The argument in one line.
Claude Code outperforms Cursor and Windsurf because Anthropic, as the model builder, developed superior tool-calling infrastructure specifically optimized for its own Claude model rather than relying on third-party implementations.
Read if. Skip if.
READ IF YOU ARE…
- A developer currently using Cursor or Windsurf who wants to understand whether switching to Claude Code is worth the friction and learning curve.
- A programmer building side projects or MVPs who values model quality over UI polish and is willing to trade some ergonomics for better code generation.
- An engineer deciding between AI coding tools right now and wants a technical breakdown of why model-provider alignment matters, not just marketing claims.
SKIP IF…
- You're already deeply embedded in a Cursor or Windsurf workflow and satisfied — this explains the theoretical advantages but doesn't address migration costs.
- You need a practical tutorial on actually using Claude Code — this is positioning and reasoning, not a step-by-step setup or feature walkthrough.
The full version, fast.
Claude Code has structural advantages over Cursor, Windsurf, and VSCode because the company building the underlying model is also building the tool layer — eliminating the gap that third-party IDE makers were trying to close with homemade integrations. Cursor, Windsurf, and VS Code all run the same model; what differentiates them is their tool scaffolding. When Anthropic builds both the model and the tools, the two optimize together rather than in parallel. The broader AI coding landscape maps into three categories: context-window coding agents like Claude Code and Codex, autonomous agents like Devin that handle larger scopes, and review tools like CodeRabbit. For vibe coders starting from scratch, the practical recipe is: start from a template, have Claude Code generate a markdown explanation of the codebase architecture, then build features on top of that documented foundation.
Members feature
Chat with this breakdown.
Modern Creator members can chat with any breakdown — ask for the hook, quote a framework, find the exact transcript moment. Unlocks at T2: refer 3 friends + add your own API key.
Create a free account →Who's talking.
00:00hostRiley Brown
00:28guestRas Mic (Michael Shimeles)
Where the time goes.
00:00 – 00:28
01 · Intro — "Claude Code is better than Cursor"
Riley sets the thesis cold: many people now think Claude Code is the best AI coding tool. He brings on Ras Mic to defend it and walk through what makes Claude Code different.
00:28 – 02:38
02 · Why Claude Code is best
Mic frames the lanes: OpenAI is going B2C / companion, Anthropic is going programming. He attended a Claude Code conference where Anthropic told him 95% of their own engineers use Claude Code internally — the rest are 'rebels' on Cursor, Windsurf or Vim.
02:38 – 07:41
03 · Whiteboard — Cursor / Windsurf / VSCode all use the same model
Mic shares his screen and draws an Excalidraw diagram: Claude 4.0 sits on top, three forks (Cursor, Windsurf, VSCode) all call into it. The model is just a knowledge store; the differentiator is the tools each IDE builds — file_reader, write_file, context handlers. That's the 'secret sauce' Cursor and Windsurf have been fighting over.
07:41 – 12:41
04 · Using Claude Code with another IDE
Mic's actual setup: pin Claude Code top-left in the terminal, keep Cursor chat open on the right, use Claude 4 in Cursor as the 'assistant who answers questions about the code base' and Claude Code as the 'executor that ships changes.' He recommends pairing both — never raw-dogging Claude Code alone in VSCode.
12:41 – 19:58
05 · Future of AI coding tools & Claude's SDK
Cursor raised $900M and Windsurf is rumored OpenAI-bound; both need their own models now because the tool moat is gone. Anthropic has shipped a Claude Code SDK (CLI today, TypeScript and Python SDKs coming). Mic's prediction: niche vibe-coding tools — React-specific, WordPress-specific — will win by wrapping Claude Code and adding fine-tuned UX.
19:58 – 21:16
06 · Codex vs Claude Code
The clean-line distinction: Codex runs in the cloud, Claude Code runs locally. Cloud-based means OpenAI controls what you can install — Mic argues local wins for power users because you can install whatever you want (and he says 'I could start malware' as the punchline). But Codex could win the mass-consumer market if ChatGPT plugs it in as a tool.
21:16 – 26:49
07 · Code review tools — Devin and CodeRabbit
Mic adds a second axis to the whiteboard: code review on one end, agent on the other. CodeRabbit is the pure code-review side (great free tier, catches bugs and security misoptimizations after a session ends). Claude Code is the agent end. Devin sits in the middle — 'AI software engineer' for feature work on existing codebases. His vibe-coder stack: Cursor or Windsurf + Claude Code + CodeRabbit.
26:49 – 35:32
08 · Building an app with Claude Code — the starter-kit playbook
Mic's recipe for vibe coders: stop debating Supabase vs Convex vs Firebase, it doesn't matter until PMF. Every project needs the same base — landing page, auth, DB, payments. Find a solid template (he has two — React + Convex, Next.js + Postgres), download it, then tell Claude Code to analyze the codebase and generate a markdown file explaining how the app works. That markdown becomes the knowledge base Claude Code uses to build features. He demos vibecheck.run — his own CodeRabbit-like web app, 70% written by Claude Code in roughly four hours over two days.
35:32 – 37:08
09 · Future plans — "we're going to build it live"
Riley locks in the sequel: next episode they'll build a full app from Mic's starter template using Claude Code end-to-end, with auth and payments, and ship it live with a public link. Mic's closing line: have fun with the tools, most are a couple bucks a month, don't take it too seriously until you have to.
Lines worth screenshotting.
- Cursor, Windsurf, and VS Code all use the same underlying model — the only real difference between them is which tools their teams built on top.
- Claude Code's structural advantage is that Anthropic — the model maker — is now building the tool layer itself, eliminating the gap that Cursor and Windsurf had been winning on.
- OpenAI picked the personable B2C companion lane; Anthropic picked the programming lane — and that lane choice is now showing up in how both companies build their tools.
