980: AI Coding Explained

Wes and Scott talk about the state of AI coding in 2026—from editors and models to agents, skills, slash commands, MCPs, and more. They unpack what these things actually do, how they overlap, and how to use them effectively without overcomplicating your setup. Show Notes 00:00 Welcome to Syntax! 01:39 The tools: editors, terminals, GUIs 05:27 Wes’ and Scott’s current AI setups 13:17 Picking the right model 18:58 How exactly do agents work? 22:32 Subagents and parallel workflows 24:29 Brought to you by Sentry.io 24:54 What goes in agents.md (and what doesn’t) 26:47 Skills vs agents Skills Superpowers 34:03 Slash commands as reusable prompts 36:02 Hooks and keeping your code from going off the rails 38:00 Plugins and bundling your setup 39:24 What MCP is and why it’s powerful 40:54 Cloud agents and running jobs remotely 43:47 Choosing the right AI tool 47:41 Sick Picks + Shameless Plugs Sick Picks Scott: ULTRALOQ Bolt Fingerprint WiFi Smart Lock Wes: St. Denis Medical Shameless Plugs Syntax YouTube Channel Hit us up on Socials! Syntax: X Instagram Tiktok LinkedIn Threads Wes: X Instagram Tiktok LinkedIn Threads Scott: X Instagram Tiktok LinkedIn Threads Randy: X Instagram YouTube Threads

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Scott Tolinski?

  If you're new to AI coding, or you think you know AI coding, we are going to break down all of the different types of files, techniques, and then jargon that you're going to need to be effective at AI coding. But mostly, what the heck is all this stuff? Where does it go? How do I use it? And how do you become effective? My name is Scott Tolinski. I'm a developer from Denver. With me today is Wes Boss. What's up, Wes?

• 0:25Wes Bos?

  I'm doing great. And like, there's a lot to unpack here. There's, it seems like every single day, there's some sort of new file, or some sort of new word that we're trying to learn. Like, we were talking about cursor rules many months ago. And now, nobody talks about rules files. It's all agents.md. And there's also, what? Skills, and subagents, and hooks, and MCPs, and plugins. And there's all kinds of stuff. And we gotta break it all down for everyone.

• 0:53Scott Tolinski?

  What happened to us? We used to talk about cursor rules, and now it's all agents.md. We got, we used to sit and talk about slash commands. Yeah, let's get into it. First and foremost, I think what we want to do is we want to start off talking about some of the different types of tools. Then we want to get into what all this stuff is in terms of, like we said, whether that is models and agents and subagents and those types of things. And it's mostly going to be breaking down, like, what are all the pieces here? What do you need to know? What is actually important? What is effective? You know, in another video, what we'll do is really break down our real detailed tips on being effective with these tools. Once you know what everything is, where it all goes.

• 1:36Scott Tolinski?

  But we're going to do our best to make sure that you know where the stuff is. So first and foremost, the tools that you're using for AI coding. The primary ways that people are interacting with their code these days are via their text editor. So something like VS Code Copilot. You have Zed, you have Cursor, you have something like Kiro. You have these normal text editors with AI features. Then you have the terminal-based ones, which are either CLI-based or TUI, which is like a terminal UI. Something like Claude Code or Open Code TUI, where it is in your terminal, and you're typing into a text box, right? And then you have something that is like a full-on GUI, which is like a desktop app,

• 2:21Scott Tolinski?

  something like the new Codex app, or Open Code has an app. And these things are much more like a chat box, like a chat GPT inside of your your app and an app that you can, you know, drag files into and it has more information and a nicer UI typically than the TUI. So Wes, what are you using as your main tool for AI coding these days?

• 2:45Wes Bos?

  Yeah, the tools I'm using for AI coding right now are constantly changing because I'm trying to learn them all, but I find myself migrating to, when I'm like working on existing application, I find myself migrating to Cursor and mostly Agents. If I'm trying to fix something in the CLI, I find myself going for either Cloud Code or Open Code and just saying like, hey, I'm working on X, Y, and Z, just for like short little spurts. And then like the TUI itself, I'm curious if you think this way as well, is like, I do love a good TUI, but not for like long-running things. A GUI is just so much more flexible and I don't think you can ever replicate the experience that you're getting with a GUI. So I always try to go a little more on the GUI side.

• 3:31Scott Tolinski?

  Yeah, man. So the tools that I use for AI coding right now are specifically GUI-based. I am a GUI man. I use a GUI over a TUI. I've stated as such, but I like, me personally, Open Code, I use the desktop app or the Open Code web interface because I can run Open Code web on my Mac Mini, visit it in the browser and it's a GUI. And then I'm connecting to my app in my normal text editor to code and read the code, et cetera. But I'm, the tools that I'm using to do the AI coding themselves is pretty much exclusively Open Code's GUI. Reasons why I like it?

• 4:14Scott Tolinski?

  I think GUIs offer things like when you drag an image, it shows the image, stuff like that. It shows better diffs. The diffs are nice. The UI, you can click on something and it flies open and closed and the keyboard shortcuts.

• 4:27Wes Bos?

  Information density is better.

• 4:28Scott Tolinski?

  Yeah.

• 4:28Wes Bos?

  You can see the terminal when it's running. Yeah.

• 4:33Scott Tolinski?

  Yeah, I greatly prefer it. Yeah.

• 4:34Wes Bos?

  I agree as well. Plus, like, I find one thing with the dedicated GUI apps that are not code editors. So we're talking Codex, Open Code Desktop. I find the dance between the code and going back to the app kind of weird. And that's why I still prefer like a cursor or like a VS Code fork. Because that dance between the agent did something and no, I want to correct something. I want to see how it's done. I just feel a lot more comfortable in that space, at least at the moment right now.

• 5:08Scott Tolinski?

  Yeah. Yeah, I do think it's interesting. I think it's, as long as I can track what's going on and I'm able to monitor it. It also depends on the application. What am I building? Is this something I care about? Is this something I don't care about? Is this just, you know, quick and easy, get it done kind of stuff? So for me, that's just how it goes.

