Why Your AI Assistant Still Feels Generic
Vibing with AI Code, Episode 16: Brendan & Baird

Brendan: Hey, how's it going?

Baird: Goddag, Dan

Brendan: Yeah, not too bad. Busy with things as always. I feel like that's always the intro when we start

Baird: Be weirder if you weren't busy, I guess

Brendan: Yeah. Yep. I mean, that, but that's why we've got AI helping us out, right, to and solve that. And interestingly enough, like I had a conversation today about the command center thing that I built, but just

Baird: Yeah

Brendan: kinda… In a lot of ways it was just more of a generic like how can AI help but also like not more problems? i- in a lot of ways, like as I was talking, I was just sitting there thinking like my, my command center thing is tailored to me very much so at this point, is, the concept is still like very gen- general, like generic. Like the, the concept itself, like anybody could take and run with. The only difference is that, you know, it's got months or six, I don't even know how many

Baird: Yeah. Yep

Brendan: of data of mine now. And what I, I ended up talking about, and I think you've got quite a bit to say based on pre-show this sort of thing too, is like th- there's something missing it feels like with just having a generic bot, even if it does have all my data. 'Cause like for instance, I still struggle to get it to help me in ways that, say, like a secretary could. Meaning, if I wanted it to give me like a, an actual to-do list for the day, it fails almost every time. it might get one thing, but it's gonna have multiple things I don't want, in some cases it's just like this huge list that is way too much for anybody to use, and that becomes like a job of its own,

Baird: Oh, for sure

Brendan: to manage that sort of thing.

Baird: Yep

Brendan: And I, I mean, I was thinking about it from the standpoint of like, okay, what… And this was during a conversation, so I didn't solve it, but it was kinda like… It was almost like I landed on the concept of, okay, there's something there, like this is too generic. There's gotta be some way to tailor it even more than just data. I, I, I don't really have an answer there, but I think you might have some thoughts or maybe some stuff you've been doing recently that, you know, may- maybe is moving towards that direction a little bit more. So I'm kinda curious what your thoughts are.

Baird: Yeah, I, I mean, it's… I've had the s- same thoughts and feelings. Like, there's been so many, you know, you had Grok Bot come out, you've had OpenClaw, you've had, like, all of these things that might be able to get there. But every time I've played with any of it, it's like It just doesn't feel personal enough. And I know it's like, oh Yeah you could, like, OpenClaw is supposed to learn as it goes and, like, get better over time. And I've seen a lot of people have gone away from OpenClaw now with things like Grok Bot and Meta Muse and Instinct coming out. It's like, oh, these ones are now amazing 'cause they're simpler and they've got a little gimmick here and all this type of stuff. And, like, I've run into the same issues for so long of just, like, trying to use AI in general. Of it's really, really good at some things, but then it's really, really bad at, like, knowing what's important. You know? Like act, like, you can try to teach it, and you can try to, like, give it more data and stuff like that. And I think after a while, part of the problem is that it's actually, like, too much data, if that makes sense. It's like, and, like, you still need the context, but it's like the layer that I think is missing is, like, how do you curate that data? Like, how do you teach your AI to think about that data and actually, like, work with that data? Because what I figured out is, like, the more you give AI without really good instruction and really, like, thoughtful intent, it'll give you something back, but it's usually, to your point. Like, I've, you know, had multiple versions of agent assistants that are like, "Here's your to-do list," and it's like 35 things long. And I'm like, this is not helpful. And it's, like, my morning digest in Slack, I get, and it's like 35 things, and it's like there's no priority. There's no actual, like, thought behind how this stuff actually makes sense. And so, you know, I keep coming back to, like, building my own agent and, like, not even, like, a in-Claude Code or anything like that, but it's like, you know, I've been interested in Eve, Vercel's agent framework, and just their kind of ecosystem. And it's so easy now to essentially give Claude your, you know, like, the, essentially, like, the documentation through their skills and whatnot, and then just have it start working on it. And, like, at work, there was a thing going on where it was like I have a project manager that's helping out right now, and we were debating, like, is it worth having, you know, paying somebody to do that job, or can we figure out something else? And so,

Brendan: Mm-hmm

Baird: in a day, I thought through, like, what do I actually need? Like, what's the problem that I'm trying to solve now? And it's like, where I work, I'm literally, like, the only technical person, and the problem is, how do I get the stuff out of my brain and into a format that non-technical people can actually, like, understand what's in process, what still has to come? Like, what am I actually doing without me having to send a bunch of Slack messages or respond to a bunch of Slack messages?