- A model by itself is useless without tools — it's a resource store that can answer questions but can't read files, execute code, or make changes without a tool layer.
- If Anthropic owns both the model and the tools, every Claude improvement automatically improves Claude Code — Cursor can't benefit from the same model upgrade the same way.
- The recipe for vibe coders: start from a template, have Claude Code generate a markdown explanation of the codebase, then add features on top of a foundation you understand.
- Codex, Devin, and CodeRabbit each occupy different positions in the AI coding stack — Codex for agents, Devin for autonomous long-horizon work, CodeRabbit for review.
- Asking Claude Code to generate a markdown explanation of any codebase is the fastest way to get oriented before touching a single file.
Steal the whiteboard.
Riley Brown's interview format
Don't argue your take — diagram it live, then place every competitor on the diagram.
- Hook with the spicy take, then immediately hand the floor to the expert. Riley speaks for 28 seconds in the whole opening — Mic carries the rest. Authority by proximity.
- Force a screen-share within the first 3 minutes. Mic gets out of his face cam and into Excalidraw at 02:38 — that's what makes a take feel earned instead of opinionated.
- Build the explanation incrementally. The whiteboard starts as Claude 4.0 + Cursor + Windsurf + VSCode, then gradually adds User → tools (file_reader, write_file) → memory layer → context. Each new arrow rewards you for staying. This is the visual equivalent of escalation comedy.
- When the category gets crowded, add an axis. At 21:14 Mic draws a single line — code review on the left, agent on the right — and lands CodeRabbit, Devin, Codex and Claude Code on it. One diagram replaces ten minutes of words.
- Always end with a sequel hook. They locked in the 'next episode we build it live and ship a public link' commitment — the audience now has a reason to subscribe, not just to watch.
- Have a 'photocard' frame ready. The Riley Brown / Ras Mic side-by-side card with name labels appears 8+ times — it's an instant identifier viewers can screenshot. Brand it once, deploy it whenever you cut to wide.
- Let the guest plug a real artifact, not just channels. Mic demos vibecheck.run (his actual product) at 34:00 — the demo IS the proof of the thesis, and it gives Mic something to take home from the appearance.
Terms worth knowing.
- Claude Code
- Anthropic's terminal-based AI coding agent that runs locally on a developer's machine, using Claude models plus first-party tool calls to read files, write code, and execute multi-step programming tasks.
- Cursor
- An AI-powered code editor built as a fork of VS Code that lets developers chat with large language models, generate code, and edit files inside a familiar IDE.
- Windsurf
- An AI-native code editor competing with Cursor, also forked from VS Code, that bundles its own agent loop and recently shipped a custom in-house model called SWE-1.
- VS Code
- Microsoft's free, widely used open-source code editor that serves as the base many AI coding tools fork and extend.
- Coding agent
- An AI system that can autonomously read, write, and run code over multiple steps to complete a programming task, rather than just suggesting single completions.
- Tool calling
- The mechanism by which a language model invokes external functions — like reading a file or writing to disk — to take actions it couldn't perform from text generation alone.
- Context memory
- How an AI coding tool selects, compresses, and feeds relevant pieces of a codebase into the model's limited context window so it can answer questions or edit code accurately.
- Context window
- The maximum amount of text a language model can consider at once; larger windows let a model ingest more of a codebase before having to summarize or drop content.
- Fine-tuning
- Further training a base AI model on a narrower dataset so it performs better on a specific domain, framework, or style of task.
- Model-agnostic
- A product design that lets users swap between different underlying AI models rather than being tied to a single provider.
- Background agents
- A Cursor feature that runs coding agents asynchronously in the cloud so they can work on tasks while the user does something else, positioned as a direct competitor to Claude Code.
- Excalidraw
- A free online whiteboard tool used to sketch hand-drawn-style diagrams, often used by engineers to explain systems on a call or in a video.
- Auto mode
- A Claude Code setting that lets the agent keep working through a task without pausing for approval between steps, including writing and running its own tests.
- SDK
- A software development kit — a packaged set of libraries and APIs that lets developers build their own applications on top of an underlying platform or tool.
- MCP
- Model Context Protocol — an open standard for connecting AI models to external tools and data sources in a consistent way, reducing each tool-builder's need for custom integrations.
- Codex
- OpenAI's cloud-hosted coding agent that runs tasks on remote machines rather than on the user's local computer, letting users delegate work from a chat-style interface.
- Devin
- An AI 'software engineer' agent marketed for working inside existing codebases — building features, reviewing pull requests, and integrating with team workflows like Slack and Git.
- CodeRabbit
- An automated code-review service that analyzes pull requests on GitHub, flags bugs and security issues, and produces ready-to-paste prompts for fixing them in an AI coding tool.
- Auto-GPT
- An early open-source experiment built on GPT-4 that chained tool calls together to pursue goals autonomously, helping popularize the idea of agent loops.
- v0
- Vercel's AI app-building tool, notable for releasing a custom model trained specifically on the Next.js framework rather than relying purely on general-purpose LLMs.
Things they pointed at.
14:40toolWindsurf SWE-1 model
19:58toolOpenAI Codex
27:12toolTempo
32:05productNext.js Starter Kit (Ras Mic)
32:05productReact Starter Kit (Ras Mic)
Lines you could clip.
02:38
“Underlyingly, they all use the same model. So the model in and of itself is useless — it's just a resource store.”
04:47
“What makes Cursor, Windsurf, and VSCode unique is they have tools that they've built.”
06:50
“95% of their engineers use Claude Code. Very few are using Cursor and Windsurf. The other 5% are just rebels who want to use Vim.”
12:00
“If the model provider is finally building the tools that make tools like Cursor and Windsurf better — who better to build it than the one developing the model?”