• 5:27Wes Bos?

  So let me ask you a question. You're working in Open Code Desktop and it did something for you. And you're like, that's not really what I want, but it's just, it's like a quick fix. Like it's faster for me to just to like go in there and code it. What's your process for that? Like that you need to open a cursor? A lot of people just jump into Vim because they're in the terminal, but you're in like a GUI. So what's your process for that?

• 5:49Scott Tolinski?

  If I want to go back and forth between my code and my AI coder using Open Code, using the GUI. The thing I do is I put the Open Code on one side. I put my Zed, my text editor on the other side and I'm following it. I'm watching the commits. I'm looking at diffs. I'm moving to files. If it's doing something I don't like and I can correct easily. It's so easy in that way for me to just tab over and make changes and then go back and forth with it like that. So that's how I do it and I find it to be decent. It's obviously not as nice as something that's built directly into the text editor, right? But I just like my tools the way they are.

• 6:32Scott Tolinski?

  And because I have them all set up and configured to my liking, it's a nicer experience for me to have those split app experiences here. So in addition to this, inside of editors, there's usually like a chat window. So VS Code has Copilot Chat. I believe you can install many other extensions to do an inline chat window. And this flow of working in the inline chat is going to be very similar to how what we're describing with any of these other terminal tools or even GUI tools where you're talking back and forth with the AI. Hey, can you make these changes? Can you do this? Can you do that? Reference these files, those types of things. But there's also tab completion where you're typing and it's like a fancy autocomplete.

• 7:14Scott Tolinski?

  So in the tab completion is usually for a little bit like the AI agents often are doing way more. They're going off, they're creating files, they're making big changes. Where the tab completion is usually for spot updates and like, I need to change this and edit this. And it is kind of predicting what I need or writing new ones. Some people prefer to completely stay in the realm of tab completions. But I have really moved away from doing tab completions as I go. And because I've gotten more effective with the chat based flow of things.

• 7:45Wes Bos?

  Yeah.

• 7:46Scott Tolinski?

  And then manually just editing things instead of relying on.

• 7:49Wes Bos?

  I agree. I was, I was the same way as well as as the like, the way that you puppeteer your chat as you get better as that, at that, and as the models get better and, oh, that's actually what I wanted. I find myself moving a little bit away from the tab completion, more into, into the agent. And then the tab completion is more just like, I need to fix this thing one by one. Or like, especially CSS. I find myself using tab completion in CSS so much because so much of the AI CSS that gets kicked out is, is awful. Or like, that's not what I want. That's what like, I'm just trying to like, design something that looks nice or feels good. Typing what I want into a box and then hoping it pops out.

• 8:33Wes Bos?

  And it just added a whole bunch of like hover opacities to everything is not it. So I use a tab completion tons in the CSS, more design, designing things. And then the more like logic based, flow based stuff, like writing backend JavaScript code, that's way more doing in the agents.

• 8:50Scott Tolinski?

  Yeah, I think about things like your MAD CSS website because it is so like. It's so deeply designed. You put craft into that design. That design, if you didn't have that done in Photoshop and just say, hey, go make this, I feel like that design would be impossible to dictate to in AI to do today, at least with my skill level. Can you, I can't imagine being like nailing that design aesthetic or being able to experiment with that and really nailing it. So I totally get that, especially because like the things that we're good at, we're good at. And I oftentimes am like, what are you doing here with this CSS move? You know, let me do it.

• 9:28Wes Bos?

  Yeah. Oh, man, that drives me nuts. It's true. Like even the last two sites I worked on, the Matt CSS one and the Sinhacks one, I was battling it to stop writing BEM selectors like nine times.

• 9:44Wes Bos?

  I eventually had to throw it into like the agents file because like it was searching for a way to make structured, repeatable CSS. And I don't think it understood like the scoping that I was using. And it just kept going back to making BEM selectors for everything and not using nesting.

• 10:01Scott Tolinski?

  Yeah. Yeah, there's so much there too, in terms of even like context when you get into, AI has to keep all this stuff in context. And still, like when you're dealing with big projects with reusable CSS and like the more you add there, the harder of a challenge it gets for the AI. So because of that, you know, dry CSS and dry things like that, I think are definitely still a challenge for it. And I wonder like what the exploration into that space looks like in terms of clearing some of that up. So that is the tools here. Again, you can use many of these tools just by installing them. Like Claude Code again is a CLI. Now, people really like Claude Code.

• 10:42Scott Tolinski?

  I do find it to be less featured than something like an Open Code CLI. And it is directly attached to using Anthropic and a Claude Code subscription. So for me, if I'm the type of person who is using some of these tools, I personally invest my time in learning tools where they're not tied to one specific subscription or model. Because the models change so rapidly these days that my tool should sit in front of the model choice. So I personally am not investing a lot of time into Claude code as a tool because I would much rather have a tool that lets me use what model I want to use, when I want to use it.

• 11:26Scott Tolinski?

  And therefore, that's why I use open code. But there's other great choices for this. Pi, I mean, we just had Mario on from Pi. Pi is a great choice for this. Have you seen Charm? Westwards, I believe it's not Charm. Charm is the overarching.

• 11:41Wes Bos?

  I've heard of it, have not used it.

• 11:43Scott Tolinski?

  Charm Crush. Charm is like a bunch of terminal utilities and Charm Crush is their AI CLI tool. It's beautiful. It's cool. It looks great. As far as two E's go, this is top tier stuff and it looks really super cool. It's got a lot of personality, I'll say that. But again, same kind of thing. I think it's a great choice to give this stuff a try. And folks, if we don't talk about your favorite tool in here, there's too many of these things. They're changing all the time.

• 12:09Wes Bos?

