VIBE CODING AND THE EVERYTHING CLAUDE REPO

E111A Vibe Coding and the Everything Claude Repo

Transcript

Dalton Anderson (00:01.432) Welcome to Venture Step Podcast where we discuss entrepreneurship, industry trends, and the occasional book review. I'm your host, Dalton Anderson. In this episode, we're gonna be touching on two key topics. The everything Claude code repo and anti-gravity. With the goal of this episode to discuss some of the things that are going on and.

how to best utilize the tools that you have around you, either for free or things that cost money, to get the best output from your AI assisted coding. Because it's gonna be a thing, it's not going anywhere, it's only gonna increase in its efficacy when it comes to usage. Majority of code later on this year,

will be AI generated. And so how do you get to a point where you trust the output faster than your peers? Or how do you get to the point to where you're operating lean and don't have the capacity to review all the code that the AI agents create and you feel like it's in a good spot? That's what we're gonna be discussing today. The first thing that we'll like to do is we're gonna be building off of a previous episode.

In a previous episode, I discussed creating an app and the features of a product that Google has called Gemini, Jules, or it's not called Gemini, I messed up there. It's called Jules and it uses Gemini to create a cloud version of your app. It makes a virtual machine using Google Cloud servers. It takes your repo, runs it.

does an analysis of what it needs to do, your request, and then it builds it and codes whatever feature it is, tests it, sends you screenshots and a report. And that's all in the cloud, so it's a bit slower. But the new version they'll be coming out with is going to be the next step of where we're going, which is a automatically iterating agentic system, but we're not there yet.

Dalton Anderson (02:22.872) So how do we get close enough? Well, we're gonna be discussing what I think. I'm not gonna say I think our best practices because who am I? I'm gonna discuss what works for me. But first, we're going to kick off some agents to work on our behalf during this episode. And the first thing I would like to do before I do that is I wanna show you the app that we built last time.

So the app that we built last time is just a simple UI. It has a text box and the essence of it is you put in a claim, you describe what happened and then the system will ping via API, Gemini. I think I used 2.5 Flash because it's so cheap.

It's basically free when you're just sending this low amount of It's gonna be free forever, just heavily subsidized at the moment. It sends that request over, classifies it, tells it what it is, what type of claim it is, and whether it's fraudulent-ish. Without too many rules, what I am gonna be requesting

is some increased capacity on the UI front. Let's make it pretty minimalistic, flashy. And then the goal of the backend is to also allow us to process a file, an image, and then use a image model. Probably gonna use doc.

like document AI workbench from Google, just because it's in anti-gravity, it's easier. And I already have API slash Google authentication with Google. But I'm not necessarily specifying all those things. I'm just telling it, hey, I wanna make the UI better, use some flashy colors, pick a color, do your thing. And then I just said, hey, I wanna be able to process a file and then I wanna be able to detect whether or not AI

Dalton Anderson (04:44.482) had any play in this? this image created using AI? Is that true, false? I didn't specify how it's gonna do it. I just told it that's what I wanna do. So we'll see building that out during the episode, how it goes. I have no idea. I have not.

I have not done that before. So within here, I set this up to create a couple things for us. So this is, let me backtrack. So this is antigravity. Antigravity is a heavy fork of VS code. So it looks similar to VS code. If you're not familiar with VS code, VS code is pretty dope. It's an open source code editor.

that was released by Microsoft way back in the day and it's pretty much goaded and a lot of people use it and they use it for good things. They use it because it has the distribution, the plugins, it's heavily customizable, it's open source, people like it and it's good. There has been instances of companies creating

a rehashed version of VS Code like Zed. I really appreciate Zed. I like Zed quite a bit, but when it comes to productivity, you got to go with what's the most productive, not necessarily what you like the most. If I had a choice between Zed or anti-gravity or VS Code, I'd probably choose Zed. Zed's built ground up in Rust. It's way faster. It utilizes your GPU instead of your CPU. It does a lot of cool things.

on a optimization basis, but this is anti-gravity. It looks like VS Code, feels like VS Code. Really the only difference is that you have this agent manager system built in to where you can run multiple agents at the same time. I think up to 15, I've never ran that many. One of the key issues with running a lot of agents at the same time is that since the code base is not in the cloud,