17:58
“Cloud Code is the first time where I'm like, yeah — this this is real. Like, this is powerful.”
20:40
“On my computer, I can install whatever I want. I could start malware.”
27:50
“When I see vibe coders debate which database platform makes more sense — it doesn't matter. Believe me. Until you hit PMF and it scales, then you can worry about it.”
28:20
“It's gonna come down to grit, determination, and discipline in actually building the one feature that you need to make a solid app, which is where I think a lot of people go wrong.”
34:55
“I would tell Claude Code to analyze the code base and create a markdown file explaining the flow of the app — and that file becomes the knowledge base for every future feature.”
35:48
“70% of this code is written by Claude Code. This took me two days — two hours each day, four hours total.”
Where the conversation goes.
00:28 – 07:41denseWhy Claude Code is winning
02:38 – 07:41denseTooling is the moat (whiteboard)
07:41 – 12:41steadyHybrid setup — Claude Code + Cursor
12:41 – 19:58denseSDK + niche wrappers thesis
19:58 – 21:16steadyCodex vs Claude Code (cloud vs local)
21:16 – 26:49denseDevin / CodeRabbit positioning
26:49 – 35:32denseStarter-kit playbook + vibecheck.run demo
Word for word.
metaphoranalogystory
00:00Many people think that ClaudeCode is better than Cursor. So I brought on an AI coding expert, Ross Mike, to break down what ClaudeCode is and why it's changing the world of AI coding. We talked about how ClaudeCode is different than tools Cursor and Winsor.
00:15We talked about how to get started building with ClaudeCode. We looked at an app that Mickey built fully with ClaudeCode. And we also talked about why Anthropic is building ClaudeCode the way they are.
00:27Let's dive in. Ross Mike, in your last YouTube video, you said that Claude code is the best programming agent, and I've actually heard other people say the same.
00:38Do you really think that it is better than tools like Cursor? I think so. I believe so.
00:44Not only have I used it, like, very much in the last two weeks, but I also understand why I think that way. And I wanna kind of explain if it's okay with you, share my screen and sort of, like, just get into Excalidron diagram why I think so.
01:00More than anything, one thing that we all understand is that Claude is the best model for programming. Right? I think we all understand that all the big providers have sort of picked their lane.
01:10Like, OpenAI seems to be going for the more personable b to c feely, you know, human companion type vibe, and then Anthropic has picked its lane where it wants to go into the programming lane.
01:21And I actually was at a Claude code conference where they announced Claude code. And the big difference really is if you're using Cursor, Windsurf, zed, like, whatever it is.
01:33Right? I know everyone has their camps. Like, we have Cursor here.
01:36We have Windsurf here. And then let's say you're, you know, an OG and you you're using Versus Code. Right?
01:44If you're using any of these, like, tools, you notice that some people are like, oh, I think Cursor's better.
01:51Some are like, oh, I think Windsor's better, and some are like, oh, they're the same. I just use Versus Code. Underlyingly, they all use the same model.
01:59Now the question is what really makes them different? Why could one user think Windsurf is better than one user thinking Cursor's better? Now there is some sort of anecdotal point to it where, like, maybe someone just had a good run, and they think it's better, and someone had a bad one with the other tool.
02:14But, really, the the difference is they're all using the same model, but they're all building their own tools.
02:22Right? So the model in and of itself is is useless. Right?
02:25It's just an image it's just a resource store. It has all these all this knowledge. Right?
02:30So you can ask it questions, but if you tell Cloud four, read this file in my code base and tell me what it does, it has no way of doing so. So what makes Cursor, Windsurf, and Versus Code unique is they have tools that they've built.
02:46Right? So for example, like, you use Cursor and you you look at the tools it calls, like, you'll see something like file reader.
02:54Right? And then, like, you or you'll see something like I don't know if you've ever, like, seen, like, Cursor or Windsurf when they're about to push an update.
03:03Sometimes, like, you see gibberish. If you look at the gibberish, it'll be, like, writing file. Right?
03:08So you'll see that what makes Cursor and Windsurf different is that although they're using the same underlying model, they're developing their own tools.
03:19Right? So Cursor has its own file reader tool, and the way this works is maybe I can, like, illustrate this better is, let's say this user, right, goes to Cursor.
03:31Right? This user is using Cursor, and this user asks cursor, cursor, can you change the color of this page to green?
03:40What happens is cursor fires that off to Claude four o, and Claude four o is gonna say, oh, the user wants me to change the color of this page. I need to know where the page is. So what Claude is going to do is Claude is going to fire off the tool that Cursor has provided for Claude, and that tool is going to read the file, give the contents back to Claude.
04:01Now Claude is then going to be like, okay. This is where the file is. This is the code that I need to change.
04:07Now I need to change it. So Claude then fires off the other tool, and this tool allows it to write to the file. Right?
04:14These tools are developed by Cursor, Windsurf, Versus Code. And how they've implemented these tools, how these tools work together, the efficiency of the tools, how they handle context memory, this is sort of the that Cursor and Windsurf are fighting for, which is why if you notice, Windsurf developed its own model and Cursor raised a bajillion dollars, and they're probably gonna develop their own models.
04:38Because right now, what they're fighting over is this. This is this their differentiator right now because everyone is using Claude four. Right?
04:46But what happens is when they have their own models now, they can have their own models that are trained a certain way, that becomes an edge for them. That's a business podcast. But, underlyingly, this is how Kerser and Winsthorf work.
04:57They have tools, memory, different ways of handling context. Like, they'll even maybe fine tune specific models for certain things. And, you know, working at a an AI startup, I know this.
05:09It's not just one giant Claude agent. They probably have multiple agents, maybe even using different models.
