Transcription
Context engineering is the biggest buzzword right now. And I actually have the king of context engineering, whether you believe it or not. And his name is Eric Provencher. And Eric is actually the founder of this app called RepoPrompt. And for those who don't know the OGs of AI coding, Eric came up with this really, really brilliant app and a way to actually talk with your code base and grab as much context as possible and allow you to rearrange things so that way you can get the most out of these language models.
Whether you believe it or not, whenever you use any of these AI coding apps, Cursor, Windsurf, you name it, they're grabbing your code base and kind of rearranging and grabbing those pieces out. And then they're sending those things over to the language model so that the language model can give a response. And so you'll notice sometimes after various updates that you may not get the output that you really want.
So Eric had this problem and said, I should make an app for this. I want to see the tokens that I send over to the language model. And if I rearrange my responses and do things in a different way, I can actually get a better response from this. And this was something that, you know, he discovered many moons ago, almost, you know, it's probably a year ago, but in AI, it probably feels seven years.
And so I'm extremely delighted to have Eric on the show because we're going to talk about RepoPrompt, Context Engineering, MCPs, and many more things. And since this is a live stream and you're actually watching this right now, feel free to drop any of the comments. And we're going to go ahead and have a Q&A section towards the very end of this episode. So I'm going to make sure I go through those. I'm going to prioritize my members first. So big shout outs. This whole stream is powered by my members and people who support the channel. So yeah, without further ado, Eric, thank you so much and welcome to the show. I'm excited to have you on.
Thank you, and what a great intro! I really appreciate that. Yeah, man, you have really been cooking with your app. And for those who don't know, RepoPrompt, repoprompt.com. Tell me a little bit more about kind of your app and as far as latest updates and kind of what's been happening in this, in this world of context engineering and stuff.
Yeah, yeah. I mean, so as you were saying, I started this up a long time ago at this point. It has been actually just about exactly a year since I started working on it. And, you know, at the time, you know, the space was very different. Cursor was very nascent compared to where it is now. You know, the best tool, Sonnet 3.5, just released when I started working on it, and that's when it, the timing became better, and it was easier to, to kind of work, you know, with low latency responses, and you can get a lot more done in a moment.
And the thing is, at the time, the only way to get a good way to use the full context window of Sonnet was by using the Cloud Web app. You would go onto the Cloud website and you would put your prompts in and you would get a response. And that was the only way to use the full 200K tokens. If you're using Cursor, you're at the time stuck to 30K tokens or something like that. So it just wasn't the full experience of the model, and that's why I built this app.
So the app was very simple at the time. You just open up a folder, you select the files you want, you type in instructions, and I would take that and put it in a big XML block that I would paste your clipboard. You can paste that in, you would get good responses. And for a while, I would just code that way. I would just see what the model said, I would go in and I would manually paste responses in.
And then that was the second problem. It's, okay, well, how can I get the responses from Claude back into my code? And so I started working on a feature for this diff apply stuff. And there's a whole XML feature that's in there now that is still great. But now the latest thing with everything supporting MCP is being able to just have AI agents talk to the app directly. And so a lot of the features I've been building over the last year are now fully exposed over the MCP tooling so that you can have Claude kind of use the app as if I was. So it can select files, it can build prompts, it can write my prompts, it could access the file tree. It even has some more powerful tools, search that are, you know, what you get with Claude Code. And in some ways, it's actually a little bit better because I can actually have multiple repos open and have it reference all kinds of stuff across my repo with its search. And even using the chat directly, so I can have O3 in the app and Claude Code can go have a conversation as if it's invoking chat GPT itself. And it's delegating work to...
Bro, hold up. You're cooking super hard right now, and there's literally a five-alarm fire in this kitchen. We need to kind of unpack...
That's fine. I've been going on. You've been telling me what I'm doing. So I'm getting digested.
Yeah. Bro, bro, bro. Okay, okay, okay. For those who are saying, he's basically saying that with his app, when you have it loaded up, you're going to have your code base loaded up and everything. But one of the things that has changed in recent times from what had happened before and what makes this really killer today is the fact that you said now an agent can go into your app. And it can reason over multiple code bases. And so for those who aren't familiar, RepoPrompt will let you allow you to select basically as many code bases as you want to kind of load into its workspace. And then from there, you can actually select the language models that you want to have via either an API, via copying a prompt and putting it into a chat window or something like that. But beyond that, in the agentic workflow, you can also have various models apply changes to your code as well. And so you don't have to be stuck with one specific model. It could be as simple as, you know, just doing one single API call, having a chat, or you can actually go into very little details of saying, I want these specific models to do those different things. And those tunables are actually done by you, the engineer. And so if you really want an extreme power tool, yeah, you have this in this one type of thing.
So tell me more a little bit about this agentic workflow that's kind of new. It's new to a lot of workspaces and coding things. And to me, a lot of people are talking and trying to find alternatives to agentic coding because they don't want to pay $200 a month. Right. So, yeah. Tell me more about kind of what this looks like for your app.
Yeah, I mean, so the key thing is, right now, if you want really good value, you have to kind of pay Anthropic for the Cloud Code, just a Cloud subscription. You can do Pro if you want, just the $20 a month use on it, or you go to Max if you want to get into the weeds with Cloud Code. And that can be expensive, $100 or $200 a month. But there's really no alternative to paying them to kind of get a good agentic workflow because Sonnet is really the only good agentic model at the moment.
So what does that mean? What is an agentic model? The thing is, the way these tools work is an agentic loop is very simple, really. You just give some tools, a bucket of tools, a toolbox to a model, and you're, here's your tools. Here's my prompt. Get to work. And what is it going to do? So it's going to look at the tool set. Okay, what tools do I have here? So in RepoPrompt, I give it, okay, you can pull out the file tree. So what's the file tree? We've already loaded your repo. We can go ahead and see all the files and folders in your directory. Okay, it gives a good idea of what the project is. Okay, so from there.
Okay, so Claude is, okay, let's go look around. You know, you asked about this prompt feature. Okay, well, what files seem they might be related to the prompt feature? Oh, well, let's do a search. And it'll do a search and it'll figure out which files are where. And then, you know, the search will return, okay, well, here's some lines of code that are referenced in your thing. Let's go read those lines of code with a little bit of extra window. And then all of a sudden, you know, it has an idea of what's going on in your code. It has some idea of how things are working, and then it can go ahead and actually make changes to your code.
So you're just giving it the tools to go discover what it needs to know and then actually making changes. And the thing that's really interesting with Claude versus other models is that it's really built to figure that out itself. Whereas if I'm using O3, I find it performs better if I do it and give it tools and then it reasons over it. Because if you're letting it kind of discover things, sometimes it'll do an okay job, but Claude is the one that is really engineered and reinforcement learned to make sure that it's able to call tools very precisely and well. And it doesn't being given too much context. It's, no, no, no, I don't want what you gave me. I just want to go find what I need to find. So it validates, it double checks itself. It'll read things multiple times. So it's not super efficient sometimes. But yeah, that's the thing.
