in a Tesla with fsd, but I've heard folks say the same thing there where like, if they're using it all the time, they actually like suck at driving without it. And it's like, right, you know, like more and more, that's going to kind of be a thing where, yeah, you.

Know, that's, that is the thing where we start to like, to like. We're like almost living in like one of those sci fi novels. Right. Like, everything being super, super safe. You know, I live in Japan. Everything. You used to live in San Francisco. Everything's super safe in Japan. And there's one reason I like it, but you do lose some. Some freedom in. In that. So. Eric, it's been awesome and maybe we should, you know, tell people where they can find you and where they can find repo prompt and.

Yeah, so. So I'm, you know, a puncher with a V on exos, like pvn, C H E R on X and most, most socialists. My handle all over. So you can reach out there. My DMs are open if you have questions and repo prompt, repo prompt dot com. So you can just head over there and find the app free to download. And nice Discord community too, if you want to hop over there and send me some messages and tell me what you think, like, please do.

Cool. I think I'm on there too, which I'm not very active, but.

Yeah, it's all good. Yeah. Well, thanks for having me on, Nathan. It's been great chatting with you and it's been great. Yeah, appreciate it.