05:17Right? Maybe they have, like, a specific model and they use Gemini, like, 2.5 flash because it has a large context window that ingests the entire code base and gives, like, summarized context to Claude four because Claude is expensive.
05:32I say that to say this is the edge that Kursor and Windsurf have. Now where does Claude code come into play?
05:41Claude code, although not a similar interface, meaning they didn't fork Versus code, Claude code is essentially built on Claude's model.
05:52Right? But they, this time, have built their own tool calling. And this is sort of the understanding I have.
06:00Like, I spoke with some of the team at Claude at the conference, and first and foremost, they're going all in on this. Second of all, they told me 95% of their engineers use Claude code.
06:10Very few are using Cursor and Windsurf. The other 5% are just rebels who want to use, like, VIM or, like, maybe they'll use, like like, you know, Cursor or Windsurf.
06:21But majority of their developers are using Cloud Code. It was an originally an internal tool. Now they've made public.
06:28And the thought process I have is the one developing the model will probably know the best way to develop the tools. Right?
06:37And this is why from my personal usage again, I don't have, like, data, like, intrinsic data for you to say, oh, this is the percentage. But for my personal use, especially with larger code bases, Clawr code seems to understand code better and execute code better, and it does feel more agentic.
06:55Right? You can allow it to continue on. Like, I think it's called auto mode or something.
06:59You enable it. And not only will it continue till the task is task is complete, but there are certain instances where it will try to write tests, and it'll test the code. Right?
07:09It does feel a lot more agentic than Cursor and Windsurf and Versus Code. And Cursor launched with one point o background jobs or background agents.
07:18I think that's what they call it. And that's what I would say is a direct competitor to Cloud Code. So the reason why I think Cloud Code is is better is because the model provider is finally building the tools that make tools like Cursor and Windsurf better, and who better built it than the one actually developing the model.
07:36So that's kind of my spiel and my thesis as to why I think ClaudeCode is better.
07:41When you use ClaudeCode, do you use it within Cursor? I've heard some people use ClaudeCode in Cursor. And so, like, is that just, the ultimate stack?
07:48Because, like, I tried to use ClaudeCode, and I actually asked Cursor how to set it up. And I actually got it set up, but I didn't have the normal IDE. So it was hard for me to go to the right file, like because I was, like, uploading images, and I was trying to give it context of the file that I wanted to reference, but, like, there was nothing to look at.
08:05So can you talk about why people use Claude code within Cursor? And then, like, why you think this like, what do you think of this interface of Claude code versus cursor?
08:16It's definitely very geared to developers. And my actually favorite setup is don't know if you can see my cursor is the way I kinda have it set up is I will have Claude like, I'll open Claude code, and and I I can show, like, how, like, you set it up.
08:35But I just wanna give you, like, a quick rundown on how I have it set up is I'll pin Claude code to the top left, and then I'll still have my terminal here, and then I'll have my cursor chat.
08:47And what I've really been enjoying is I'll use cursor like, I'll use Claude four o for asking. So, basically, like, I'll build my understanding of what it's doing, how this code base works, or I'll build, like, my questions.
09:01Like, this is, like, the assistant I ask questions, and Cloud Code is sort of my executor. So it it is true in a sense where I think, especially for a Vibe coder, it's best to mingle the two. I don't think you should use Cloud Code by itself with, like, a Versus Code or just raw dog it on terminal.
09:16I think it becomes very confusing. But if you couple it with, let's say, your Windsurf guy, Windsurf and Cloud Code, or you're, you know, a cursor person, cursor and Cloud Code, I think you get the best of both worlds.
09:29And here's here's the best part. I I know there's a lot of compare comparison videos, but you can use both, right, which to me seems like a great compromise. Right?
09:38I can use cloud code. And if it feels like it's not my type of thing, I can use, like, Cursor Chat.
09:44But I think what makes Cloud Code special is the model provider is building the tools. And I don't know if I've just been on a hot streak, but I've had instances where, like, the regular, whether it be cursor or windsurf just would not get it done. Quad, even though it took longer, got the task done.
10:01And the thesis that I'm going with is the same thing I shared is I think the model provider building the tools for the model, they just end up knowing what works best. Right?
10:11So I highly, highly recommend if you're gonna give Cloud Code a shot, definitely couple it with your favorite ID, whether it be Cursor or Windsurf.
10:19Because you have to remember, I mean, these companies are hedging. Right? Like, Cursor and Windsurf, at least up until a few months ago, are building their models to be are building their product to be model agnostic, to use any of the models.
10:32Now with Windsurf, it's interesting because Windsurf is rumored to be bought by OpenAI for $3,000,000,000 and now is Anthropic cutting them off of tokens, uh, intentionally or unintentionally.
10:44I don't wanna get into the rumors. I don't think it matters that much. But, like, if you think about it, it makes sense that the the company like, if you really think the edge is in the tools that it uses, like, whatever like, if quads ends up being the best coding model in the world, you would think that Anthropic would be the best at building those tools around it because they can build the model specifically for those tools.
11:11Exactly. Exactly. Yeah.
11:13They they have the And it They literally have the cheat code. And then it just makes sense for them to go all in at that point, because if they can own programming, like, that seems you know?
11:21And that that makes sense to me that Claude just owns the programming side, and OpenAI owns the kind of the the the become your best friend personality side or whatever hardware they're building. So that's really, really interesting.
11:34Kursor and Windsurf, I think they are starting to see that
11:38Claude is sort of inching in the direction where at first, there was a big separation. Right? Claude was like, we'll just provide the model.
11:44You two battle it out with UX. You two battle it out with actually getting users. Now it seems like Claudel's stepping in that front and Kerstro and Windsurf.
11:52You also wanna I think the moat for them becomes the model. Right? V zero launched, like, their own specific model that's trained on Next.