So yeah, I just wanted to give that primer there. That's quite amazing. And if you have any demos, I would love to see anything that you've got pulled up. Maybe we can kind of get that screen share going on. And in the meantime, I can kind of share a little bit for those who are wondering. This is kind of the RepoPrompt website. And actually, Eric has some actual videos and stuff that that are on here. So, you know, talking about the visual context engineering, there's a cool clipboard XML workflow, there's a lot of different workflows, and it kind of gets all over the place. And the AI context builder is kind of something that you're talking about a little bit about just now, right? About it kind of going into your code and figuring out that context. And, and for me, you know, the prompt layer library is actually pretty awesome because this is where you can actually get those architect prompts and the engineering prompts. And I actually have a demo video where I'm actually comparing, you know, Grok versus O1 Pro at that time, and then the other models. And I, to use this because I actually can feed the language model, which is hungry with all these tokens, all of the direct input into Grok and O1. And at that time, Grok actually took the top spot for architecting a code base for, you know, I'm going to make a database, what are all the different relationships and all the different entity diagrams, and it just really took right over. So yeah, this is really, really cool.
Let me go ahead and see if I can get your screen up on here. And I, oh, I think it might require Chrome on your side. And then the, the screen share type of thing. Yeah, I forgot I'm in, so it doesn't work there. Oh, it should work. Yeah, it's a Chromium-based browser. So yeah, I, I don't see the button on my end.
Oh, to share screen? Yeah. I see, I see. Okay. Y'all, maybe? We have to come back. Whoops. Let me. All right. Let me see. Okay. It's all good. This is live, folks, so that's how you know we're cooking live. Yeah, yeah, yeah. No, it's all good. So, yeah. And let me see if I can pull up RepoPrompt on my end, too, and then kind of have that in the demo here. Kind of do some things.
Yeah, I've got it ready to rock if ever I can share my screen. But let me see. I wonder if you can open a new browser. Yeah. That's what I'm going to do now. Okay. Cool. I'm going to go into Chrome Direct. Chrome. All right. There we go. Cool. Let me just go ahead and let you in here. Oh, yeah. Cool. I'll have you take over the guest one here and then I'll take over. Yeah. It's, it's you never left. Okay. We're good. We're ready to rock now. Okay. Yeah. Yeah. Sounds pretty good. Okay. Dang. That was pretty smooth. I appreciate everyone for hanging out with the, the technical assistance here. I'm just kind of updating my last screen there. Perfect. So that I get your name. And then, what's great about this whole thing is that I'm able to then just get this window up and then I'm going to go ahead and assign this into this spot and here we are. Here's your screen. This is great.
Okay, great. So just to give context here, so what I did. So I've got the RepoPrompt open here. It doesn't have anything selected. It's just got my website open and then I added the RepoPrompt project in here as well. So I just said, hey, can you make a call? Just take a look at what's here. And it was, okay, great. So I got a really good understanding just from the file tree of the website and the app. So I can ask it, can you take a look at the documentation specifically around MCP and read the code in my repo to update those docs with the latest tooling? Is where most of the action lies.
All right, so I can have it get to work. So it's able to see all of the repos I've got open. So right now it's doing a search. So, you know, you can look for paths. It can look for content as well. So you see, Claude likes to make multiple searches. It finds results. You can see, it's getting all these file paths, the relevant line in code. It's making some selections. So it just updated the UI, actually, which is interesting. Wow. And now it's able to bulk read files because of the file selection. Yeah, sometimes you'll see, actually, the UI updates before Claude because it's doing stuff and there's a little lag on the Claude desktop, which is pretty funny. There you go. See, it's just getting to work. It's just using the app you would, getting to work. It's finding the relevant tools. So see here now it's, reading code about RepoPrompt specifically, before it was reading code about the website. So now it's a good understanding of the current code base and, you know, the website.
Now it's actually I don't want to do the work myself. So it's asking O3 and it's set to O3 high, which is probably a bit intense. Yeah. So, you know, so sometimes, you know, it will it will try and use things kind of a little bit, a little bit funny. There's still some prompting to kind of figure out in terms of the best way for it to use the chat tools. But, I do find this pattern where it will try and delegate work. Claude is very much, oh, I don't really want to do that work. I'll just ask the AI model to do it. But, the thing with this is that it needs to kind of make sure that the context is sufficient. So now it's just kind of gave it a bit of a vague task, which is not, this is not optimal. But, yeah, you can see, it's able to kind of use both models. It's able to do the file editing work. And then Claude can review what was changed and it can actually make further edits. So Claude has its own editing tools here. And yeah, so you can read all these things. So yeah, that was auto-applied. The chat completed. And then Claude should be able to see the response there and continue from there.
Whoa. So kind of break this down in slow motion from the beginning. You have Claude Desktop set up, and that's configured with your RepoPrompt MCP, right?
Exactly, yeah. So Claude Desktop becomes an MCP hub, so you can actually have different tools. So one I actually really like is the Xcode Build MCP. A lot of people, context, seven. And the thing that's really great with this is that, you know, basically you're able to make Claude Desktop into Claude Code as it can get to work on doing, you know, pretty much everything Claude Code can. Because it's able to kind of read files, do edits, and do all that work. And so you're able to kind of turn this just into your hub. And, you know, Claude Code is great. Not everyone wants to use a terminal though. And, and the terminal, you know, you have one directory at a time, so there's some limitations there. But yeah, no, just being able to give it the tools it needs to work, Claude Desktop is actually a very powerful agent driver at the moment.
Oh, I had no clue, dude. So you're saying I can use Claude Desktop right meow?
Yeah, Claude, Claude Desktop in this whole piece. And Claude Desktop is the MCP that connects to RepoPrompt. So that's going to basically sort of behave like Claude Code. But you're saying Claude is really good, it could already start acting like someone who is, you know, what Claude Code already does and says, oh, I need to do this coding task. Oh, I see that MCP server. I'm going to go get everything I need to do. And it follows the same agent that workflow.
Yeah, you're using the model, you know, and they prompt, they prompt Claude, they prompt Claude pretty well. The thing you won't get with Claude Desktop in this workflow is just having rule files and certain things which are quite useful. So I think there's some scaffolding to be built around that. But, but yeah, it's funny, I'm seeing as it's using this now, it's just really likes to kind of just delegate the work out to O3. It doesn't want to do the work itself, which is pretty funny. You can turn off certain tools. If I were to turn off the chat tools and try it again, it would just kind of just use it directly. But see, it's telling it exactly what to do. It do those edits. Okay, great. It doesn't want to handle those things. So it's able to kind of follow along. Because RepoPrompt updates its files right away, there's just a file watcher. It's able to update everything, so everything kind of stays in sync. And what's cool is...
Is there a way to make your windows a little bit bigger or something like that? Maybe, I think some folks are saying.
Yeah. Yeah, on that side. Yeah. I don't know if this font scaling is not going to...