Brendan: Yep

Baird: And so, like, literally in a day- I was like, "Okay, what do I need?" So I plugged into Granola to pull in my meeting transcripts. I plugged into my Gmail, Slack, I, my Google Docs, and then I also plugged in, like… And the cool thing is it's all reactive, too. Like, it's not like it's just sitting there polling. It actually, like, reacts to when things happen, and then it actually goes out and figures it out. But one of the coolest things was plugging it into, then, Claude Code itself and having hooks. So, like, as session turns end, it actually will send off that context to my agent to piece together what am I working on, you know, like, what are the things that are sitting out here from all my other context sources, and what am I actually working on? And it can take those and start to figure out, like, in a nice UI, where it shows me statuses, like, throughout the day, where it's like, what's on target? What is at risk? What, you know, like, showing me all this context in a really nice UI that's, like, very simple. It's like, the projects I'm working on, what the status of things are. And then I can go in and I can see, like, what are all the sessions that I've seen? Like, what's, you know, like, the summary of where it's at. And, you know, this was the first time where I'm like, man, like, this actually is coming close to giving me something that is extremely useful. And, you know, we can talk about it a little bit more, but, like, you and I started playing a little bit with, like, Instinct recently. And I started to kind of see what I could push Instinct to do from that perspective, and it was like, there's still that missing piece. Like, I couldn't quite figure out how I could make it so that it would fill that kind of personalized gap. Like, there's some cool things about Instinct, but it's like, again, a little more general, where what I needed was something that was, like, tailor-made. A- and this goes back to, you know, the harness. Like, this is a lot of what it is, is like, there is a harness around what I needed to do that I then built up to kinda hit that.

Brendan: Yeah. Right. And j- just to touch on Instinct really quick, like one that had too is like I, I started the onboarding, which made sense. It's gotta ask questions, it's gotta get to know you. And, and almost immediately though I was just like, what is this gonna gain me that what I currently have doesn't get?" Like, I've already got all my projects tracked in my Claude

Baird: Yep

Brendan: all my personal stuff there. Really the only thing that I found, and I, I even asked the Instinct agent this question. I was like, "Okay, I'm kind of already doing this, so like, what can you give me?" And it- proactivity was the biggest thing. Like, it's gonna keep hounding me and stuff. I could build that in. I have done that before

Baird: Sure

Brendan: my work Claude stuff. I mean, there is something… And, and I guess the hard part would be like, for me it would be kinda nice if I could just say, "Hey, Instinct, can you just like take over for the Claude Command Center maybe?" Like, I, like, I don't actually wanna do that necessarily, but it's like that was the only thing I could think of is like I don't wanna redo all this. I don't wanna go through

Baird: Yeah, for sure

Brendan: all the things that my Claude Command Center knows. by the way, I really just need to name that something better, 'cause at this point it's too much of a mouthful,

Baird: CCC.

Brendan: Yeah.

Baird: C cubed

Brendan: Right. A- I mean, the, the other thing that I was gonna say that Instinct offers but I, I've kinda already got now with Claude Code is, you know, being able to interact with it on my phone I'm away from my computer.

Baird: Yep

Brendan: what I've been doing to do that is I just leave my computer open and I use remote control, so I mostly get that still. it, it is… A- and it did feel generic too. That was the other thing. Like, as I was onboarding just like- I, I feel like there needs to be something tailored or you're gonna have to build it, like you said. A- and that, that comes back to the work stuff, where it was like I was having that conversation, and it just feels like everything is so generic unless it knows more than just the context of what you're working on. But I guess one example would be, like, having your roadmap or something like that, that you can basically give it, like, bigger goals than just, like, daily to-dos. if it could, you know, maybe know about the, say, like, three things, 'cause honestly, anything more than that is too much.