12:00Js and stuff. And I think, like, for tools like this, if they build their own specialty models based on, you know, all the training data they get from the users or they see what people really like to use, from a business perspective, it does make a lot of sense. They also get to raise a lot of money to justify doing that.
12:16Right? So this is sort of how I've been thinking about it. I don't know if you've used Windsurfers.
12:21I think it's SWE dash one. It was actually a decent experience. I didn't really get to use it as much, but it was very fast.
12:28It it it felt like the tabbing felt very fresh. It felt like it was a pleasantly good experience. So I do see cursor dash one or Windsurf version two coming out soon.
12:39So that's why I'm kind of like like and if you're like an AI builder, like, Cloud Code has an SDK coming out too where you can literally build on top of it. So it it seems to me they're they're definitely coming for everyone's lunch and Wait.
12:51Talk about that for a second. You said it has an SDK, which means people can build their own custom interfaces for it?
12:59So, yeah, instead of, like, you know, your your your tempos, your lovables, your bullets, your vibe code, like, instead of, like, building the tools for your agent and all that stuff, they're literally like, right now, it's terminal based, but their team said, if I'm not mistaken, yeah, the SDK currently supports command line usage, TypeScript, and Python SDK coming soon.
13:20So they're entering a world where they're telling the up and comers these guys will probably own always build their own stuff, but the up and comers, like, you focus on UX, we'll handle everything else for you. Right?
13:32So you can imagine yourself building a very clean seamless UX that literally runs on Claude code, and you end up having a model that competes with all these other guys.
13:44Right? So it it is an interesting I think there's gonna be some use cases for that that aren't, like, just broad coding things.
13:52Like, I feel like there you like, there's gonna be, like, microcoding tools for one specific niche that might be on the mobile device or I don't know.
14:02I've never thought about this before, but my mind is running with a bunch of different ideas. Because if you you could basically create a Claude code wrapper. And if you think Claude code is as good as Cursor, then, you know, we're gonna get, like, something that is as powerful as a Cursor wrapper.
14:16And that's terrifying. Like, you know, as someone who is we're building a a a a vibe coding startup, you know, it's it's terrifying to think that anyone can create a Claude Code rapper, but it's also very fun and interesting and maybe yeah.
14:29I don't I don't know. I don't even know what to say to that. I mean, what do you six months from now, what do you think that looks like?
14:35Like, what do you think people are gonna build on top of of Claude Code? You're right. What happens when everyone has access to the greatest agent, coding agent, and it can build on top of it?
14:45I I do think
14:47niche or niche, I never know which one is the correct pronunciation, specific, like, platforms are going to win.
14:53Like, for example, let's say you have a vibe coding tool that's specific to React. Right? Or you have a vibe coding tool specific to, like, WordPress or whatever.
15:04Right? I think, like, the niche tools are going to start to win because you can sort of, like I I I they haven't mentioned any fine tuning, but I can imagine, like, if you could fine tune the model to specific, like, types of programming or frameworks or stuff like that and you build, like, specialized user interface and you have content.
15:23Right? The model is definitely the model arena is definitely getting taken over by Cloud Code.
15:29And unless one of these guys and that's their goal, and that's why Cursor raised 900 freaking million dollars is because the only thing you can really build to withstand this behemoth is build another one.
15:39Right? So Build Anthropic. Yeah.
15:42That that's literally all you could do. Right? Because, like, at some point, like, the tool calling like, at first, I remember when GPT four just came out.
15:50We were, like, on 3.5 to four. Like, people were building, like, these open source. Like, I remember one of them was called, like, auto GPT, and the uniqueness was, like, being able to, like, have tools.
16:04Right? And tools were a big deal. Now tools have become normal, and then we have the MCP era, and, like like, this moat that these tools had is disappearing very, very slowly.
16:15And on top of that, they forked Versus Code. We can all fork Versus Code. Right?
16:20So the the the the big differentiator they can have is to build richer UX on top of what they have, which Cursor made a big bed background agents. They I don't know if you saw. They have their code reviewer, which is like a CodeRabbit alternative.
16:34Right? So they're building around the IDE, but then they're also going to build a model, which makes sense. But for And also they're also building a mobile app too.
16:44Like, Kurz was building a mobile app. I I I didn't even know that. So that's insane.
16:47Yeah. Yeah. Yeah.
16:48I follow their head designer on on X,
16:51and, yeah, they he's been posting it for a couple months now, some sneak peeks of what it will look like. So they're building a mobile app. So, yeah, I think yeah.
16:59I mean, it's it's probably time for Cursor to expand out into other niches and and try and figure out other ways to get some sort of edge. Because I do every single hike, like, at at our company, our back end developer was like, oh my god.
17:14Claude Code is so much better than Cursor right now. Like, he is, like, blown away, and he's just like, I literally tried Cursor and Windsurf for thirty minutes, and then Claude code just one prompted it.
17:25And, like, he'll just, like, enter a prompt and then just, like, walk away for seven or eight minutes and come back. The crazy part. Like and that's where, like, it actually
17:33feels agentic. I know we've been saying agentic this whole time, but Cloud Code is the one. And I sound like a shill.
17:39By the way, mind you, they're not paying me. I wish they were. But it genuinely it it's the first time I felt agentic, and it I'll I'll see it's the first time I felt like, yeah.
17:49Programmers who don't use these tools, yeah, like, you're you're you're definitely going to get, like, outshipped. Like, I know we've been saying this this whole time, but for the most part, I'm like, man, a really talented senior developer can, like, go and see. But Cloud Code is the first time where I'm like, yeah.
18:05Like, this this is real. Like, this is powerful.
18:09So one question I have for you is, like so Codex was recently released, like, in the last year by OpenAI. What is the difference between Codex and ClaudeCode? And then something like Devon, what about those three tools?