Let me scale up the font, actually. Appearance. If I go large. Large, yeah. Yeah, there we go. And does that one do command plus or something for Claude Desktop?
Yeah, there we go. Perfect. Okay, cool. Yeah, thank you for the comment, by the way, Carlos. Yeah, and some other hooks. Yeah, yeah, yeah.
So, so yeah, I mean, you know, at the end of the day, yeah, it's just gonna keep going. Whoa, that's the thing. Sonnet, Sonnet just keeps keeps working as long as now it's finished, finally. So you're gonna see what's this cool? It's gonna see the prompt just update dynamically. It's telling me what it did so I can review it. Perfect. There you go. So it got to work, told me everything it did. Yeah. So it just updated my site. It did quite a lot of stuff actually. I didn't finish updating there. But yeah, that's the thing. So you can give it a bunch of tools and being able to have leverage different models is really useful.
The thing that a lot of folks will say right now in the current state of agentic tooling, you know, as I was mentioning, Sonnet is the one that people use a lot for agentic tool calling because it, it can read the tools, it knows the description, it can handle them properly. It won't error too much. A lot of folks are having issues with Gemini CLI at the moment because Gemini is messing up the name of the tool or the parameter order or different things that, that are just kind of being screwed up and it's just not handling them as well. And so the thing that you want to do is is try and find ways to get Claude to be able to use the intelligence of other models. And, and the only other tool I've seen right now that's kind of doing that is is AMP, which has an Oracle tool they just released.
Yes, I was going to ask you about that because there's that planning feature. And have you used it? And then are there differences and what do you feel?
Yeah, yeah, yeah, yeah. So, so, so AMP's planning tool basically, I have to see how they, what the kind of context they give it, but I think they just kind of have, you know, Claude prep a prompt and kind of go ahead and invoke it. What's nice with the setup here is that because there's this, selected files feature, Claude can, and you saw it do it early on, as it's, starting to manage selection and picking which files to update, you know, it can build a prompt with the context. So it can assemble, okay, I'm going to add some of these files. So, you know, as a user, I can go ahead and add some of these things here and pick these files. And you see there's a prompt here. And what's great is that then when it does request plan, it uses all the context that's shared plus the context of the instructions to then trigger, you know, a plan from O3 to get some, you know, a bird's eye view to kind of get it unstuck when it's dealing with tricky issues. But when you have a chat situation here, you can see basically Claude is having a multi-turn conversation and it's using this chat GPT to kind of select the feature, select the files that are relevant, and then kind of going through and and and making edits to the to the code. And, you know, that's just a new kind of thing and I don't think anyone else is doing right now where you have a chat that's optimized for O3 where the context is kind of in a well-assembled block and then you have, you know, Claude that's running the agent loop and it's able to kind of work with both in their preferred environment right at the start. I was, you know, Claude, Claude needs to find its own context, it needs to discover things for itself. But O3 needs to be spoon-fed a little bit more to do its best work. So finding the best ways to mix and match these models and these tools, that's the secret sauce right now. That's the thing.
You do, dude. I, I'm still blown away just on this single concept. First, just MCPs, right? And watching one model talk as Claude Desktop. And so the advantage that maybe you could get using this RepoPrompt setup that you maybe couldn't using something like AMP is that if you're paying Claude a subscription for Claude Max, right? $100 a month or $200 a month, you know, calling an MCP tool call to go across, you're going to pay the API cost for OpenAI, you know, O3 or whatever. But you're still having the intelligence of one model, which is clutch, right? And so it's just doing this loop back saying, okay, I'm going to send you back everything that changed. So you're not eating the context as far as all these tokens for consuming, you know, the entire code base. That's actually done on the O3 side, right? It's the main model, which is Claude Sonnet has, or Claude Opus, whatever the Claude things you do, has just given the tiny instructions to RepoPrompt as an MCP and saying, hey, I want you to use the O3 model and see if you can accomplish this task. And so from what it looked in your chat window, it was kind of building that conversation separately and not consuming that back into the main. So you're not eating up your limits in some ways too.
Yeah, really interesting approach. Wow. Yeah. Yeah. The thing that is most expensive for tokens is all of these tool calls, right? That, that is extremely expensive because every tool call, you know, if you're doing this over API, you're sending a new request. And, and you have caching which reduces the cost of that request, you know, by 75%, but it's still 25%. So each tool call for all the previous tokens, you're paying again for those tokens. And so if you don't have a subscription, Claude, where that's kind of just free, you're paying a lot. It could cost you in some cases up to $50 a day, depending on how you're using it. That's why everyone's kind of going and trying to use these Claude subscriptions because it's really the only way to get a lot of use out of Claude, because it's just such an expensive way to work. It burns tokens.
I think the future is having really good small agentic driver models that are able to do this tool calling to kind of figure things out. And then just having smarter models kind of do engineering work that is then looped back in. And, I really think this, this flow of having O3 do some work that it gets validated and there's this back and forth between different models. I think that's just a way to kind of add kind of gasoline to this, this fire. I, I, I think, you know, going forward though, you still, we have to be careful with all of this because, you know, this tool, it can, it can go and it'll do a lot of work. And sometimes it does great work, but, you know, often, you know, it does things that are a little sidetracked. It'll keep going for longer than maybe it should. It'll try and add things that maybe aren't necessary. So you do want to stay, you know, in this driver's seat in review. And one of the things I'm working on actually is a way for, every time Claude makes a file edit through this tool, that I can go ahead and review the diff. So, if I open this up here, I can see this is the diff that Claude did for this file. And I want to be able to, review those changes every time. You know, just as it goes through. You know, so there's a lot of work into figuring out the right UX for, staying in control of the whole process. And I think, interactive code reviews as Claude is tool calling, that's going to be, a really nice unlock for that.
Okay. Yeah, and for those folks who are just joining into the stream or maybe kind of catching up a little bit, maybe just give us a quick summary of kind of what RepoPrompt is. And somebody was just asking, what's the difference from this versus GitHub Copilot agents or something like that?
Yeah. So, I mean, RepoPrompt, you know, first and foremost, it's a context engineering tool and it's made to help you, it's made to help you pick and understand what about your prompt and your context is important for a given task. And, you know, when you're selecting files, you know, a lot of folks will use things like RepoMix or other tools where they basically just zip up a whole GitHub into a string. And that's nice if you have super huge context windows, which people Gemini for. But in reality and practice, giving spurious info to a model is just not going to help you solve problems. It's going to give it a ton of extra garbage to deal with. And there's a term recently that's been coming around called context rot, where, you know, there's a mistake in the context. There's something wrong in the context and the model, because of the way that they have these attention mechanisms, they kind of glob onto those errors. And that's why if you're chatting with a model and you're, hey, can you fix this problem? And you're, oh no, that wasn't the fix. This is the fix or do something else. It'll kind of loop over itself trying to go in the same direction. And that's why people are, no, no, actually just roll back, try again and start again because you want to chop off the rot. You want to kill out the bad answer, redirect it from a higher point and have it go in a different track. So being able to control the context window, all that stuff, that's a huge, a huge thing that I built this app for.