  Yeah, I agree with you as well. Like, being able to switch between models. Obviously, Anthropic's had the lead for so long, but OpenAI just came out with their new Codex model, which in my testing is actually pretty good. It actually beat it out. I was trying to make it, make some 3D printed gears that would mesh together. So there's quite a bit of math to that. And it did a fantastic job. But I want to be able to switch and I want to be able to like use somewhat standards based. A lot of things we're talking about today, like agents.md and skills and whatever, these things are starting to become like standards. They're all a little bit different, but the sort of whole ideas that you're using between all of these tools are starting to become the similar.

• 12:54Scott Tolinski?

  Yeah, I mean, I want my tools to be able to flip it and reverse it, you know? I want to be able to choose what I want with what I want and still have my tools adjust, right? It's just like text editors. Like, I would hate if my text editor was like, this text editor only allows React files. You know, I would not use that tool. So that, to me, is the distinction there. In terms of these tools. Okay, cool. So let's talk about the stuff, all of the stuff. I think that the most logical place to start with is the models. Now, we don't need to get into what AI models are specifically, but as far as coding goes, I personally believe the model that you choose will drastically change your experience with AI coding in 2026.

• 13:39Scott Tolinski?

  Because they all function very differently, even the best ones that are very good, still function very differently. And you need to interact with them in a different, different way, at least in my experience. So Wes, what model are you using to code with primarily in 2026?

• 13:57Wes Bos?

  Yeah, I've been on the like Opus models for a long time and I've been, right now I'm on Opus 4.6 for everything. And then I have been dipping a little bit into GPT-5.3 Codex. And then also every now and then I'll just try a random one, you know, try the Grok one and see how it is. And it's, it's really funny to see the like, the tells that they have of, of every single model. Like one, obviously the purple thing was, was really big for a long time and Anthropic and OpenAI have tried to beat the purple out of these models in their training. It still sort of ekes up, but now-

• 14:32Scott Tolinski?

  The bad gradients.

• 14:34Wes Bos?

  The bad gradients. Now you're seeing like in code, you're seeing the like, the comment with, where it does three line comments and it does a bunch of equal signs on the bottom and the top. That's like an Opus tell. And then like, I think another Anthropic one that's happening is it, it just like these, makes these little pill icons for everything with rounded corners on it. It loves to do like a hover and then slightly scale up and then, then make the opacity. There's all these like little tells. Every time I look at some code, I can almost tell which model it's been using.

• 15:06Scott Tolinski?

  Yeah, yeah, I know. It is very funny. The AI model that I'm using today to code is currently, I use Codex 5.3 for detailed JavaScript work. I do that because I find it to be way more precise. I find it to be way more reliable in its output. Now, it follies a little bit on Svelte 5 code, but using the Svelte 5 tools I have, like the Svelte 5 agent from the Open Code plugin that I use, those things, it doesn't, it's not a problem. Without that, it is a problem. But I find Codex to be more exact. And even though sometimes it requires me to yell at it to, "Hey, actually do the thing," instead of you telling me to do it.

• 15:50Scott Tolinski?

  Because sometimes I'll be like, "You should install this plugin." I'm like, "Well, you can do it. Why don't you do it?" And then I use Opus 4.6 for more exploratory work, more creative work, brainstorming work, those types of things, and I find it to be better at those tasks. I find it to be more creative, more, less exact in those types of ways. And because of that distinction, I am using both of those two models as my primaries. It just depends on the task at hand. And I know people out there get way more in-depth with that, where they, they know like, "I'm using this model for this, and this model for this, and this model for that." Like, that is too much for me personally.

• 16:30Wes Bos?

  There's also like benchmarks, tests, evals, all of this stuff that were like, you can actually get scores from it. Because I think a lot of this, like, you'll listen to me being like, "Ah, it's a little bit better on X, Y, and Z." A lot of this is just touchy-feely type of stuff.

• 16:44Scott Tolinski?

  Vibes.

• 16:45Wes Bos?

  Yeah, it's all, it's my vibes, yeah.

• 16:48Scott Tolinski?

  I think there's a lot of experimenting that needs to happen. But if you're out there and you, you haven't really tried AI coding, or you tried it last year and you found the results to be frustrating, I would say, give the latest Codex a spin, and I think there's a world of difference. I think it, it feels incredibly different than it used to.

• 17:08Wes Bos?

  Anthropic also released this fast mode. Did you see this, where it's two and a half times faster, and then it's six times the cost? Which is

• 17:18Scott Tolinski?

  That's crazy. None of that for me.

• 17:21Wes Bos?

  Yeah, well, like

• 17:22Scott Tolinski?

  I'd rather wait.

• 17:22Wes Bos?

  It's it's crazy to think like it can get two and a half times faster at at what cost, you know, like I I'm sure they're even losing money on that. They're probably sticking a whole bunch of GPUs on it. But kind of neat that it it might get that fast at some point. Once we can catch up with all this hardware.

• 17:40Scott Tolinski?

  Yeah, I know. It's already shocking. Even the speed right now, I oh, I will say back to the models thing. Codex, in my experience, even though it's gotten better as of late, it was way slower. So like when I first moved from Opus 4.6 to Codex, I was like, this sucks. This takes forever. I I'm not gonna come back after lunch to see the results for this. But in in reality, I I think the the results ended up being that much better. I had to go less back and forth with it. So the time it kind of evened out at the end of the day, but now they have made it faster.

• 18:12Wes Bos?

  A lot of the speed also has to do with the the tool that you're using, right? Like if you use the same model in in several different like earlier on, I'm not sure if it's still still the case, but Copilot in VS Code was significantly slower than Cursor with the exact same model. And it's just the way that these tools are doing things like like the tool still has to figure out what pieces to shove into context and and to figure out what what to go along with it, you know? Is it is it going to use the TypeScript server to grab all the types? Is it going to look at your clipboard, stuff that you recently deleted? All of that stuff is is sort of sent along with the with the request and to figure it out. And also, there's there's also like local models as part of this as well. Some of the tab completion stuff might not be using the model that you have selected.

• 18:57Wes Bos?

  It might just be have it happening locally.