18:21Because it seems like they don't have kind of the standard interfaces like these three, uh, tools that you have here. What is the difference between these tools? I'll explain the difference in Codex and Cloud Code.
18:33So the underlying difference is Codex works on the cloud,
18:36and Cloud Code works on your local development, like, on on your computer. And there's pros and cons to both. I think the right bet is local just because like, and especially if you're building, like, serious applications and there's maybe certain things you need to install.
18:56Like, you're limited to what OpenAI allows you to install with Codex. Right? Versus with Cloud Code, I mean, on my computer, I can install whatever I want.
19:04. Right? So, like, that's, I would say, the big differentiator and, obviously, the models that they use.
19:10I've used Codex before. Codex is pretty good. It's not Cloud Code, though.
19:15Like, I will just say this. In terms of, like, sheer power and output and speed, I definitely think Cloud Code takes it. But, again, when I look at Codex and them betting them making it on the cloud, it makes sense for, let's say, if you're a Vibe coder or you're not really comfortable.
19:32Like, I know many Vibe coders are comfortable with Cursor and Windsurf, but let's think even broader. Most people are not. Right?
19:38And if you can go from your, like, friendly chat assistant to being able to tell it to build something and they almost make you feel very intuitive and clean and simple, then Codex wins big with that market. Right? So
19:51I would say Right. You could almost make Codex a tool that ChatGPT could have access to in theory if it's in the cloud.
19:58Like, you could say, hey. Create an app, and you could literally just do this in ChatGPT, and then it just says, okay.
20:03Running that tool, and then maybe it sends it back to you, and it's just a web app link that you can open up right there. And you can look at it, you go back to ChatGPT and say, no. Change it.
20:13And, like, that is that's pretty big, and I don't I don't count that out for OpenAI.
20:18If there's one thing that OpenAI has proved to us over the year is that they have knocked the park out of consumer. Right? Perplexity held search down, and then ChadGBT came with their search massive.
20:30Right? Deep research came massive. Right?
20:33ImageGen came massive. Right? So I definitely wouldn't count them out in terms of, like, being able to build consumer products or things that consumers will wanna use.
20:43I just think in terms of, like, if I were to pick one, I would go with Cloud Code because Cloud Code just seems to and I think it's a it's a design choice that they went for the terminal because they don't really want to build the UX layer for you and all that stuff. Use whatever you're already using, and this is truly an agent that you just install in whatever you're using.
21:04Right? So I think it's a purposeful design choice, and that's what makes, in my opinion, Cloud Code better than Codex. It's more for me, for you, for someone who's a power user.
21:12Devin is different, though. I would put Devin in the same say, this is code review.
21:20Like, this is one spectrum of code review, and let's say this is another spectrum of, like, agent, I would put Devin somewhere around here.
21:32Right? Maybe maybe in the middle. Right?
21:34Somewhere, like, in this spectrum where I don't think and I haven't used Devon heavily, but based on the people I know at different companies who work at, they'll use Devon for speech feature specific team things. Right? So you already have an existing code base.
21:48Right? It's functional. It's in production.
21:50You have your Git workflows. You have a way of deploying to production. You have a staging environment.
21:56Devin integrates well with that. You can tell it what to do on Slack, and it will basically build you a feature.
22:03Right? Or it will review an existing feature that you've built. Right?
22:08You have then CodeRabbit. Right? And I'll maybe make this clearer for everyone to see.
22:13I would put let me just clear this up. I would put CodeRabbit here.
22:20And, by the way, I would 100% use CodeRabbit. Like, their free tier is pretty good if your thing is open source.
22:26And if it's not, I think it's a couple dollars a month. Like, CodeRabbit is a great code review. So, like, whatever they do at I know they're not an agent, but what what at CodeRabbit, it is great.
22:36I can't tell you the amount of bugs I've caught using CodeRabbit. So excellent at what they do. Cursors So it it just just to clear things up.
22:44I've used CodeRabbit. I've actually have CodeRabbit. I used it, like, once.
22:48And to my knowledge,
22:50it's like when you ship a new feature and you upload it to GitHub, it will basically look at the changes you made and be like, hey. Like, this is good or like, no.
22:58You should change this. You could do this better, and it helps you it's like one layer to, like, catch mistakes or something like that. See, the thing is it's great because if you've noticed this, like, Riley, and I like I love your feedback, is whenever you start off a project with any model, really,
23:14it always starts strong. But as it continues to go on, like like, either performance goes out of whack or the model just started losing like, it's easy to start. It's hard to finish.
23:23Right? And there's a thing with, like, maybe the model has too much context. This is why they tell you create a new chat, start fresh, and all that stuff.
23:31With CodeRabbit, it gets your code all at once, and it just I think because it's not being repetitively used like the other, like, tools, it just is great at catching, like, either flaws or security mistakes or misoptimizations.
23:49Right? So I think So you're saying that, like, when you're using Cursor, because you need to manage your context and you can't speak into the same agent window for hours, that, like, it's gonna lose track of what you're doing in while you're using Cursor.
24:03But this is kind of a good thing. Like, once you end your full session, which might be 10 different agent chat windows,
24:10this will see it all at once and analyze it all at once so that you don't miss the big ones. Okay. So this is my technical theory.
24:16Right? It could be something else, but based off my experience, CodeRabbit, like, it's just been like and this is from, like I'm talking from also other developers. I'm using it at big projects, and it's just been great.
24:27And it's something simply you install, and it literally one thing that I love I don't know if I have a repo I can show, but it will give you a prompt to give the agent. Like, if you're using CloudCoder cursor, it will tell you, this is the bug that I found.
24:41Copy this prompt and paste it in whatever tool you're using, and it will fix it for you. Right?