I see. But, you know, as the realities of people wanting to use, you know, tools like Cloud Code and, you know, have this kind of full auto driving for coding, I had to kind of, you know, figure out the best ways to kind of give this tool set to a model, Claude. And that's what I've done with the MCP setup where, you know, Claude can can use all the tools that you can read through your files, you know, apply edits with some of my, you know, very advanced editing tools and just handle it it would.
So what's the difference with Copilot agents? Actually, the funny thing is with this MCP tool, I can actually plug this into Copilot, and it's actually faster to use the RepoPrompt tools to go ahead and edit files than it is to kind of use the built-in Copilot tools. You know, the read file tool is able to read much more at a time. It's able to kind of – the search tool is better versus the codebase embedding stuff that Copilot does. So, you just turn off – you can turn off their tools and just plug in the RepoPrompt tools. And you can keep the ones you want, the linting stuff with the language server, perfect. Keep those. And then all of a sudden, you've got the RepoPrompt agent in VS Code, which is really nice. And yeah, I think that's what's really cool with this. It's you're getting a sense of, I'm working on an agent too. And this is a preview of that, what it is to work with the RepoPrompt agent in all these different tools using your existing subscriptions.
Whoa, whoa, whoa, whoa. Okay, lots of alpha that's going on here. So take me through a little bit of that GitHub flow because you kind of dropped some interesting hints there. Somebody could be paying, I think it's probably $20 a month or something to have access to this GitHub workflow or GitHub agents or whatever it is. So if they plug in RepoPrompt to VS Code and have this enabled, they're just using their credits there, is that kind of what it is?
Yeah, that's right. Okay. So the RepoPrompt tools, it gives it the ability to read files, search your codebase, your multi-repo setup. It could do apply edits with search replace, which is what Cloud Code uses. And it can even do multi-search replace where it's doing a bunch of calls in a single tool. So you're able to get more mileage out of these tool calls because I know that Copilot, if you're using Sonnet, for instance, you have a limited number of credits for that. If you're using 4.1, it's different, but 4.1 isn't quite the tool calling model that Sonnet is. So yeah, you're able to kind of use the tools more efficiently, more rapidly get to your end state. And yeah.
Okay. Let's kind of break down a little bit of this context engineering concept that has been thrown around. I think Andrej Karpathy is the one that really is just putting this into the mainstream. And I feel if you've been working with AI and doing AI codings, you've already kind of have figured some of this out. And this word is kind of the thing that everyone's, I call them the McKinsey analysts are kind of starting to unite around, you know, as far as language to toss to their, you know, people who don't do this type of stuff. And so explain a little bit more of, you know, what this means and what does it really mean to you? You know, because I feel you've been working so deeply into this space and, you know, in terms of rearranging things and we've all been kind of doing our own little workflows. To me, it feels, you know, I stick my little antenna out this way and kind of pray to the AI gods and make sure that I, you know, I fall into the right, you know, context before I get my output.
Yeah, yeah, yeah. So I think ordering and all that stuff, you know, it's still very important, especially for reasoning models. So the big thing that I've learned, you know, working with models is that you want to kind of build your prompts in a way that kind of separates instructions from context. I think with agent tooling, though, you kind of have to let go of a lot of things. You just have to vibe a little bit more because the agent just needs to do what it's got to do. You got to think through your prompt, okay, well, here's what I think is important. You want to give hints to where to find things, but you don't want to give it too much info because when the agent is working in a message, it's going to fill up the context window as it goes. You don't want to pre-stuff it too much. You want to really stay minimal and lean at the front, you know, just to kind of give it that room to kind of fill up its context with important info.
If you're working with other tools, though, if you're on ChatGPT and you want to use O3 Pro or, you know, a tool that, you know, you don't have that ability to kind of have it discover context in the same way. So you have to kind of prepare your things a little bit better. So that's why, selecting your files, you know, that's a key thing in terms of ordering. I actually have a tool in RepoPrompt where there's a prompt order menu and you can reorder things. Some models prefer instructions at the top, some at the bottom. Personally, I find it works really well with instructions at the bottom and you want to separate it from file content. So you want to have key XML tags that will kind of say this is the files that you're reading. Here's, you know, some code maps and file tree. Here's my instructions at the bottom and you can duplicate them. So sometimes duplicating instructions actually helps a lot. So it's an option as well. So just I, you know, to a fault, I think with RepoPrompt, I really want to give people the tools they can use to control how they want to try and do things. People are always, oh, I don't your defaults. I want to do this. You know, there's no shortage of power users out there and they always want to, you know, be able to change every little thing. So you've got to give them those tools. But sometimes it becomes a little overwhelming for other people to kind of deal with that. So, yeah.
Yeah, let's talk a little bit more about some of these language models. I know that a lot of people, to quote the million token context window from Google. And then I was at a dev day last year in October for OpenAI. And Sam Altman on stage asked, okay, imagine the context windows go to 20 million to 100 million and pretty much infinite, basically. You know, what type of apps would you build then? And so while that's very dreamy right now, talk a little bit more about the actual effectiveness of some of these windows and your experience using them.
Yeah. So right now, if you're using most of the off-the-shelf tools, you've got O3, which has about 100K tokens. They say 200, but actually I discovered recently that when the model reasons, that's part of the token window. So if you want to maximize this allocation, you need to reserve 100K tokens for it. So your whole prompt and output needs to fit in 200K, which is which is quite limited. So that's why you got to really prep your context, let it cook, and then get a response. You know, if you're using Claude, you know, they have a full 200K window available on top of the the rest of it. You know, that's great, but Claude's a model that needs a little bit up front, needs to discover its context, and then kind of it rolls for a while and then it fills up and then it does, you know, if you're using Cloud Code, you'll see compaction or things that, there's all kinds of ways to trim the chat history and do stuff that to kind of optimize that. And, you know, Cloud 4, it really needs, it doesn't pay attention to a lot of things as well to that are older in the chat. That's why I've seen, I don't know, you've probably seen all of you listening who've used this, you'll see Claude will read a file and then it's okay, well, actually, let me, I already read that section, but let me reread that little part again just to make sure that I got it right, you know, and it's rereading the same stuff over and over. And so it's not making efficient use of its context window because of just the nature of it.