Dalton Anderson (07:05.986) that they can run over each other. Whereas in the last platform that I talked about, Jules, in the previous episodes, that platform is in the cloud so it can understand the code base and does its changes within the code base. And then you could be working on three different things at the same time when one task finished, the repo or the repository, basically the code.

is updated, then the other workers, when they finish, they can see that the code has been updated, and then they can refresh their work to reflect and fit with the new repo. That's how it works if it's in the cloud, but if it's not in the cloud and everything is local, you can run into problems if you're running so many agents. And it's also a lot of copy and pasting. You get fatigued, you get copy paste fatigue. Like instead of writing code and getting cognitive,

You're just like, this is so much planning. There's so much going on here. I just need to be able to kick things off and not copy, paste, copy, paste, copy, paste, and approve everything. So that's where this comes from. But I will drop into the agent manager. So the agent manager has workspaces. So you can see that the claims triage engine workspace, this has a couple of chats. We're going open up a new chat.

You can use audio recording and it's pretty cool. It's like I've listened to your audio recording and I have understood what you said. I think it's neat, but there's a couple different modes on how you can relay your chats and these modes cost less or more tokens and it really depends on what you're trying to do. So there is a planning mode and as you guessed it, it helps plan. And then there is a fast mode which

It's more direct, doesn't need planning. It just needs to get stuff done. And there's a couple different models. So there's Gemini Pro High, and there's Gemini Pro Low, and then there's Flash, and then there's the Sonnet, Opus, and then GPT-OSS. Those are the models we have available. These models have credits, and Anti-Gravity is very generous with the credits that they give you.

Dalton Anderson (09:36.346) I have think the ultra subscription at the moment testing it out. Wanted to test a couple tools. So I subscribe to that and yes, it's expensive, but also I'm utilizing everything except my video generation credits. Maybe I'll get to that one day, but so many credits per month, it's like 25,000 credits. But those are the models. The credits are generous.

and

You specify the model, so in this instance, I'll say I want the low and I want the planning model. Because what I did earlier was I asked it to plan out some epics. And if you're not familiar with epics, it's just like big projects.

And so this epic is to create a flashy reactive UI. And this next epic is to create a AI claims classification and escalation process. And so what I've done in the past was I would kick off a whole bunch of chats and I would say, okay, I have five tasks and I would just kick them all off five, five at a time. And so I'd batch them all through.

I would look at their plan and then I would bring it back in. Whereas now what I like to do is I like to set up projects and these files, markdown files, to specify, okay, this project is Epic II, it's AI claim image classification escalation. All right. I tell it, hey, I want you to iterate through that, run through, do your thing.

Dalton Anderson (11:25.112) plan out and kick off and build, test it, do your thing, come back. And it runs through all these requests instead of me saying, okay, after this one, now do this, now do that, do this. It just knows, okay, now I've gotta go do the next thing within here and it will cross it off. That's how I have it set up. So let's kick this off. It did enough explaining, so let's say.

I've never tried this while I'm recording, so let's see if we can record and record at the same time. All right. Hello. I would like you to take a look at the file Epic to Claims classification, and I'd like you to iterate through these requests and follow some of the commands and rules that are set up within the project and put together an implementation plan that I'll approve.

slash provide commentary on, and then we'll go from there. Thank you so much.

It's funny, I say thank you, please, whatever. I did see somebody today on X build a whip for Claude, where a person would whip Claude to start working. And there was a comment saying, people are saying please and thank you to these AI models, and this man built a whip and is whipping AI. And there was jokes like saying, this is the first person when

we reach Singularity that becomes like AI Slave.

Dalton Anderson (13:05.24) Okay, so we're sending that. And so the next thing we'll do is we'll kick off the UI piece. Hey, I'd like you to see Epic One Flashy UI and review the requests, put together an implementation plan. I'll go in and approve it. Please iterate through the whole process. Once it's completed, pick up the next task and then identify which tasks are being completed and are in progress.

You don't have to specify all that stuff by just doing that because I'm recording. I would just say pick up this. would normally just type it probably. But so it's gonna go through, it's gonna look at this. And so why this is going on in the background, I'll put this over here. I wanna walk us through the everything Claude code piece. So everything Claude code is this repo.

that is by FN-M and it's a very popular repo. It's got 143,000 stars, 21,000 forks, which is a lot. So what is this? There's a lot going on and so I took this and I didn't realize how often he was updating. Like it's got a thousand commits, but this is something that somebody called out a little bit ago.

been using it for some time now. I'm not sure how long, but a while.