• 18:59Scott Tolinski?

  Yeah, totally. Cool. So let's move off of models and into some of the language that you hear, one of which the words that you hear is agents. And to me, I know a lot of people, especially people who are used to using just ChatGPT or whatever, they hear the word agents and they think it's something special. But ultimately, the agent is basically the AI that's going off and changing the files and doing stuff for you. So if you're in a chat and you say, hey, go modify this file, that is the agent going in the agent. Is going and modifying the file. That's really all it is at a high level, right? Ultimately, though, agents can have different configurations to them. You can have custom agents that have default context embedded in them.

• 19:41Scott Tolinski?

  So therefore that agent is uniquely equipped to handle certain problems because it is brings in its own context. Agents can be scoped to tools sometimes, so that way you can like have an agent that only has access to some tools very specifically and not access to other tools. And then you can have an agents that are like scoped to specific models. So we mentioned that some people get really deep into using very specific models for specific tasks. And you could have an agent that is really good at one task. It's a scope to that specific model that's good at that task. And it is scoped to the tools that used to complete that task and has the context there as well.

• 20:22Wes Bos?

  You might have a code review agent. Anthropic in Claude Code has a web design agent where they specifically try to beat the purple out of the prompt. I think it says something like, don't use like overused AI things like gradients and overly, you know, like it tells it to do that stuff. So if you're, if you're constantly finding yourself telling it to do like, like that BEM example I had, right? I was like, stop using BEM, you know, like that's probably something that you should be sticking in your agents where you're telling it exactly what type of of coding you like to use, rather than just having to put that into your prompt every single time.

• 20:57Scott Tolinski?

  Yeah, yeah. And a good example of this, I've mentioned before is the Svelte open code plugin has a specialized Svelte file editor that again, has access to the Svelte MCP stuff. It is always looking only at Svelte, Svelte TS or Svelte.js files. It automatically knows to, you know, fetch documentation and validate the code. So it's been given specific instructions on how to interact with those files in the best way to get you the best results. And you might be out there and feeling like this is too much. Like, oh, I just tell the thing to do the thing and it does it. Once you get into some of these things, I do think that specialized agents. Do make an impact. Personally, I have found that this is something that makes an impact for me personally.

• 21:42Wes Bos?

  Claude Code has this thing called output styles and like, I've never used this. This is typically something I would put in the agents where it's like, don't yap, basically. When you're, when you're telling me what you did.

• 21:53Scott Tolinski?

  Yeah.

• 21:54Wes Bos?

  Yeah. You make me like this, like whole book report. I ain't reading that, you know, like, and like, you don't have to be nice. You don't have to use a whole bunch of flowery words. Like, just, like you said, keep it terse.

• 22:04Scott Tolinski?

  Yeah, I have mine, mine, mine, especially I like that about my Claude Bot, my Open Claw setup was that like, when you're creating it, it asks you to like, how would you like me to, to respond to you? And I said, like bulleted lists.

• 22:16Wes Bos?

  Yeah.

• 22:17Scott Tolinski?

  You know, keep, keep it minimal language to express your, your topics. I don't need any extras and stuff like that. And I do think that that makes a big impact for sure, because man, I hate it when it gives me a giant paragraph. Like, just tell me what you did or didn't do and what needs to happen next.

• 22:32Wes Bos?

  All right. Next up, we have subagents, which is, oh, and honestly, I haven't used subagents a whole lot. But from what I understand, and maybe Scott will, will tell us what he uses them for, but like, basically you have parallel processes. You want, you want some specific agent to go off and do one thing and sort of come back to you, delegate one specific part of it. Then you can, you're basically puppeteering a whole bunch of stuff. So what are you using for it?

• 22:56Scott Tolinski?

  Yeah. And that, that's really it. So like the, the agent is the one that you're talking to, right? It's the one that you're going back and forth with. And the subagent is oftentimes when you need to spawn off a process to go and do a task. Like, hey, go and review this Svelte file. It is going to spawn the Svelte subagent to go and work on it. And that's going to have its own thing and it's going to go off and do that. Meanwhile, I can continue interacting with the main agent to do other things. And then tell it to go spawn off other subagents. So that's the big thing there. And, and subagents, in the same way as agents can be tuned, you're giving them default context. You're potentially giving them, you know, scoped tool access, or you're giving them specific models.

• 23:39Scott Tolinski?

  Or again, MCP instructions and stuff like that.

• 23:42Wes Bos?

  Yeah. So when you are defining an, an agent and a subagent, those are, those are essentially the same thing?

• 23:48Scott Tolinski?

  In many ways, they are. Yeah. Yeah.

• 23:50Wes Bos?

  So like when you create an agent, you don't then decide whether it's a subagent or an agent?

• 23:55Scott Tolinski?

  Oh, no, sorry. You do.

• 23:56Wes Bos?

  Just an agent. The fact that it is sub is because it just goes off and does it for you while you continue on?

• 24:01Scott Tolinski?

  No, my understanding in different platforms, they treat this differently. But in in open code specifically, you define if something is a sub agent or not. Okay. Like in many of these chats, they give you access to like have a dropdown of which agent you're talking to. I'm talking to the build agent, the plan agent, the spell agent, whatever. Those are your agents. And then the sub agents do not show up in that list. They're not something you're going back and forth with. They're going just specifically. So you are defining them as sub agents specifically. And if you want to see all of the errors in your application, you'll want to check out Sentry at sentry.io/syntax. You don't want a production application out there that, well, you have no visibility into in case something is blowing up and you might not even know it.

• 24:45Scott Tolinski?

  So head on over to sentry.io/syntax. Again, we've been using this tool for a long time and it totally rules. All right. So the first file that many people are introduced to when working in AI is the thing that's usually created when you initialize a project. This is often now referred to as the agents.md file, which is just a markdown file that is going to prime the AI with context in every new session. So this file, I'm actually so annoyed that nobody's come up with the standard yet. It seems like we're kind of standardizing on agents.md. But like, there's cursor rules, Kiro has its own thing, Claude.md, Copilot Instructions.md.