24:47So I think this is one of those quick wins that if you haven't been using the program, they're not paying me.
24:54I wish they were. Hopefully, they do one day. It's it's a simple addition.
24:59And then for agent, we'll have Claude code on this side of spectrum. And then Devin is somewhere here where it's great at code review, but it can also build features.
25:13But I haven't heard any stories of people building Devin, like, something from scratch.
25:19And to the company's point, they called it, what, an AI software engineer.
25:25And more than likely for, like, big companies, like, a software engineer is not going to be building something from scratch. Right?
25:32They're probably gonna build and maintain features. Right? So if that's what they're building Devon for, that's perfect.
25:38So if you're a Vibe coder and your product hasn't hit product market fit and you haven't hired developers, I don't really think you need Devon. I think you get you something like a cursor windsurf plus Cloud Code if you wanna be fancy and CodeRabbit to review your stuff.
25:55I think you have a solid stack. K. That's yeah.
25:58You heard it here. And then so codex codex would be to the right of Devin Yeah. So codex left of Cloud Code.
26:04Yes. Something like somewhere right here. And mind you, not saying it's more of a code review, but I would just say even though I would align it similarly with car to code, I think Cloud Code is just more powerful simply for the fact that Codecs is cloud based.
26:19And that's the problem with cloud where, let's say, you wanna spin up a server. Let's say you wanna have multiple servers. Let's say you wanna you know, whatever that it is that you wanna do, there's a certain limitation you have with cloud versus with your own local machine, you can run and install whatever you want.
26:33Or even, like, security reasons. Right? Let's say, you know, I wanted to build a, you know, blockchain project.
26:38Why? I don't know, but let's say I do. Right?
26:40There there are certain security things and keys and stuff like that that I just wouldn't want on the cloud versus with Cloud Code, it's all on my machine my machine.
26:49So if I were to be launching into a long project, right, like you and I are gonna do in a future episode, where do you start? Right? I think that is a big question for VibeCoders.
26:59I think what, you know you know, Lovable, Tempo, Bolt, the whole nine yards, all those companies do a really good job of doing is kind of giving you a place to start from. You know?
27:09Like, if you just open up Cloud Code, it's hard to know what to do. I know you can open up Cursor, a fresh project on Cursor, and you can say, build all the files, and it will just do it.
27:18I guess, like, I guess in the next episode, when we when we build something from scratch using only AI and Cloud Code, how are we gonna start and, like, how are we gonna think about it?
27:29Here's the way I I I think of this. Right? And and this is stuff I've done for myself without AI.
27:35You realize that most projects have sort of the same base foundation. Meaning, you'll need a landing page. Right?
27:42Simple. You'll need authentication. Right?
27:45You'll want to authenticate the user. You want to store some user information. Right?
27:49You want the user to be able to log in with Google, email, whatever it is. Right? You'll want a database to store this information.
27:55Right? And here's one thing I wanna say. The one thing that actually frustrates me, Riley, is when I see vibe coders debate on which database platform makes more sense for you.
28:05It doesn't matter. Believe me. Until you hit PMF and it scales, then you can worry about it.
28:10But until then, bro, you use super base, you use convicts. By the way, I prefer convicts. But whatever it is that you use,
28:17doesn't matter. Right? It it it it doesn't matter.
28:19It's the , you know, like, where, like, you know, the the people that mean, it's like it's like if if if the the database because I looked at the numbers. Right? It if when you're using Firebase versus if you're using Supabase or or InstantDB, you know, there's a lot of different tools you can use.
28:37It really doesn't become that big of an expense until you're really you're making us so much revenue that, like, hiring someone to come in and figure that out for you is gonna be very easy. Right?
28:50Yes. Especially if you're if you have the proper monetization in the app that you're creating. Right?
28:54Like and so, yeah, it doesn't really matter. And even I think the cursor windsurf debate is kind of not even worth it, really. Because at the end of the day, it's it's gonna come down to, like especially if you're vibe coding and you're not using AI, it's gonna come down to, like, how meticulous you are.
29:08Like and, like, it's gonna come down to grit, determination, and then discipline in, like, building actually building the one feature that you need to make a solid app, which is where I think a lot of people go wrong.
29:20They try and build a a ton of different features into an app, and then they just get lost when they should have and could have easily built that one feature. The I, like, I call it the that is at the center of their marketing, and they just could have marketed that one feature really well.
29:35And once you kind of have that revenue coming in, then you can worry about expanding it out and even hiring a team to come help you. We need a DB. We need auth,
29:44and we need to accept payments. And one thing I'll add, like, if there's one thing I would focus on is, like, there there's a term called DX developer experience.
29:55I would focus on tools that have that focus on ax, like AI experience. So, like, what do the AI models what are AI models good at? If you notice, TypeScript is a big thing with AI models.
30:08And the reason why the models will do good with TypeScript is because with TypeScript, there's a way to check if things are done right with types. Right?
30:16And the and this is getting technical, but the AI has a way to go through the code. And even though it might be hallucinating, it can check if it's done things right based on the rules of TypeScript.
30:27Right? So, for example, for me, I usually use the database provider or the back end provider convex because I realized that AI is really good with writing Convex code just because it's TypeScript.
30:38Right? At the end of the day, whatever you use, until you start to really, really scale, you won't feel it. Once you scale, you've made money, you're rich, congrats, hire some developers.
30:48You know? Send back the money to to the developers who who who built these tools. But I say that to say, if I were to build something using AI, I would first find a solid template.
30:59Now, Riley, I don't know if you have a template. I have two that I use on my GitHub.
31:04I have a React starter Back
31:06when we first started, yeah, we had templates, but they are out of date at this point because they go out of date relatively quickly. But I saw that Yeah. Yeah.