So but effective context window, you were saying as well, and that that's another thing to consider with this is that these models are flawed. They're not whatever it says on the tin in terms of what you have as context window. You can't just fill up the context window and expect great results. You have to be cautious of every bit of information you're giving it because if you cross certain thresholds, and in many cases, that's even around 32K tokens, the model's response quality just degrades. Gemini is the only one that seems to be able to break that kind of, you know, pull, the gravitational pull and is able to kind of exceed and go beyond the 100K, 200K, about 500K barrier and use it effectively. But even then, you know, it's not perfect and, you know, that model has other flaws in terms of be able to call tools. So, so it's, it's all, it's all a challenging thing. But I think going forward for coders, if you're working with these models, the thing that is important to look through with large language, large context windows is is the ability for, you know, longer running tasks. So, you know, as you let, you know, Claude work or Gemini CLI work, you'll see, you know, it's going to read things and then reread them and then look at diffs and then it's going to keep going. And, and, if you rely on summarization, you lose important information. But having longer context windows, you're able to really kind of let it go longer without having to degrade that data. But even then, you know, as I was mentioning, context rot is a thing. So if there's errors or bad things that happen, you know, you still have to do some cleaning. You know, that's actually in the repo chat, it's not agentic right now, but I, I do a lot of work to make sure that the chat history is clean and efficient. For instance, only the latest version of files is sent to the model. So there's not a long running history of changes that is going to be seen. It'll see what the model wanted to do between file changes, but it's not going to keep that whole history every time. So it keeps each message kind of tight and focused, which lets it kind of work, you know, a lot longer in the same thread.
Wow. Super amazing insights and different. I feel you're so in the weeds of this and this is super helpful, especially for people who are prompting. And that also means, I guess, if you're using the Cloud Desktop, if you're using Cloud Code, you know, slash clear would create a new chat window basically or opening a new chat window, that type of thing would basically kind of reset that so you can do the type of thing. And so can you explain a little bit more about what this behavior of context rotting is and maybe what, what, why some of the language models prefer that in terms of how, how it goes about, you know, these weights and stuff, maybe at a higher level, right? We don't have to get super technical, but, you know, why does, why does it keep anchoring on something that, and we don't even know what that looks like in your experience?
Yeah, I mean, I think it comes down to a key trait, a key engineering trait of transformers, which is called attention. So if you're using a language model, attention is that mechanism that allows the model to focus on bits of context for its response. And, and it's not a perfect system. And, you know, it's why you'll see as well, if you give all of the information in the world, these models are trained on all of the information in the world, but they're not able to kind of solve cancer and do things. They're not able to bring their attention to every single piece of information that exists at the same time when forming an answer. So when you have a context window, if you have bad data in it, it's going to get latched on it because the context window is, it's really short-term working memory. So context rot is that problem where you'll have bad data that's in there, a failed tool call, or, or just a mistaken reasoning. And, you know, the way that it's set up, it will, it will kind of get stuck on that idea. It's part of its thinking. It's, it's almost a subliminal message for it at this point where it's, it's, it can't escape the idea that that that's there. And that's so it has to consider it.
So when you do clear and you do certain things that, you're, you're, you're removing the data and you're starting from a clean slate and you're giving it the context and the problem to solve and you're letting it run. And that's why you'll see even some of the Anthropic researchers, they say treat it like a slot machine, which I really don't love because I don't, coding as a slot machine is an idea to work, but you just try your prompt, let it run, and you see what happens and you can take the results or drop it and try again.
Yeah, I mean, that's a way to work. It is valid. Even OpenAI researchers will say, you run O3 10 times and look at the variance in the output. There's going to be a lot of variance. And then, and if you take O3 Pro, what a lot of folks think it is, is this just a bunch of O3s. I think they're sequential though, where they'll kind of run through the problem and then run through it again and then run through it again. And then you, that's why it takes so long, but then you get a much more reasoned response at the end. So that's actually the Anthropic theory, I think, is that a lot of the researchers say the more tokens you generate, the closer you're going to get to some type of truth to verify. And so that's kind of why that's that's actually why a lot of their everything about them is just about the tokens, just produce the tokens.
Yeah, I mean, I, I have mixed feelings about that personally. I think, you know, if you're in a world where, you know, you work at a big lab, Anthropic, and your only way of working is to burn tokens and you're, yeah, of course burning tons of tokens is great. But for the real world and people living in it, those tokens aren't free. And there are better ways to work than to just burn tokens and hope you get to a good answer. You can plan better. You can think through, what am I giving the model? And I was talking recently to one of the engineers at Sourcegraph about this. Because their philosophy is always more tokens is better. And I'm, well, there is a balance because efficiency does matter. If you're, and efficiency, and also asking yourself a question, if I'm burning tokens, is it because I'm lazy and I want the model to do more work for me? Or is it because I am really getting better results? And sometimes, you know, burning more tokens does lead to better results. But often it's just because people are lazy and they don't want to do the work of planning what they want to put into their.
Prompts and planning their tasks properly and thinking through, you know, as an engineer, what it is that I'm trying to solve. Thinking about your problem for a second and just doing that work ahead of time, it makes everything 10 times more efficient for you and for the model. So, yeah, that's right. You gotta ask that question. I this way of thinking because a lot of times before I'm even getting into cloud code, now I'm using O3 in ChatGPT to kind of help me find libraries, do some initial research, so then I can synthesize it before I actually start to vocalize what I really want. And a lot of the times I don't really know what I want. So I'm going to another model and in essence, and asking these types of questions. And my favorite part of an O3 type of thing is this web research type of thing that it can do simultaneously and kind of speed up that workflow. And so I think that's an interesting approach that no one really talks about in the mainstream. It's, if you want an idea before maybe you put it into cursor, before you put it into any of these models, it's use another AI system to help you get some clarity, to better kind of focus your, you know, outcome or whatever thing that you want to do. And so I feel that, yeah, yeah, because there's there's multiple ways to approach the problem, obviously. And I think experimenting with them is going to be probably the biggest bet right there. What are your thoughts right now? I know Grok is kind of on the cusp. Tonight at 8 o'clock p.m. Pacific Standard Time, you're going to be 1 a.m., right? It's, what, what do you think is going to be the key here? Because when Grok 3 came out, I was really blown away with how you can, you know, ask these different questions and stuff. And and and and I guess kind of with that, have you noticed any any difference in using extended thinking with Claude's on it or any of these thinking models? Yeah, yeah. So I think the thing with extended thinking and just to go to to Grok also as well, is that I find personally extended thinking works really well if there's something to think about. Which is why I O3, when you prep your context, you give it a task, and it's able to think through the problem that you give it. That's the best way to use them. If you're if you're using a model, Sonnet, I actually almost never turn on extended thinking, just because I find that one, it kind of goes in loops a little bit. And two, it's thinking about stuff before it sometimes even has the relevant context. So just letting it kind of figure out its context and then solve the problem or then delegate the thinking to something else. That's also a good way to go. I know Anthropic recently released interrelief thinking. And I think Claude Code kind of does this also, where it's able to kind of think between tasks after it's gathered context. That can be useful. But, you know, generally, I don't find that it adds sufficiently to the results because you do want to just kind of let it work with what it has, find the context, and then kind of organize things. And sometimes just stopping it and then moving along and, you know, finding more ways to kind of using, I think this is going to be a thing that's going to be bigger and bigger as time goes on. It's human in the loop tool calls. So it's making a tool call, having the human kind of look at the tool call, put some thoughts into it. And then in the response, it'll see what you said. And then it can keep going and kind of redirecting it. So that kind of workflow of, we're steering the model as it runs. That's a good way, I think that we'll kind of progress. You heard it here first before Andrej Karpathy talks about it. Yeah.