• 25:29Scott Tolinski?

  And like, some platforms allow you to customize and configure what these files are and where they live. If we could just all standardize on one, that would be awesome. But again, this is a dedicated place to put context that should be there in every session, no matter which agent you're using, no matter what you're doing. This is like, this is a TypeScript-only codebase. This uses Svelte 5. Like, this is a Svelte Kit app. Whatever. The stuff that every single session and anything should have at the top of it, this is it. Now, the one thing that I think, when this first came out, this was like the only tool that you had in your tool belt to like steer your codebase. Yeah. And so because of that, people just threw the entire kitchen sink into here.

• 26:14Scott Tolinski?

  I think that the more you put in here, you know, again, that is additional context that is being Set in every single chat. And we've all had the experience that the longer you go back and forth with the chat, the worse it gets because that context just keeps getting bigger. So you're probably going to want to put just the stuff that you need in here for it to be effective at its job and like that's it. You're not going to want to put like just, you know, 10,000 lines of text in here about your project because it's probably going to have the opposite effect for you.

• 26:47Wes Bos?

  All right, so let's talk about skills then. And like, quite honestly, a lot of this stuff overlaps each other and you might not be sure where to put it. But another idea that we have here and probably one of the more popular ones aside from that, like agents dot M D is, is giving your agent more skills. And these skills are, the skills are a set of like capabilities that will extend the functionality of the, the AI tool that you have. So this is where you can teach it how to do something specific. And before we get into like, like what you can make it do, I think one of the, the key things here is that the AI won't necessarily reach for a skill unless it needs it.

• 27:32Wes Bos?

  And that means it's not stuffing all of the stuff into context unless it absolutely needs it. So it's, it's able to figure out what skills it has. And then if it thinks that one of those is, is helpful, it will sort of bring that into context and, and give it a little bit more information. I find that to be really helpful as well because the like context stuffing, giving it too much can certainly, it's, it's expensive and makes it slow. And then it, once you have too much stuff, then everything gets kind of muddied.

• 27:59Scott Tolinski?

  Yeah, yeah, totally. And that's a, that's a problem, I think, that people experience when they're getting into this and, and normal, normies, normal people experience that context issue. Again, like I said, and they're just using chat GPT. My wife was just lamenting about Gemini being like, I've just been going back and forth with it and it's doing so many like weird things in these ways. I'm like, you gotta start a new chat. Like, uh, the context is, is shot at this point. Yeah. So skills to me was one thing that I didn't, I had a hard time like kind of grokking what skills were useful for, but it's really just a way to like.

• 28:37Scott Tolinski?

  Instruct the AI how to do a thing. I got hung up on like, oh, does this be along in a slash command or a skill? Or is this like better suited for an agent? And we can talk about some of that. But ultimately, like, skills are a way to define your set of instructions or best practices for doing one specific task. And I think if it's something that you go back and forth with and on this task, maybe that's more appropriate for like an agent definition. I'll talk more about that later. But have you dove really deep into any specific skills? Or is there any skills that you're liking?

• 29:12Wes Bos?

  The skills is something that I have been using more than the agents. And quite honestly, I don't know the difference between an agent that is preloaded with all of this stuff and just a whole bunch of skills that you can load up into a project. Because so Versal has this website, skills.sh, and essentially, you can use their little CLI and it just downloads a markdown file in into your project and puts it there. And then, and then your AI will pick up that markdown file, right? Just a set of instructions of how to do things. So I was like going through it. And the one of the most popular ones was like a remotion one, like, here's how to use remotion to make a video with code, right? That absolutely blew up. And everybody started using

• 29:56Wes Bos?

  it. And it also will, it's not just one markdown, it's, it can be multiple markdown files, meaning that like, you don't have to stuff everything in context. But if it thinks, oh, I need the docs for 3D, or I need to know how to do something for 3D, then it will download, go ahead and download the markdown for that one. So one skill that I used, and this is something that everybody was saying as like, the best skill ever is this thing called superpowers. And this is a GitHub repo that you can download and put it in. It's, it's, it's, the superpowers is a complete software development workflow for your coding agent, built on top of a set of composable skills, and some initial instructions that make sure your agent uses them. So I installed this thing,

• 30:37Wes Bos?

  and essentially, it gives it the workflow of building out an app. It brainstorm, it uses Git worktrees, it writes a plan. It queues up all your subagents. You've got test-driven development. It does code review and then finishes the development branch. And it says like, anytime you're gonna go and work on something, follow this very rigid workflow. And here are all the skills that you have to do that. So I, I installed it and I loaded it up with Claude Code. You can like, you can install it with, with Claude Code or any of these, these tools. And I said like, go and make me a website. I, I gave it like a, a word document with all of the details about our cottage. And I was like,

• 31:20Wes Bos?

  make a website for our cottage, you know, because every now and then we let a friend stay at the cottage and they want to know all the information, you know, like, what do you do with the water? And like, where, where are the keys? And all of this information. I was like, it would be nice to have it as a website with the, with the images and everything. So I gave it that document. I gave it all the images and then I let it, I let it rip. And it worked for three and a half hours. Made a plan, TDD, get work trees. Three and a half hours later, it kicked out a two-page HTML website. And it was $26 in anthropic thing. And I was like, this is not good.

• 32:05Scott Tolinski?

  This is not good.

• 32:05Wes Bos?

  And it feels like you just did a whole bunch of nothing work.

• 32:09Scott Tolinski?

  You're pushing text around.

• 32:11Wes Bos?

  Yeah, well, they just love the structure and they love all the, like, the process of everything. And then the output of your beautiful process sucks. So I spent, I let this thing run forever. 26 bucks later, a very mediocre website that I probably could have, um, chat GPT'd in, in three or four minutes, or like any of these other things. So I think a lot of this, like, people getting obsessed over getting the perfect AI thing dialed in, you know, all your different agents and subagents and skills. I think they're really handy and I use them quite a bit, but I think you can certainly go overboard and end, end up costing yourself in money and time.