Baird: Yep.

Brendan: two's enough.

Baird: For sure

Brendan: being, like, that's something I've, I've sorta tried, and I think I just need to get better at giving it the right,

Baird: Yeah

Brendan: or whatever. But I, I don't know. That- that's just, like, another thing that I, I, I really felt starting with Instinct was like I don't… this doesn't seem appealing to me. Like, maybe three months ago before even Remote Control existed, or at least I started using it, like, maybe it would've been appealing.

I, I don't know. It just… don't even know what Instinct's audience is, and I haven't used Grok Bot. I haven't used Muse. So, like, I don't know how different they are besides, like,

Baird: I think they're pretty similar.

Brendan: But

Baird: like, like what I've seen is it bounces from like, everybody loved Grok Bot when it came out because it was a little more like there's some personality to it, you know? And like, it was, and it was also simple. I, I think that's the biggest thing when you look at mass adoption is, is it simple? Can I just jump in with no technical knowledge and get something to do something? A- and,

Brendan: too with Instinct as well. It's like this- all… Well, again, I haven't used Meta Muse,

Baird: Yeah

Brendan: Grok Bot and, and I'm assuming Meta Muse the same way, and Instinct, and others maybe. To me, it's like OpenClaw or even the Claude Cl- Claude Code setup that I've got like, running Linux. are, like, Windows or, you know, macOS or iOS even, or something like that where it's, uh, you know, f- the fine edges are rough, sanded down or whatever. um, uh, Jesse Hanley actually put this maybe a better way. He, like, he called it the play school. Like, it's just,

Baird: Sure

Brendan: super simple, like, blocks you put together, like, instead of having to, like, build the entire thing yourself. Like,

Baird: No, I

Brendan: so

Baird: is, this is actually, like, a replication of my career to date. It's always been like, I've built websites and apps and all that type of stuff, and the question's always been, do I take the off-the-shelf thing or the do it yourself thing, or do you do a custom thing that literally is exactly what you need versus shaping something else to work well enough to get you where you want? And it,

Brendan: Right

Baird: the same conversation again, is like, a Grok Bot or a Instinct is like the off-the-shelf thing, like the Squarespace of agents essentially. Which again, is gonna get a lot of adoption, but you're gonna have to kind of work in its own system. You know? Like, it's only g- it's gonna have opinions, it's gonna do things a certain way, and you've gotta be cool with that and be able to, like, make that work for you. Where, you know, my whole career has been like, how can I shape the thing to do what I want it to, and have, like, that intent behind it? And I think there's room for both. I, I think it's not like a one's the right way and one's the wrong way. It's more of a, like, what are you actually looking to get out of an agent setup, and, like, what do you actually need from it? And then it's really just, like, playing around with it and seeing, like- Are you getting the right results? Like, is it actually giving you what you need? And, you know, my gut says the fact that a lot of what I've seen is people are jumping from Grok Bot to Muse to Instinct is like, they're probably not. You know? Like, it's really cool, and it's that AI thing where it's like, oh, this is really awesome, like, on the surface. But once you scratch past the surface, then you start seeing the flaws. And that's the other, you know, part of all of this is like, and I still struggle with this, is letting the AI just handle everything for you and not being an, like, true director of all of this leads to so much trouble. Like, if you're doing simple stuff or if you don't need any taste or any, like, actual direction or human involvement, sure, like, but you're probably not solving an interesting problem. You know, like what I've seen is… A- and, you know, this is something we've been doing on the, you know, work we've been working on together is, like, what I really love right now is leaning into transcripts. So like, a transcript of anything we talk about is like a working session essentially, versus just us having a meeting. Like, it used to be you'd have meetings and you'd be like, "Oh yeah, we wanna do this, this, and this," and we talk about it, and then you'd have to go and execute. You know, actually leaning into that human element and feeding it back into the AI and then directing it, like that's that kind of customization I think you're talking about in a different form, where it's like using our actual knowledge and the things we talked about as the driver for getting work done versus like, "No, let's plan it completely from, you know, fragments of our r- you know, memory of what the, you know, meeting was actually about." Or, like, we have to go off and program it all, you miss some of the details. And so like that's one of the things I'm learning is like getting that as like a integral part of your workflow has made a huge difference in both how much I'm doing, but also how much quality there is in what I'm doing. And like, again, like this is getting a little off topic, but I think it goes back to that finding the right context and then finding the system around it is such an important piece to getting all these systems to actually work the way we want them to.