I'm using it because it writes better code. It basically provides guidelines and rules for the agent. I did a fork from my other code base, is

Dalton Anderson (14:53.698) which is private and then, let's see, yeah, which is private and then I added in some of these rules. But the big thing here is when you think about what happens when you request information from a model, the model does something called RAG and that is like retrieval, was it retrieval augmented generation?

Basically, it's a complicated mathematical process that has documents and information in this vector database that

by the math tells the model where to go and retrieve information so it doesn't get overwhelmed with all the information that it has at its disposal. And it's done at runtime to reduce the amount of tokens used and makes it more efficient, one, faster, and it optimizes the retrieval of information.

I'm not saying this is rag. What I am saying is it's similar to rag and I would think about it as like an iterative context and file directive. And so when you think about it, I had a glitch there. said when you, when you so many times. But it's this iterative.

file and its iterative context and file direction or routing and what I mean by that. So in this dot Gemini, there's this folder called commands and so this folder commands has some commands and different things going on. But then there's this rules, styles, hooks.

Dalton Anderson (16:56.334) pattern, security, and when it has anything related to those types of questions or those types of things going on, it will reference those files. Instead of having a massive spreadsheet or a Word document that the AI needs to read through, when it has a question about something, it's able to just retrieve the exact thing it needs. Okay, how do I build a feature?

How do I do database migration? Okay, what stuff am I supposed to be working on? All right, what are some of the of the controls that we have in place? What are our rules? Okay, this is our rule about architecture. Here is our coding commandments. Let's look through that. How should I structure a database? All of those things.

If you had it in one large file, it's a bit overwhelming when it comes to the context window because you're filling it up with the code base. You're filling it up with the information that you're requesting it to do. You're filling it up with its own thoughts because it's got to think about what you're talking about. And then on top of that, you've got this massive file that you want it to reference all the time. And so it doesn't scale very well. And what happens is it starts to hallucinate and

This context retrieval allows the code to be crisp and viable in, let's make sure I have, open up the, crisp and viable. I want to make sure I don't have any approvals that I have to do. Implementation plan review. Okay, so this put together, I'm stopping for a second. There's a lot going on here, but this is the implementation plan. You can add comments on what you'd want to change, but.

It's gonna use the Stitch MCP, which is the Stitch Model Context Protocol. Stitch is a UI generative platform that Google came out with, I think a couple weeks ago. And so I asked it to use it. It's made some cool stuff, so I figured why don't we try it out on the show. It could be a hit or it could be a miss. And normally I should be able to...

Dalton Anderson (19:19.726) put approve, but it's not letting me so I will just I'll just say looks good.

Dalton Anderson (19:35.97) All right. And then.

This one, it's got another plant. Summary doesn't allow me to prove this one either. Weird. And so it goes through what it wants to do, how it's going to name things, the process. And if you didn't like something and you're like, OK, I want to use XYZ model, you can make the suggestion and the change and it will update the implementation plan.

I don't need to make any changes because this is not going to be production. So we'll say it looks good.

Dalton Anderson (20:17.422) All right, so these are both cooking.

I'm pretty focused on full, okay. So it's gonna run through. So let's back this down. Okay, so what this does is it allows you a...

streamlined way to retrieve information and implement rules and guide rails of things to do and not to do. And what that does is it makes the model more efficient when it's retrieving information because it's not reading through a massive document and it provides a better code and code that you can trust, which is important.

but this is very, very detailed. has like the skill.md is massive.

I didn't put this together, I put together the other stuff because I thought it was important and missing, but it's more architectural.

Dalton Anderson (21:24.726) on how I wanted to set up certain projects.