• 32:57Scott Tolinski?

  Yeah, totally. Especially money and time, that's for sure. And I find the same thing. The 80/20 is like largely that like that 80% of it you can get with very minimal amount of tools. Like again, like if I'm working in Svelte files, I need that Svelte thing so that the output is good. So like for me, just straight up YOLOing it without that is a less effective thing. But the moment that you're adding too much, again, you're just like passing compute around. And it is, it is like, what are we doing here? What is going on? So I haven't really also gotten into skills that much either. Well, I may not either.

• 33:32Wes Bos?

  Well, I have. That's the thing. I haven't gotten much into the agents. I'm into the skills.

• 33:37Scott Tolinski?

  Yeah, I've gotten more into the agents than I have into the skills personally. But it is something that I'm adding more and more here, but I'm doing so in a more minimal kind of cautious way, rather than like, ooh, let me head to skills.sh and install all these ones. Ooh, good. Okay, this one looks like my planning one, my TDD one, all that stuff. So yeah, I have not gotten that deep into them. But again, a skill is primarily useful when you want to accomplish a specific task. Another thing to accomplish specific tasks would be something like a slash command. And the slash command is something that you use manually. So like a skill is something that I would say, hey, AI, go work on this. And the AI is able to determine and use the skills.

• 34:18Scott Tolinski?

  You could tell it to use the skill specifically, but like it's not like calling a function, right? Where a slash command is more like calling a function, even to the point where open code slash commands have arguments in them and can have arguments. And those essentially are running a dedicated single prompt that arguments or whatever, that is essentially a snippet. It's the equivalent of copying and pasting a prompt that can have arguments. It's a calling a function. So that's what a slash command is.

• 34:48Wes Bos?

  I'm a big fan of these slash commands because like I often will have a whole package.json full of all of these different scripts that I would very rigidly run. And it's nice to have have a slash command that doesn't have to be as rigid. It can just simply be natural language prompt where you can do things like add a new page or add a new route, scaffold this this thing out, lint, run the tests. And you can just put these reusable prompts under. These nice quick slash commands and then as you're using the UI, you just hit your slash command and it will auto-filter them. I'm a big fan of that, like probably looking at all of these things today. I would probably say skills slash commands and couple MCP plugins are probably my most heavily used.

• 35:33Scott Tolinski?

  I'm kind of stoked to hear you say that because I have a bunch of my slash commands tied directly to my stream deck and I also really like using slash commands. And when all this stuff comes out, I keep on just wondering like, ah, this feels like I have a better time doing this with the slash command, but at the same time, like, nobody talks about them. And there's no hype around slash commands. There's all this hype around skills and whatever. So I do like slash commands myself again, because they feel a little bit more like I'm in control and I like that.

• 36:02Wes Bos?

  Next thing we have is hooks. This is something that Claude Code has. I know Open Code has a similar idea in their tools or plugins, sorry, not tools, via plugins. Essentially, this allows you to run some code on a specific event, you know, like save a file before you do something, before you make a change, after you make a change. And this is really helpful for running linting or formater, TypeScript checking. A lot of these tools will simply just try to run TSC after it's done, a specific prompt for you. But if you're trying to keep your code quality in check, having these extremely rigid tools that will throw errors is super important.

• 36:44Wes Bos?

  Otherwise, it just goes off the rails and starts making a whole bunch of slop.

• 36:49Scott Tolinski?

  Yeah, yeah, totally. And I first really got into hooks with Kiro because their interface for it was really nice. Like you can say, alright, the package.json updated, now update my readme. You know, it's essentially things like that. And when I was at the VS Code Insider Summit, that's one of the things I told them I wanted the most was a way to do hooks. They mentioned it was like a really tough problem to solve, but I do think that hooks are greatly underutilized, but at the same time, since not using Kiro, I haven't been like going nuts to re-implement them in anything.

• 37:22Wes Bos?

  I'm just looking at the list of events from Claude Code, you know, session start, user prompt submit, pre-tool use, permission request, post-tool use, post-tool use failure, notification, subagent start, subagent stop. So like, If you're building out your own little UI for these types of things, you're probably going to want to listen to all of these different events. But I think the regular user is probably not building out there, building out these things. It's the same thing with like a lot of our code quality tools as well. A lot of us are simply just users of them. And then in people that build Oxlint and ESLint or whatever, they're the ones that figure this stuff out for us.

• 38:01Scott Tolinski?

  Yeah, let's talk about plugins. Now, a lot of things have the idea of having plugins. Specifically, I know Open Code has plugins. Plugins in Open Code can do a number of things where they're adding like, one of them I've seen, I asked on Twitter the other day, actually, it was like, what are you even using plugins for? Because I'm not using plugins that much. And people are saying they use like a super memory plugin that gives them access to a specific type of memory. There's a handful of plugins to make Open Code like use Claude code conventions in various ways and things like that. The only plugin I use personally is the Svelte one that I mentioned. There's a Svelte Open Code plugin and all that's doing is it's adding the MCP, the MCP tools, and it's adding the subagents.

• 38:46Scott Tolinski?

  It's like basically wrapping up a bunch of the features that we've talked about in this episode into one package. And that's it.

• 38:53Wes Bos?

  Yeah, it's shareable. You can wrap it up. You can put them like with Claude Code, the plugins can contain skills, agents, hooks, MCP servers, all of that type of stuff. And you may want to share that with people that work on your team. Like you can share it in your Git repo. But if you've, if you've got multiples, right? Or if you're starting a brand new project or you're doing anything, you may want to just find a way to bundle this on up. ESLint can be plugins, but they can also contain rules and all kinds of stuff like that. So very similar. And then finally, we have MCP. MCP, we've talked about it many times on the show, but an MCP is a way to programmatically connect your AI

• 39:35Wes Bos?

  to interact with external apps and external services. So the way that it works is that you, you put an MCP server in your tool that you're using, and then it will call that MCP server and figure out what are the tools that are available to me? And that might have some like- A very common one is context seven, just looking up docs, the ability to search for and download documentation for specific libraries. You can bring that in. Another MCP would be using playwright, the ability to use an external browser and sort of puppeteer it for you. That's different, sometimes is different than than skills and agents, you know, there is some overlap between the two.