Brendan: Yeah, exactly. A- and, and it is… It's getting to the point too for me where, like, these new things come out, and I don't really feel the need to go try all of, all of them

Baird: Yeah. Yep

Brendan: You know? And I was even thinking about this the other day, 'cause I still have Comet, uh,

Baird: Sure

Brendan: uh, Perplexity on my, my machine. And it's just like, I, I haven't opened that in

Baird: Yeah.

Brendan: even know how long.

Baird: Well, you, you don't need to anymore. Like,

Brendan: I know

Baird: there's a, a browser baked into your Claude Code or your codex, so it's like, why do I need…

Brendan: though.

Baird: Sure, yeah

Brendan: Like, they are building… a lot of ways, like, they're all kind of building different things at different times, but they're all kind of

Baird: Yep.

Brendan: parity with each other.

Baird: Yeah

Brendan: like, one part of it. it also just doesn't feel like, like you don't need all these tools, at least not for what I'm doing yet. Maybe if you're running a full company or something, or you got a bunch of companies, maybe you need of these tools mixed in to kind of like, I don't, I don't know, maybe Grok Bot is doing customer support help or something, or maybe Instinct is, and then you got… May- maybe it's vice versa and maybe Grok Bot's doing coding. I don't know. But, like, even then, I don't know that you would need both. Like, you probably just need, like, maybe… You probably, honestly, you could probably get away with just Grok Bot because you got Cursor in the mix there, and so, like, that's maybe one solution. I haven't, you know, I don't even know where Meta Muse comes in as far as coding. I'm sure it can do it, but, like, I

Baird: Yeah.

Brendan: haven't

Baird: I think what I've seen is Meta Muse is more like shopping and, like, more of the non-technical side of things.

Brendan: Yeah,

Baird: you can do anything, but yeah

Brendan: Mm-hmm. And I think there's a need for that too, but I mean, my, my Claude setup can do that sort of thing. Like, I literally am trying to find hockey pads now, 'cause yay, hockey season, and need to buy them. But, you know, I, I don't wanna do that work, so I told my command center like, "Hey," I went and tried, or I had my son try these on. are the sizes I need. You go find me the best deal." And because I don't wanna go on Facebook Marketplace or, you know, any of

Baird: Yep

Brendan: so yeah. A- and that, I guess, that all just comes back to, the fact that these, th- these are gonna fly because people know about Meta, they're gonna use their tool because they're already in that ecosystem, or it's more simple. You know, much more… A lot less people are going to do what you and I do, and that's totally fine, 'cause we aren't trying to sell it as a product. We just want something that works for us. And we also know, and even OpenClaw people are, you know, willing to put in more effort they want that extra power. So, you know, it's not surprising that that's the case and it, it, like that's kind of where we are now. But, um, yeah, I, I don't know that I have a whole lot more to add on that specific subject, but it, it is kind of an interesting place to be with it. And like, I still feel like there's work for me to do. I don't wanna give up what I've put… And it's not even like a, a sunk cost kind of thing. just like, I don't think I'm gonna get anything more out of these other tools than what I've already got today, w- why switch?