This is really good stuff. And if you're getting into vibe coding or if you're getting into just coding in general, it's probably good to think about these kinds of things and make your own rules. And if you're working on a project and you utilize this repo to everything Claude code, I've changed it to everything Jim and I code because I heavily forked it and changed it up. So I made it made it my own, but

Dalton Anderson (22:00.556) The one thing that this does is it will recognize when there is a gap in its rules or it's encountered a new situation or the project has gone to a certain direction where a new rule is required. And that's important because you don't have to constantly recognize when things are awry. You can prompt it, yes, of course, you can add new rules. But it also recognizes that,

There's this new thing popping up that we saw when we were validating the build that I missed. We need to add it to the rules. And so it's added and then you'll approve and make sure that the rules don't get too large. But the rules are not, I would say the rules aren't specifying which tools to use or what to do. It's more or less a guide. doesn't tell it step by step what to do. It just says, Hey,

don't do these things. And there's some instances where it does tell it exactly what it would like to do, but it doesn't necessarily state exactly what to do when it comes to coding. Like it says, prefer camel case for JavaScript and TypeScript module file names and keep skill or command directories and give me a case.

and it specifies its preferences, but it didn't say absolutes. I think the only absolute would be like sandboxing and not accessing information or editing information that it shouldn't be doing so. But the continuous learning is such a great piece of this whole process. And then there's also workflows. So there is workflows that the agent will set up. And so it will read through all these rules and it will set up workflows.

So this one says, initially, is agents with core project context, spec, state, and commandments. So these are some important files that it deemed that it should be reading through. I didn't set this up. AI did this themselves. I said AI did this themselves. The spec MD, state MD, commandments, and if any terms are unclear, read the glossary MD, and then provide a brief summary of the project and state how it impacts the current task. Okay.

Dalton Anderson (24:30.422) and this is implement a feature. How do you implement a feature? And so this is saying it's going to look through the read context workflow to ensure that the spec and state are understood. And then it will go through X, Y, Z steps. But once you have all this stuff set up and the agent will know and read through your rules and will set up its own workflows, which is beneficial since you don't have to manage it.

see if there's any approvals that are needed here. Approval.

Dalton Anderson (25:14.862) Okay, so it did the UI already. So we'll see how this goes.

how we're going on time. Okay, so since I'm sharing my screen, I've got the outline. Just wanna make sure I got everything on. yes. So another thing that this does is it helps implement these sub-agents to where it will delegate small tasks to look things up.

instead of having the agent doing everything, it has like a managing agent in these small task agents.

to where it delegates certain things to look up tools or retrieve information to bring it back to the actual agent that's managing the task. It's like an overseer. Okay, so now we're in this live demo section and we've gone through everything Claude code, highly suggest that you read through the repo and look through what was done and think about the thought process and how you can implement that within your workflow.

And also think about how the contents is structured within the everything Claude code and then how that can apply to your day to day. It doesn't necessarily have to be coding related. It can be related to document retrieval of knowledge. Like you have training documents, you have operation procedures at your company, you have a whole bunch of documentation. How do you

Dalton Anderson (26:48.36) structure it in a way with rules that allows AI to glean the information at the right times to maximize the utility of what you're building at your company or how do you maximize it for yourself? Kind of depends on your situation. I don't know what's going on with you, but I do know if you're listening to this episode, you're interested in being more efficient. Okay.

So I stalled a little bit. Let's see if this thing is done. Come on.

another run. Let's see.

So it's doing the test build. I have a strong, strong thought that this is probably not going to work first time. But it's live. I said it's live. So we're building it live. And as long as it's better than what we had before, then that is an improvement. we'll see. I'd be surprised if it's able to classify a image.

whether or not you used AI or not. see. Whether or not I used AI. Let's do this. Let's go. Gemini will say.

Dalton Anderson (28:10.19) Can you please create an image of a insurance?

claim.

I don't know if I'm going specify, I'll just let it figure it out. I don't know.

And then we have to use, I can use a normal model and then can use create image.

Dalton Anderson (28:38.894) Testing AI photos. I'll just call it PICS, testing AI PICS, I don't know.

Dalton Anderson (28:58.616) Okay.

So this is going through, let's see how quickly this image can generate. Okay, so this looks like a flood claim.

What is that? Okay, let's say, okay, looks good. Looks good.

But can you please make it so the person take it uses so it uses the person's POV while person's phone?

as the POV of the damage. This is an odd way. This is a very odd way to structure this. And if you're listening right now, it's a woman who took a photo or looks like they're trying to take a photo, but it also shows her, or not her, but it shows her phone to the user.

Dalton Anderson (30:19.7) okay i guess work with some other girls scene

Dalton Anderson (30:30.712) Run.

Dalton Anderson (30:35.714) So I've got the image downloaded. Let's just get this finalized. This thing's been working. I'm wasting my AI credits for you guys. All my AI credits are going away.