• 40:20Scott Tolinski?

  I think that's what's confusing about all this stuff is the overlap between all of the different pieces.

• 40:25Wes Bos?

  I think it's because everyone's just trying to figure out like what this looks like. And then the next time like one of these new models comes out, they may just be trained, some of these things might just be trained away, meaning like you don't need X, Y, and Z anymore because it simply is is way better at figuring out what it is that you wanted. But like, like even like Sentry, you know, you could Sentry has an MCP server and you can hook it up to that and say, hey, I'm working on bug X, Y, and Z, or check my, check my Sentry. That's one thing we didn't really talk about is this idea of like cloud agents where I was laughing the other day because I was running something on my computer and I had to go bring my daughter

• 41:05Wes Bos?

  to ballet. So I tethered my phone and I left it open on the passenger seat thing while I was driving. I was like, if this was a cloud agent, then I could have closed my laptop and then opened it up when I got there, right?

• 41:18Scott Tolinski?

  See my tail scale video, folks. I did a video on tail scale. I'm running open code on the Mac mini on my dev process. If I on my computer tell open code to work on a problem, I can shut my laptop, I can do whatever, I can pick it up on my phone later and then finish it or see what's going on. And it's just going to be just cranking away on that thing. So, man, it rules.

• 41:43Wes Bos?

  And a lot of these, these things will have or do have the ability to have cloud agents, right? Like, I know Cursor has it for sure. Oh, does open code? That's probably their like monetization platform, I'm going to guess, is the way that they're going to make money, you know, like open code is a business, but their tool is open source.

• 42:04Scott Tolinski?

  They have a subscription service for models in compute specifically. So if you wanted to say like, I want to subscribe to all of the models you would. It's called Open Code Black, maybe is the new one. They've changed it. But the being able to serve it over the web, one line command, Open Code Web or Open Code Serve. Bingo, bango, you got it on your IP. If you're on a tail scale network, you just log into that and you're there.

• 42:29Wes Bos?

  I don't think that most people are going to do that, though. I think that most people.

• 42:34Scott Tolinski?

  Man.

• 42:34Wes Bos?

  I don't think so.

• 42:35Scott Tolinski?

  Developers are lazy.

• 42:37Wes Bos?

  I will put money on it that Open Code is going to roll out Open Code Cloud.

• 42:41Scott Tolinski?

  Sure.

• 42:41Wes Bos?

  Where you can just run it in their sandbox on their, their cloud. And, and that will be great, right?

• 42:50Scott Tolinski?

  Not me.

• 42:51Wes Bos?

  One thing like Sentry can do is like I hooked it up to my cursor cloud. And then like if there is, is a problem, I can just like sick cursor cloud on it right from my phone. And then it will submit a pull request and you can merge it. That's, that's a really neat workflow. And I think that I bet they will be going that way. I bet that's a big, like, I don't know that a lot of these big model providers are going to be making a ton of money on these models because eventually we're going to get a lot of these open source models that come out that are good enough or just as good. And then all of a sudden someone says, well, either I host it myself because I have some like $10,000 supercomputer

• 43:33Wes Bos?

  or like the, I think, I think the like profit that they're making on these models is going to be like razor thin. And there's going to need to be other things like paying for whatever cloud at the end of the day.

• 43:47Scott Tolinski?

  So let's go back real quick to some of this because we were talking just about like the confusion and overlap. And one of the questions I posed was like, is this better as an agent or a skill? Or what is this better as? I do want to just like touch on that because I did a little bit of, of diving into that for my own personal ideas here. So like the, here's, here's a question that I had personally. Is it better to have an accessibility agent for completing accessibility tasks? Or an accessibility skill that any agent can use? Or is it better to have an accessibility skill that an accessibility agent can use? So, skill, agent, or agent plus skill? Like, what is the ideal combination there?

• 44:30Scott Tolinski?

  And the answer that I personally landed on is it basically depends on what you're doing. Like, if this is a one-off thing and you're saying, "I wanna, hey, go fix accessibility issues," then an accessibility skill is more than appropriate for that. And that, for the most part, that is the answer. The answer is, "Hey, just go off and fix these issues." But if this is a longer-running thing where you're like, gonna go back and forth with it, like, "Hey, I want you to go and do a full audit. I wanna then talk about that audit. I wanna then go maybe potentially fix things based on that audit. Or I wanna dive into that more, or go back and forth." Then you're in the land of like, "Oh, this should be an agent, right? That is taking care of this." So it depends on what you want there.

• 45:15Scott Tolinski?

  And then, that third thing of an ally skill and an ally agent, that actually could work, and it could work well, where the skill might be, "Run these specific tools." And then the agent could have instructions to run those specific tools as a part of your process, or as a part of its fixing process. But that does feel like you're getting into potentially overkill territory, but it really depends on what you're trying to do. So, if you're feeling like there is overlap here, again, specifically for me, skills was always one of those things I just had a hard time, like, really latching on. "I, but I have an agent for this. Why would I need a skill?"

• 45:52Wes Bos?

  Yeah.

• 45:52Scott Tolinski?

  And that's really it. If you're needing to go back and forth, agent makes sense. If it's a one-off thing, a skill makes sense. If the agent can use a skill, like, that is also something that absolutely can work. At the end of the day, experiment with this stuff, and I think that is a decent answer.

• 46:08Wes Bos?

  Yeah. The decent answer is, nobody totally knows.

• 46:12Scott Tolinski?

  No, I know, I know everything.

• 46:14Wes Bos?