Baird: Yeah. I, I think for you, the only other place you could potentially go is, like, where I have been working, where it's, like, fully custom software. And, you know, one of the, the benefits too is, like, if I really wanted to, and I wanted to spend the money, like, I could connect my current agents that I've built into a phone number, and then I could text it. Like, it's just a different modality. Like, it, it's really just like spinning up a Twilio account, but it just costs more money. Like, I already have it connected to my Slack, and I can talk to it there, and I've got a chatbot in…

Brendan: of thing up

Baird: E- exactly. So it's like… A- and the funny thing is, like, if you start… Like, I've got it connected directly into my Claude codes as well, so I can be like, "Hey, Claude, go talk to my agent and figure out what I need to do next for this," and all that context is there. And, and so if you do it right, and, like, again, like, you can do this with your command center as well, where I think it's harder to do unless you're connected to, like, a Notion or something like that with your Muse or Instinct or whatever. Like- You know, just me personally, and you know this about me, like, I like to have control of these types of things versus, like, I wanna use a ton of third-party tools. And part of it's just, like, to do it right, you've got to spend so much money on all these paid accounts that give you nothing other than, like, a little bit more usage or something like that. And it's like, especially these days, spinning up the pieces that I need is so much easier than it used to be. Like, I always had the, the ideas of like, "Hey, I'm gonna build, like, a little bit of this and a little bit of that," and then I would never do it because it would take so much time and effort to do. Nowadays, it's like, you know, I just did it. Like, let's build a Kanban board that's a bit like Linear, and I built it in, like, a night. And is it Linear? Of course not. It's nev- never gonna be that. But it's like, it is MCP enabled. It can actually consume stuff, you know, that I need it to, in the way I need it to, taking transcriptions and putting it at… Like, that's all I need is, like, just enough of the infrastructure to then feed everything else that I'm doing. And, like, we've come so far to be able to do it that, like, if you know how to program and you know how to use, you know, the tools, you can get something custom without having to sit there and spend months of your life, you know, at a keyboard making it happen. And,

Brendan: Yeah

Baird: the, the upkeep stuff is not as bad as it used to be. It's not like you have to be doing maintenance 'cause you can be like, "Claude, this is broken. Go fix it," and it can generally do it pretty quickly. And I think the one other thing I'll, I'll say, and I'm curious if you've noticed it at all too, is like, one of the bigger things that I've seen while doing all of this is like, you know, I use Opus 5.5 on Medium as, like, my driver. And I've noticed it's doing a much better job of actually, like, getting through things and doing, like, a solid job with it. Like, it actually gets good results, so you have to babysit it less. Like, you still have to be involved. You still have to be driving it and directing it, but you don't have to be, like, looking for the holes that it used to leave all over the place. You can more focus on, like, did you interpret what I said correctly? You know, like, is this actually what I want? And shape it that way. But you have to, you know, babysit it from a quality perspective, at least, like, for what we build, a lot less than you used to, which has helped tremendously to push some of this stuff forward.

Brendan: Yeah, I agree. I mean, and even like code in general, like not, not just UI stuff, but like code. Like I, I… It already writes better code than I ever did, so

Baird: for sure

Brendan: I've re- I've relinquished that. I don't even really pay attention to the code beyond like having a separate agent come through and say, "How does this look? Are there any gaps?" And I find that valuable because it's gonna find- that's also gonna

Baird: Oh yeah. Yep

Brendan: things I generally wouldn't find. Um, but yeah, I, I agree. Like, I haven't… I couldn't even tell you the last time it actually built anything where I was like, "Yeah, that's completely wrong. What were you thinking? Let's do this again." You know, maybe something was missed, or maybe I didn't give it something that I realized.

Baird: Yeah

Brendan: Um, like, like even a, a good example is I've got a consulting engagement that I've been working on, and, and this is purely code. It's doing Ruby on Rails code stuff. It's not like AI specific. you know, the, the person I'm working for, they are completely on board with AI coding. They're doing it themselves. basically just needed more capacity because they've got so many pull requests open themselves, they just can't keep up. And so I have do- I've worked with them in the past, and I used AI, and yeah, in the past it was always like, "Oh yeah, this is a mess." Like, "We gotta fix this." it's basically the same thing I just described. It's like it built it, and maybe I didn't quite explain it right, or I'll, I'll give it to, you know, my, my, uh, client, and they'll be like, "Yeah, let's do this and this." But it's not like that's completely wrong. It's like, "Oh, let's just tweak this."