So before I get here, I just wanted to say some folks in the chat, I was seeing they're asking for discount codes. And I just want to say, anyone in the stream right now, if by the end of the stream, I'm going to put up a code for anyone here. I think I'll name it RayFernando15, and anyone will have 15% off, and you can pick whatever sub. It'll only be good for the day. But yeah, everyone in the chat, RayFernando15, if you stay to the end, you'll be able to get that discount. Let's go. Let's go. A lot of people are asking for the code, so definitely, especially my members here, they're, we want to get cooking with this. And thank you so much for providing that. I think that's just super duper fire, and the chat's going crazy right now. That's great. That's awesome. Yeah.
And now Grok. So Grok is coming out today, probably. We'll see. Probably. So the thing, and it remains to be seen, well, Grok, what kind of a model is it? Is it going to be a deep thinking model? Is it going to be the kind of model that you prompt O3 or O1 Pro? And is it going to be more of a tool calling model, Claude? And I don't think it's going to be a tool calling model. That's just kind of the way I think about it. So it'll be the kind of model that you really want to prep your prompts for and think deep for. And I think that's fine. I think Claude is really specialized in the tool calling workflow and people really that. But I really encourage people to try these other models and find ways just going to Grok or chat.com or even AI Studio and just prompting the model in the chat flow, you're able to get a lot done with these models even outside of a tool calling environment. And I know everyone's so used to just letting the AI do everything for them, but that's very recent and it's often still not the best way. And sometimes it's not even faster because the model's going to go off for a long time. It's going to make a bunch of stuff that you have to review. And then some of it's good, some of it's not good. But I often find this, and a lot of folks will say this as well, is when you're solving a problem, and they'll say, okay, I'm going to give it to Claude, even Opus or Sonnet in Claude Code. It'll say, fix my problem. Or then you ask the same problem to O3. And O3 will be, okay, actually, this is one line of code. You just change that, your problem is solved, and it's right. And then Claude will be, okay, I engineered 15 classes for you. Here's all the files. It solves your problem perfectly. And you're, okay, well, you know, great. But that's all this code I have to do, though. You're so right. Absolutely. Wait, there's got to be a better way. Absolutely. I can do that better. Yeah. It spawns eight subtasks and just runs off for you for 10 hours. Yesterday, I was live doing that with Claude Code. And literally, I'm just, I have a value that's displayed from my back end as far as how many minutes the user has left in their actual plan. And so, you know, in the UI, there's four or five different places where that shows up. And Claude Code's, I can apply the math.round function to all these different places in your code. I'm this is super sus, bro. What is you please rethink this whole thing? It's, oh, you're right. You are so right. I was, bro, you're just gassing me right now. I feel I feel I'm the model and now it's stuck on that. And so I think there's a shortcut in Claude Code you can do where it's escape, escape, and then it lets you roll back and you can pick up the chat from right before. Bro, what a nice way to kind of trim some of the, yeah. Oh, oh, so if you hit the up arrow too, is that what it does? It trims it if you hit up and go? Yeah, you can move up to kind of go there. But escape, escape lets you kind of start from a a different point and kind of fork the chat from there, which which is super powerful. Bruh, how did I not know this? Well, what's nice with the Claude UI too, is you can just go ahead and if you're on Claude desktop, you just go through the chat and you just hit edit and Yeah. Let it stop. Yeah. Yeah. And just just start from there. Oh, my God. Damn. Yeah. My mind is blown with escape, escape.
So context, right? We heard about also human in the loop tool calls. Right. I feel no one's talking about that. You're you're hearing it here first on the live stream. You're going to be telling your friends about this. Yo, the future is all about human in the loop tool calls. We said it here first before the guys saying it. Well, basically what we're acknowledging is that at this present time right now, the language models still have a lot of work to do to figure these things out. And maybe nerding out just a little bit more because I love talking about language models and things. So in this extended thinking or reasoning type of stuff, is it really just a shortcut of pre-trained thoughts or synthetic data? So, if something has never been reasoned over and never been trained on, then, going forward, the likelihood of it, kind of coming up with a new novel reasoning is actually really low? Is that kind of what I understand? Yeah. Unless it's something. I'm yeah, I mean, the the key thing and and this is what is steps up O3 from O1 for instance, is that they gave it way more problems to think about and solve while it's training. So you train the large model, and then you go through post-training, and then you have a multi-turn chatbot, and then you want to add reasoning on it, and you basically put it in the gym, and you're okay, here, here's a bunch of different problems to think about. And then And you basically, you let it work on the problem. And when it gets it right, you keep that data. And when it gets it wrong, you throw it out. And so then it's a way to kind of discover a bunch of extra data to train the model on these extra problems. And so if you have a problem where, you know, the model was able to work through it in the reasoning, it's going to work really well in the field. But, again, you know, it's TBD if, you know, to what degree fully, truly novel problems are going to be doable by these models because it's still, you know, it's still so early with all of this. And I do think, if there's insufficient data, they're going to suck. I've tried O3 on some, really low-level problems, and they just hallucinate everything, and that's just not helpful. And then sometimes the RL and the reasoning, kind of leads it to hallucinate more. It makes it think that it solved things that it didn't. And when you're multi-turn too, a lot of the time, these reasoning tokens are lost. So as you ask the model to do something again, it has no idea what it already thought about before. So, you know, there's this whole thing that, you know, it's kind of an inefficient hack at the moment to kind of work through these tricky problems. But it does, you know, solve things really well in certain cases. O3 is an incredibly good engineer model. And that's why it's able to come up with really succinct solutions because it's able to really think about them for a while, and then it kind of throws out the old one and it backtracks and it thinks through a better way. And then you end up with something really small, and you're, wow, this one line of code is all it took to solve this problem, versus Claude just kind of going off the rails and just moving forward, a race car, just kind of go. Yeah.