Dalton Anderson (30:55.692) that again.

to a selection. I wonder where it's at in this process. Didn't think it was going to take this long.

Dalton Anderson (31:09.45) Okay, so I'll keep stalling. One thing that I've been doing on the show that is odd before I record the show, and I was thinking about how odd it was before I recorded this show, was I like to brush my teeth right before the show, where I'm minty, I'm feeling fresh. I don't know. It makes me feel better and makes me feel like the words just flow out.

versus if I hadn't just brushed my teeth, it doesn't feel as clean and fresh. It's all in the mind, right? Like there is no difference between the two, but for some reason I feel better about fresh breath, clean teeth before the show. And I've been doing it every time before a show. It could be whatever time I'm recording.

could be in the morning, it could be in the evening, it could be in the afternoon.

I just do it and it makes me feel better. You might have your weird thing, your weird quirk. Think about it, reflect, let me know in the comments on what that is. I think brushing your teeth for no reason other than it makes your teeth feel fresh just to record a podcast episode is a bit odd if I'm going to be judgmental towards my actions.

But I've been doing it so long, it doesn't feel right to not do it. So this is what I know.

Dalton Anderson (32:49.636) my gosh, this thing's still going.

Free me to do this myself would take me a while. And I couldn't do all those things without looking things up like a lot of people who coded before. Before you would just look things up. You wouldn't know how to do something. You would look it up. And then once you looked it up, you would read about it. You'd read about different methods. You'd either copy and paste code. You would change it. You would write your own.

But a lot of it was you're looking at stuff online, you're bringing it back into your code base. Okay, you got it to work. AI just does it for you. It's not as fun. There is a piece of you that kind of dies. I'm not gonna go back to coding myself, but I really do miss the experience of a hard problem, figuring out the way to do it, then the sh-

trouble and pain of doing it and then going and actually completing the task. Such a great feeling. You're slated dragon every time.

Without that, you're more of a planner and coordinator, which is fun too. You're able to build at incredible speeds, but also it's a little different. You're not actually building it, you are orchestrating it, which feels different. It does. It seems like I'm whining here, but I do miss just coding myself, but it does also open up massive opportunity for a lot of people.

Dalton Anderson (34:38.05) in society that have ideas or wanna build. Okay. This thing is still going, wow. I might just have to tell it, okay, perfect, I've successfully implemented, okay, perfect. I was like, my gosh. See, it does a walkthrough. This walkthrough is pretty simple, but it tells you what it did, and then you can go through the task and what it did. All right, so now let's open up the editor.

so brutal that we're in this situation.

Come on.

Dalton Anderson (35:18.968) focusing on this, focusing on that. Let's just get this thing ripped up and going. Let's get it going.

Oh my gosh. All right, let's try it. So it says it's good. Four inner changes.

Dalton Anderson (35:41.378) Perfect. That's what I'm talking about, baby. That's what I'm talking about. Look at this. Look at this. So we'll say...

Dalton Anderson (35:56.096) it water.

leaked from.

AC unit and there is water damage.

Dalton Anderson (36:19.064) Hmm, wow, this UI looks great. This is pretty sick. System error, failed to evaluate. Wow, context, yeah. I have, I think this isn't a UI or AI thing happening. This is a timeout error that I put in because previously,

The website type of thing that I was working on was API and wasn't through my authenticator. And so if someone else got the code on accident, they could run it and they would be able to do whatever. So I didn't want that, so I had pretty tight timeout pieces. this looks, this looks sick.

Dalton Anderson (37:18.009) Like this is really cool, I think. This UI looks great. Very, very happy with this.

Dalton Anderson (37:28.013) So we had some boops in the road, yes, because the Stitch project didn't link up and didn't add in the changes, but this from, I don't know, 10 minutes and very limited instructions, this looks great. You can, yeah, and then had a little loading and yeah, this great. Okay, well, I'm happy now. I'll have to edit and compress the middle part.

of the episode because I had so many issues, but this looks good and I'm happy with it. Okay, so.

Thanks for listening. Hope you learned a lot this episode as I enjoyed sharing some of these findings. Unfortunately, it didn't go all as planned, but this is how live demos go. It shows the realness of the issue and not necessarily everything is polished every time. Don't get frustrated. Keep learning and keep iterating. Thanks for listening and of course, wherever you are in this world, good evening, good afternoon, good morning and

Until next time, see you next week. Goodbye.