Baird: Yeah, for sure

Brendan: a better place to be for sure. It definitely has been a major improvement

Baird: I, I mean, you get there, like that was the world when you wrote code by hand too, was you would build something and they'd be like, "Oh, that's not quite it." Like, a- and I think that's part of the value, and that's what you and I have pushed on, is like the iterative nature of like don't build everything at once, build increments. Which again, like that's not new. That's how we used to build things anyway. But people got so used to the one shot with AI that you lost the value because I find it's like build a little bit, see how it works, and then be like, "Oh yeah, but what if we could do this?" That's like taking it a, a direction you didn't even notice until you got there. So if you're one-shotting it, you're missing all of those opportunities to like shape something human out of it, versus letting the AI make all of the decisions and shape it to the f- you know, final product by itself. Like, it's so much more rewarding, and I think better, you know, output if it is that like going back and forth with the machine. And like having other people as well, so you've got multiple opinions on it while you're doing that process, is like, it's, I think it, it's pretty amazing

Brendan: I, I agree. And, you know, you, you just made me think, I don't know why I didn't piece this, piece this together before, but it is very much like if you try and one-shot a, an app or even a feature and like a big fea- really big feature, for instance. Or if you try and have the AI just, like, gather all the things from Slack and email and everything and try and give you a, a priority list, it's like it's basically causing the same problem, especially when it's a full app that it builds. Because what it ends up doing is in both cases, it's just giving you this giant to-do list to then almost figure out on your own without its own help, especially in the app case, 'cause then it's like, okay, now I have to comb through an entire app and find the things that I don't like and need to be tweaked versus one feature or even one small little iteration of a feature then say, "Nope, let's fix this one thing. I don't need to look anywhere else 'cause you don't, didn't touch that other stuff." And so yeah, that, that's i- interesting to compare those two

Baird: Yeah, and it builds understanding while you go. Like, if, if you one-shot it, you don't take part in any of the decisions, so you don't have the understanding of… Again, it's not about the code. It's not about, like, do I know every line of code? It's about, do I even know the intent of what every feature in this app does? Do I even know how to test this or see if it actually was done right? If you essentially outsourced all of that decision-making in the process. And so, you know, that, that's something that I've learned, especially recently, is like the more you can keep it that iterative human involvement. And it's not even, like, human in the loop. Human in the loop to me is like, push the approve button. This is truly, like, a collaborative process where it is, like, you break it up like you would normally have done with a project, and then you build it intentionally in those small chunks, so you're making decisions for real while using it to get the understanding of it. And it, it's like really kind of going back to what we used to do, but in a whole new way

Brendan: Yeah, and then we are, you know, a layer of abstraction higher than we used to be. Instead of writing the code, are now the ones that are, are basically project managing. But, but also it benefits us by knowing how this whole process works. And yeah, it, it… You're right, like the, the understanding is the biggest thing, and without that understanding, like you're just gonna build another generic app. And whether it's to sell to other people or to use yourself, like it all comes back around to what we were talking about with these agent frameworks and, and just these agents, is like if you one-shot something, more likely you're gonna get the most average app out there. You're not gonna like it, and you probably won't ever use it, and you probably won't be able to sell it to many people, versus having something tailored to what you want, and you actually get, you know, a custom app or at least a little more custom than, than what other people would get. And, you know,

Baird: Yep.

Brendan: bespoke, which is

Baird: For sure

Brendan: I- I-- becoming increasingly, uh, you know, exciting to see in the

Baird: Yep.

Brendan: AI stuff, so.

Baird: No, for sure

Brendan: Yeah, I, I don't think I have a whole lot else to add. What about you? Is this a good place to wrap up

Baird: Yeah, no, I think we covered all the big things. I mean, we could probably talk for hours, but I think we, we covered enough for, for this episode

Brendan: Yeah. All right. Sounds good. Well, if you enjoy this, uh, like and subscribe on Facebook. Or Facebook, oh my gosh. Now I got Meta Muse in my head. I'm thinking Facebook. Like and subscribe on YouTube or Apple Podcasts, wherever you get your podcasts, uh, we'll see you next time