  And we're all figuring it out, and what makes sense. And I bet, we come back to this in six months, we'll have a bit more of a solid answer as to how these things can work. And maybe even a lot of these things will be distilled down into some sort of smaller... Piece of information that you can just load in. You don't have to worry about whether it's a hook, a plugin, a skill, etc.

• 46:38Scott Tolinski?

  You can be effective with these tools without having to go down all of these wells, but it is, honestly, it's kind of fun to experiment with. So I've been, I've been having a good time like exploring this stuff and it makes me feel like excited for the tech. I know people get down on AI or aggressive towards AI. We see it in the comments and stuff like that, but these are technical, interesting computer science problems. And I do think that there's, they're fun to work on.

• 47:03Wes Bos?

  The same thing with like all the, the Claude bot stuff, like everyone's mad about, it's hype and whatever. And of course, of course, it's stupid. Of course, people are losing all kinds of money on it. And of course it is very hyped up, but all of that is that people are figuring a lot of stuff out. They're figuring out what works. They're figuring out how to actually interact. They're figuring out security and like this Claude bot or what's it called? Open Claw now? Like we might not have this in, in a year, but I guarantee you the lessons that we have learned from this whole thing are going to be put into all of the products that we're, we're using today.

• 47:41Scott Tolinski?

  Heck yes. Cool. Well, let's get into the part about sick picks and shameless plugs. This is stuff that we're liking, enjoying, products that we like, podcasts that we like, YouTube channels, whatever we're feeling at this given point in time. Wes, is there anything in your life that has been nice for you?

• 47:57Wes Bos?

  I got one. Real quick. TV show, St. Denis Medical. If you like, like The Office and like Parks and Rec, my first recommendation to everybody is if they like that type of stuff is to watch The Paper because it's literally half of The Office characters and it's, I think it's made by the same people. It's fantastic. But if you also like that sort of vibe of The Office, St. Denis Medical is a very good show. Me and my wife love it. They just released the second season and we've been thoroughly enjoying it.

• 48:26Scott Tolinski?

  Nice. I've been out on, I watched The Office, but I didn't get into Parks and Rec. I have noticed that like every single show on TV nowadays is like, what if we did The Office but for, but for a grocery store? What if we did The Office but for cheerleading? Or whatever?

• 48:40Wes Bos?

  This is The Office but for hospitals. So, yeah.

• 48:45Scott Tolinski?

  They got a formula. Yeah, they got a formula.

• 48:46Wes Bos?

  Yeah, Parks and Rec is by far the best. The paper also, I would say, maybe even better than The Office.

• 48:53Scott Tolinski?

  Interesting.

• 48:54Wes Bos?

  Is that religion, maybe? But very good.

• 48:56Scott Tolinski?

  Interesting. Okay, cool. I'll check it. I have a sick pick is a smart lock. I hate smart locks. I have tried so many of them. I had an Akara one that my, I had to have a locksmith like destroy it and that locksmith was like, I am furious at this company for designing this lock this way. So I, I had an Ultraloq on, on one of my doors that I really like. And I, I since then, for some reason, have for my office door have like gone through different locks. The Akara one, I had a level lock. Man, the level lock sucks. I hate it. And so I was finally like, I'm so sick and tired of dealing with this thing and I bit the bullet.

• 49:35Scott Tolinski?

  I was like, you know what? I'm just gonna buy the latest Ultraloq, whatever it is. And so I got the latest fingerprint Wi-Fi. Ultraloq Bolt fingerprint Wi-Fi smart lock works with Apple Home, Siri, Alexa, keyless entry with fingerprint ID, app, remote control, keypad, deadbolt, BHMA cert, easy install, IP 65 black. I got that one and it's nice. It's great. For smart locks go, it's 130 bucks or whatever. It's got fingerprint, it's got number pad. I don't have to pull out my phone to use it. I do doo boo boo or I fingerprint it and it works and it's reliable. It has auto lock. I can connect it to Home Assistant.

• 50:13Wes Bos?

  Oh, that's great.

• 50:14Scott Tolinski?

  So Ultraloq, shout out to Ultraloq. I, I should have never left you. I should never have doubted you and just bought you for the first place on my office door.

• 50:21Wes Bos?

  I've been wanting to get one of these forever, but our front door is like a weird, has a lock and handle.

• 50:29Scott Tolinski?

  Ours is too.

• 50:30Wes Bos?

  All in one. Yeah.

• 50:31Scott Tolinski?

  That's why I don't have one there. Yeah. No, no, I don't have one there. This is for my office. And then we have like a back door. Cause yeah, there's a back door.

• 50:38Wes Bos?

  Yeah, we have one on our, our back door as well, but it always runs out of battery. It like, you need to put, you gotta put double As in it, like probably every 45 days. And because of that, we just never replace the batteries on it and it just like stops working. So I need to get something that is better than that.

• 50:56Scott Tolinski?

  Yeah, I would say that's what I hated about the level lock. You had one of those little batteries, just a single one of them and it died like every 14 days. And I was ready to like, when it died, I was ready to lose my mind. This one takes eight. AA batteries and it supposedly will last for many, many, many months, like six plus months on a battery set. And I will say from personal experience on the other one, we like never change the batteries on that one. And it's only four batteries. So yeah, it is very effective. Yeah.

• 51:24Wes Bos?

  I've been looking at like hacking mine to put like a lithium battery in it. Somebody's done it because that would last probably four or five months, which would be way longer. There's also like wireless charging locks now where you put this thing like 10 feet away and it somehow puts a little bit of power through the air.

• 51:41Scott Tolinski?

  That freaks me out.

• 51:43Wes Bos?

  Wild, but yeah.

• 51:44Scott Tolinski?

  I don't want that. I don't want laser. I don't want to stand in front of it and have it fry my balls or something.

• 51:51Wes Bos?

  Yes. All right. That's all we have today. Let us know what your skills and agents and tools and all of your things are down below that you're using. Would love to hear it. Peace.

• 52:02Scott Tolinski?

  Peace.