31:14Yeah. And I saw that yeah. You you messaged me that you have a template, and I've actually forked this before, and it works really well.
31:20And you're the first one who's started talking about payment templates, which is actually, like, one of the biggest things that, you know, people wanna make money with these tools. And so you've kind of built that in, and, obviously, there's keys you need to plug in. Right?
31:32If you have a Stripe integration, you need to plug in your Stripe account, and that can be learned over time.
31:38Yes. But, yeah, anyway, keep going. Yeah.
31:40So once like, I have two. One was with React star one's a React starter, one's Next starter. The only difference is this one uses Convex.
31:48This one uses a Postgres database. If you're not a developer, it doesn't matter. If you're a Vibe coder, I would start with the React starter kit, and we'll talk about this in the next video.
31:57I also have a video on it on my channel. It it's just easier to set up versus the next starter kit. There there's some, like, nuances that are a bit more difficult.
32:06So I would start with a starter kit. Get all my payments DB auth stuff figured out. And then what I would do using Cloud Code and this is this is this is key, Riley.
32:17I would tell Cloud Code to analyze the code base and create a markdown file explaining the flow of the app, like how it works, how do the payments work, what it used, etcetera, etcetera. The reason being is now I'm going to use that file as a base layer, a knowledge base for it to continue to build features. Right?
32:35So the plan is going to be find a template, download the template, set it up using Clockwork, tell me all the keys I need to set up, then build a markdown file telling me how the app works. And then what I'm going to do then is we're going to start building out features on top of this so we have our payments off and DB figured out.
32:54We just need to figure out the features and what we're going to build. And here's the cool part. I kid you not.
32:59Couple days, like for funsies, I would say 60% of the code is written by Cloud Code.
33:06I I wanted to build, like, my own my own CodeRabbit, but on the web.
33:13Right? And I log in with with GitHub. I'll click on analyze, and then this pulls all my public repos.
33:23I'll click analyze repo, and I just want to show you what you can do with Cloud Code. This is basically a full on editor. I I I did not I did not write the code for this.
33:32I I I Wait. Wait. You you use Claude code to make this tool called vibecheck.run.
33:37I saw I saw this on your Twitter, and I was just like, what is this? Is this a new cursor update? Like, what's going on here?
33:43This is and this is where it blew my mind.
33:4670 I can confidently say 70% of this code is written by quad code. And then I'll click, like, security analysis, and then I'll call the quad four o.
33:57I have an agent that runs through all the code and then will suggest changes. Right?
34:03And then these changes, I haven't worked on this feature yet, but I can push these changes to GitHub, create a pull request, and someone can merge that code change in. This took me two days.
34:14And when I say two days, I'm talking about two hours each day, and this was done through ClawdCode.
34:20Right? So in four hours, basically, in four hours, right, you could have done this in a four hour window. You created this.
34:28This is insane. Just so you could do the design alone. The design for the chat is fantastic.
34:33And here's the thing. And it this would've maybe two weeks.
34:37Like, I could have done this. Like, but the impressive part for me is not that it did something I could do. The impressive part is the time it saved.
34:45Right? And with Cloud Code, like, again, I skimmed through the code. I looked at it.
34:49I saw what it did. I checked everything. And, like, Riley, I I I did this in, like, four, five hours of just me relaxing and and and just chilling.
34:58And this is why I think, like, you find you a solid template. Right? You find your solid template.
35:04You, you know, get you cloud code to analyze the code base.
35:11By the way, don't judge me for my spelling. Cloud code base. And then you build features.
35:15I am telling you, you you're gonna have a great time. Will you get stuck at times? Of course.
35:19Will things get difficult? Of course. We don't need to hypothesize about it.
35:23We're gonna do it within the next week or two. We're gonna do this. So here's the here's the end of the episode, is we are the next time Mickey and I talk, we are going to build an app with Cloud Code starting from your template that's going to is it it's gonna have auth, it's gonna have payments Everything.
35:40All of it. It's gonna have everything, and we're gonna ship it live.
35:43And by the end of the episode, we're gonna have a link for people to go to to see it. Right? That's the goal.
35:48We're gonna make a whole episode. Make sure you spam, like, share, whatever it is. Hell, yeah.
35:54And then by the way, to everyone who's watching, I'm going to link Mickey's YouTube channel. I'm also gonna link your starter template, and then whatever else you want me to link, your Twitter, etcetera, etcetera.
36:04Thank you for coming on. This is actually I think it's really useful to put these tools next to each other to kind of get this mental market map, and I think it's actually useful to do with other categories as well. Like, I was trying to look at the difference between n eight n and Zapier and Make and try and figure out, like, what different tools are on the market.
36:22Like because you can't really use the tools unless you kind of know kind of it's it's always helpful to get that kind of perspective, and so I thank you. You're always, like, very good at explaining things. This is this was a lot of fun, so thanks for coming on.
36:35First of all, I appreciate the privilege. I appreciate the kind words. Thank you so much for having me all.
36:39And, yeah, like, I just like man, like, just have fun with the tools. Right? Like, a lot of them have free trials.
36:43Like, a lot of them are a couple bucks a month. Like, you know, just have fun. Like, don't take it too seriously until you have to.
36:49100%. 100%. Alright.
36:52I'll talk to you later. Alright. Bye, everyone.
36:54Thank you, Riley.
The Hook
The bait, then the rug-pull.
Riley Brown opens with the cold thesis and then steps out of the way. The whole 37 minutes is really one argument with three acts: why Claude Code is winning, where Cursor / Windsurf / Codex / Devin / CodeRabbit each fit on the map, and how a vibe coder should actually start a project today. The lesson for creators isn't the take itself — it's the whiteboard. Ras Mic builds the entire frame live in Excalidraw, which makes a take that could've been a tweet feel earned over half an hour.