So what is your go-to stack? Eric wakes up in the morning, you're writing repo prompt code. What, take me through, what, you know, what is your day to day? What AI windows do you have open? All that stuff for folks. Yeah. Yeah. So, I mean, today is interesting. I was actually starting off because I just signed up for Sentry. So I don't know if you all, I mean, I signed up for this new thing. It's called Seer in Sentry. I've had Sentry for a bit. But so, so what's cool with that is that, so say people are using my app and they, they, there's a stall UI stall or a bug that appeared that gets sent to Sentry. It gets some issues, some info about the code. And then there's this AI model, Seer, that will look for root causes for that issue. And so now lately, what I've been doing with that, I started this today, is I start with that. And then I basically be, okay, what did Seer find? It found there's all these issues. So then I go up in RepoPrompt, I put that in the app, and I run the context builder, which automatically selects the files that are relevant to the task. And then I just feed that to O3 and I let it cook. And so I've been just kind of assembly lining problems to O3 Pro to kind of think about and solve these problems, evaluate the problem. Here's the context. Here's the issue. And then go from there. So I think just starting a problem in O3, if you have the sufficient context, as always, it's the preferred way to go. And I do having ChatGPT for that because you can just open up a bunch of threads, just just set up your prompts, kind of put them out and let them cook for a while. And you can just assembly line them. And then once you're, you've got that, then you can kind of turn to Claude and either Claude desktop or Claude code and then kind of have it look through the solution, evaluate, is it exhaustive? Is it sufficiently exhaustive? Does it affect every situation, every, every case here? And then sometimes you can kind of work with that. But you want to try and find a combination of these tools. I do think a Claude Pro subscription at $20 a month right now is a really good value, especially now. I'm finding with the RepoPrompt tools, you're able to do a lot with it. In Claude Code, it won't last that long because Claude Code burns tokens you've never seen. But with the RepoPrompt tools in Claude Desktop, I'm able to get a lot of usage out of it. I'm on $100 a month right now, but $20 a month goes a long way. And then now, O3 over the API, it's really good value. You know, we're spending all these money on subscriptions, but it's getting to the point where API calls, if you're doing them right, it doesn't have to cost that much. People are used to seeing, you know, if they look at cursor usage based billing, O3 is expensive in there because it's making so many tool calls and doing so much stuff. But you don't have to do that. You can just you can use or you can just go on your own client. There's plenty Libre chat or whatever. You just put in an API key and you can just use O3. And it's really not that expensive. $20 a month of API uses on O3 goes a long way. You can do a lot with that since the price dropped. So you don't even need a subscription if you don't want to. And honestly, I would recommend people try and resist the urge to kind of have an agent do everything because you will understand more what is happening if you don't. You're able to move more deliberately. And if you're reviewing every response, you're just kind of more in control and you're able to kind of vet more of what is going on in your codebase. I find that's really important.
That's solid advice because I remember, I hadn't touched my codebase the other day, a UI part for a while, and, you know, I was doing that stuff on vacation. And so when I got back, you know, I have my founder's hat on, right? I'm doing media, I'm, taking care of family, I'm doing all these different things. And I got back. I was, oh, my God, where? And so, one of my latest hygiene tips has just been writing documentation really for myself. So after I do a bunch of chats and before I clear it out, it's, can you just write some documentation on this entire chat and summarize what we did? And so that way, I can kind of review it back and it helps me sort of get up to speed later on in the codebase because sometimes I'll forget or whatever. And it's been kind of embarrassing. The other day, I forgot to that you can actually just add one line to the Stripe checkout and it literally will produce the coupon. And, you know, so I had the the tokens flowing, burning, you know, a bunch of tokens through Claude Max. It's, yeah, it's just this one line. It's thanks. I could have probably figured that out. But yeah, you know what I mean? So that that was interesting. Yeah. Or just with the UI stuff, I forget, you know, with React. Okay, that's right. I did set up a hook for this that way I don't have to update these four places in the UI. It's all grabbing it from this one hook, you know, which is that and that value is being pre-filled from there. And so it's, yeah, stuff that. But now that I know it, it's going to be sticking in my head because I I learned about it. And I I think that happens. If people have been coding their app and not really understanding what the agent is doing, it is actually good to go review the code and play around and see what the inputs and outputs are. And it's okay. I feel I used to be very embarrassed, especially, you know, being an engineer for many years now, letting the agents code for me and then now not really understanding what they're doing. And I think what I'm realizing, Eric, is that with a new programming language, I learn better when there's scaffolding up. So if I get an existing template or GitHub repo, I'm going to learn the language a lot faster if I start poking at it, start sending it inputs, start changing codes, start adding features. And then sure, it could be ugly at the beginning, but all the gaps that filled in with my knowledge and understanding, you know, what's actually moving or debugging the app actually gives me more clarity than if I were to just study it. And that's just my way of learning, but that's how I've learned growing up. And I think that's how most people are learning these days. And I think a lot of people in my streams are learning that way. And yeah, so repo prompt as far as, is it subscription based, kind of give some people a little bit of background of maybe what they should be expecting when they're going to go check out your app here.
Yeah, I think what you just said is great. Taking, taking the time to kind of read through what's been done, to document your learnings, keep, keep in, keep in mind things that are, that are changing around in your code when the agent's making changes. It's really good. And learning by doing scaffolding is, is great to get started. You know, working on, on, on products, projects, having, having a goal in mind of what you want to accomplish and being able to have that dopamine hit of a change kind of being applied right away. Super helpful to learn. I find learning, it works best if you try to turn it into a game and you try to get those dopamine hits as often as possible. Because if you're not enjoying it, you're not enjoying the struggle, it makes it harder to motivate yourself to struggle and to try. And that's the biggest thing right now with people learning to code at all. I feel it's a little tricky for juniors who are starting out their career right now. Because they just, it's so easy not to struggle. It's so easy to just default to asking, you know, an AI model to solve problems for them. And, and, you know, when I started my career, the thing, one, one bit of feedback I got often, early on with one of my first jobs was, I was struggling in a codebase and I asked a lot of questions and I was asking kind of too many questions to the senior guy above me. And he was, dude, you gotta, you gotta struggle a little bit more to figure it out first. And, you know, that stuck with me. And, and I really kept on it. And, you know, so I would try and really struggle before I had to go and look for help. And I think that's a practice. If you're really starting out and you want to learn, try and struggle before you turn to the automated tool, to the, to the easy solution, because that's how you're going to understand what's happening. And, you know, when you're doing that, you're just going to be able to write better prompts, write better, you know, understand your problems better. You're going to be able to, figure out debugging more easily because you're able to understand how the code flows and where the ins and outs go. And you'll be able to use that context in your prompts and you'll get, you know, just overall much better products.
That's a really good point. I feel that's not said enough. And it's just what you'll actually start to discover is that there's a lot of content already on YouTube that kind of describes certain debugging tasks. So as you enter this journey, instead of relying on the agent for debugging, you're, how do I look this up in the shell command? If you just type that into YouTube, YouTube is great because it's going to start recommending videos after that. After that first video watch on debugging, it's going to be here's how to use the shell and how to pass these commands to understand what the shell is doing. It's what you can pass in dash H into any command, it's going to give you this automated help menu. I had no clue. It's and then you start to actually get into whatever stage that you're at in your coding career, you're going to learn so much more just from doing that type of thing. It's that that that's the biggest unlock. And and yeah, don't don't be afraid, I guess for anyone who's coding and if you don't know the answer, it's okay to use the AI model, but maybe also try not using it too and just doing a plain old Google search and seeing what comes up, right? So it's kind of the, maybe we're going to call this the raw movement. Can you just go raw dog it and just use Google? Well, that's great, but the problem now is Google's results are super polluted at this point. So it's kind of tricky, you know, we're getting into a tough world now, so even that's kind of out of the window. But just trying to solve things a little bit more raw before. Run the program. Read through what's happening. Either attach a debugger, put some logs. Write some prints yourself. Printing is easy to do. You just type print or log or console log or whatever language you're using and you'll be able to see, okay, that's where that is and you'll get a trace. And you know what? The funny thing with that is that language models love logs. They are They'll eat that shit up. So you give them lots of logs with your problem, they will understand exactly what's happening. And so will you, which is really helpful. That's hilarious.
So yeah, for those folks who don't know, Eric is the founder of RepoPrompt. And we've basically had this beautiful discussion just now, kind of talking, learning out a little about language models, about RepoPrompt. It's kind of what it's doing. It's a high level. And what I'm planning to do on my future live streams is going to be using the app and kind of walking you through some of this workflow because I am super impressed of this new MCP feature and watching it perform we did earlier in the show. And if you're going to be watching this, I'm going to have timestamps so that you can kind of follow along backwards because this show is also going to be sent out on YouTube. And it's so amazing just watching Claude Code and just talk to RepoPrompt and say, hey, go do the thing. And it's, O3 is smart. I'm going to let O3 figure out the things. And then it's it's not using any of those tokens. RepoPrompt is then using its own agentic sauce inside of RepoPrompt to then grab the code that it needs to give O3 do its thing. And then it sends the results back up. And Claude is basically saying, okay, cool. Task is done. And it just keeps going. It keeps going. It keeps going. Amazing. Amazing, man. Wow.
I want to just chime in, by the way. I just made that code live. RayFernando15. So if you're watching the stream today through tomorrow, you can go ahead and use that if you want 15% off any of the subs. So you can get it off the monthly and it'll work for a full year. Or any of the other tiers. So yeah, I... And if you're using this and you do want to give it a try, please do reach out to me either on Twitter or I'm on Discord a lot. I really pride myself in customer support. I know some of these things are a little daunting. The tool can be a little complex for new users and something I'm working on. But, you know, it's very powerful. Someone's asking, you know, can you just use, you know, repo prompt with only a cloud desktop? Yeah, if you have the subscription with repo prompt, you can use the MCP server. But without it, too, you don't need the MCP server. You can just build your prompts. And what's cool with the XML copy feature is you can actually just have it, output structured XML and it will apply the edits automatically. So on the free tier of repo prompt, you can go ahead and use Claude Desktop on $20 a month and get some work done. It can create files, edit files, everything you need. You're just kind of in there. There are some limits on prompt sizes, but you know, you can get a lot done with it. That's the clipboard XML flow. So if you're on, if you're trying to use, just minimize your spending, the free tier is great. If you want to go a little bit further and automate more, there's the Pro License, which gets you the MCP and a bunch of other things. You know, all the power tools are there. But yeah, you know, you don't need to spend any money to kind of get started. And AI Studio is another way to kind of get a lot of use out of that. AI Studio is great and it's free. So yeah, so thanks for signing up. I see some folks already did. So that's really great. And, but yeah, asking questions. If you're running into issues, you're running into anything, please tweet at me, DM me, it's all good. I'm gonna answer you. I have no problems with that. I'm still small enough that I can talk to people directly. Yeah, that's what's really nice. It's just literally organic engineering right here.
And speaking of organic, we have our own Alex Volkov. If you don't know, now you know. Thursday, I tune in tomorrow. This boy drops knowledge every single Thursday on Thursday AIs. They do live streams now. They are also on X Spaces. They have all of the top talent in AI. Whenever something's going on, they'll be on the show. So with Grok dropping, he's going to have the people who made the model. They get into technical details. And I would highly suggest, especially if you're into any of this AI coding stuff, definitely check out Alex Volkov. And then you can see Thursday AIs is the name of the show. They're also on YouTube. Make sure you subscribe to them. Super high quality content. I love what they do. That's actually where I catch up on my AI news every week. There's also a newsletter you can subscribe to as well to make sure you kind of check that out. To me, I was, oh, there's a new, they talk about open source models. They they just all the AI technical juices and some some sometimes some of those weeks are pretty packed, but you can always get the TLDR. And even if you're super busy, you kind of have to keep a pulse on this. And that's kind of where I actually keep my sanity. It's sometimes some weeks are so crazy with information, it's okay, let me just see what the TLDR was. What did I miss? You know, was there a new vision model that I missed? Was there something really interesting? Definitely super cracked. So Alex Volkov, Thursday AIs, I I gotta have I gotta actually have him on the show. It would be really fun to have him on. Yeah, because there's so much sauce. Yeah, yeah. And I was on his show when O3 Pro dropped, and that was that was a lot of fun. But yeah, it's really great to see him kind of hopping around here. The X kind of AI community is not even that big, you know, and it's cool to see, you know, so much information sharing. Yeah, it's really good.
That's awesome. Eric, thank you so much for tuning in. And also thank you for basically letting everyone use the Ray Fernando 15 coupon code. If you want to sign up for a repo prompt, make sure you just drop in that little coupon code and you get a 15% discount. Ends today. So if you're watching the stream on the replay, make sure you sign up ASAP to take advantage of that. But by the way, you know, Eric is very approachable, which is really awesome. It's still a small organic thing. And if you have any questions, definitely reach out, ask any questions. There's a Discord as well in his community. And if you just go to repoprompt.com, you can join the Discord and have a conversation with other people who are using Repo Prompt, which can be super helpful because everyone is creating tips on, you know, what the latest things happening. You can file bug reports there as well. And because this is Eric's full-time job now, he's able to do that. Previously, he used to be a staff researcher at Unity. Yeah, that's right. Been able to do this full-time now. It's really nice. I really I'm really grateful for that opportunity from the folks that are supporting the app and using it. And I just want to mention the Discord. Even if you're not using the app or you're on a different platform or whatever, there is some really smart people in there and I learn from them every day. And there's some good discussions on MCPs. There's some great workflows in there that people are coming up with. Some folks are building tools for themselves that they share that help you automate a lot of stuff. There's some really smart folks in there. And I'm a little biased because it is the RepoProp community, but I really love those folks in there and it's worth learning from them. So yeah, definitely join if you just want to hang out and see what's going on there. Definitely.
Thank you so much, man. It's a super pleasure to have you on. For those folks for tuning in, make sure you check out the timestamps. There will be all below. You can rewatch the stream, share this with some friends, let some folks know there's going to be a bunch of clippable moments in here with a bunch of cool stuff. And also thanks for the support for everyone who's been joining in who are members of ray fernando 1337 just by joining in on the YouTube channel. So definitely appreciate that. And I will see you tomorrow at 9 a.m. This is my summer of streaming. Depending on when Grok drops or whatever, I might even do a surprise live stream and play with it live. But definitely for sure tomorrow at 9 a.m. I will be live. And let Ray sleep. Let Ray sleep. All right, y'all. Thank you so much for tuning in. And we'll see you tomorrow, 9 a.m. Pacific Center Time. Peace out, y'all. Thanks, Eric. Cheers. Cheers. Thank you.