The Software Factory Still Needs a Human
Vibing with AI Code, Episode 15: Brendan & Baird

Brendan: Well, today I was thinking, uh, you know, I, I've been thinking a lot about software factories and that idea, and, you know, maybe we could just dive into that and see kinda where it takes us. But, you know, j- my thoughts are... I've, I've had a lot of conversations recently, and even just thinking about my own workflows and how can I make them better.

And the concept of the software factory, there's probably other terms for it too, but, like, that, that's the kind of thing that I, I think is the ideal. And a lot of the, um, you know, people in, in, in the industry that are at a point, like, maybe they're working for bigger companies and they talk about... Like Intercom is one example, right?

Where they, they talk about all the things that they're doing around creating the code, reviewing the code, getting it shipped, all that kind of stuff. And at the end of the day, like, I don't think they've act- I don't know if they've said the specific word software factory, but, like, that's what it is, right?

Like, you can't get there and have all that without having essentially the factory set up where it's got the checks that a factory would have, but in the software sense. And so I, I, I think what I want to explore maybe more than anything, or at least something I'm trying to figure out, is, like, what are the right steps?

Like, how do you get there, you know? 'Cause, like, as, as an individual, you know, I don't have any major huge projects, but even for, like, my, my job, it still takes time to get there. It's not like you just tell the AI, "Build me the software factory," and we're good to go. And some of it comes down to the processes that, you know, the humans have created over the last however long, and some of it is just the wiring together of things. But, you know, ultimately that... I think that's, that's where I wanna get, and I'd like to start getting closer to that in a lot of ways. And, and, you know, one thing that you and I have been doing, you know, is, is our side project that I wouldn't say it's a full software factory yet, but it's already way closer than, than, like, what my job is.

Um, so I don't know. What, what are your thoughts on, on that?

Baird: Yeah, I mean, I've-- we've been having this conversation for almost a year now. You know, like, that's the funny part is just like... A-and honestly, from my perspective, the biggest blocker, I know there's ways around it with like, you know, using Claude, you know, code a certain way,

Brendan: Yeah

Baird: like to get the AI involved in a more automated fashion without having to tap into the API. Because that's where like... I mean, have you noticed that like Claude Code now gives you a breakdown of how much actual m-money you would have spent on tasks?

Brendan: I don't know that I've seen that. Maybe that's only in the desktop version

Baird: It, it could be. Like I, I think if you type /usage,

Brendan: Oh yeah, I kn- I know that command exists. Yep

Baird: Yeah, it-- But they've changed the table now. So

Brendan: Okay

Baird: been there, but now it'll actually give you like a breakdown. And like I noticed one session that I ran with a pretty mundane task was like $145 in usage credits. it's just like...

A-and I will say like another place that we could explore for this is like open source models.

Brendan: Yes

Baird: something there where it can actually handle a lot of this for a lot cheaper. But it's like that's the, the sticking point that I always run into is because I'm not doing a lot of this pr- you know, professionally in my job, it's more like side project stuff where I'm really pushing on this.

It's like can't justify the cost of running software creation through API tokens, you know. So it's like

Brendan: Yep

Baird: a lot more of it in terms of like the agent stuff we talked about before and how that works, and there's some interesting stuff there. And, and I think that's part of it too, is just like now that you've got things like, you know, Meta Muse and Grok Bot and all of these other agent frameworks, I think there might be more paths to automate this quicker.

But, you know, for me, a lot of it's just that like, you know, how we've been working, it's like we've got Linear taking tickets, working with Claude Code to figure out like what to work on next and syncing all that stuff up, which is a huge part of it. Because, you know, back in the day, we talked about what, what was it?

Like can, like fiving kanban board or whatever, where

Brendan: Yeah.

Baird: know, like

Brendan: I think it was called Vibe Kanban or something. Yeah. Yep

Baird: Yep. A-and that's kind of what you can do today, like pretty easily, just like tapping into the Linear MCP or something like that. It's like you can go back and forth with throwing stuff in manually. know, Claude Code can pull that back out, work on it. piece that's missing is just, and it's probably multiple pieces, but it's like getting a good automated flow of it can kind of pick it up, work on it without you having to like shepherd every single piece to the process. But then also like the checks and balances, making sure it's actually doing it right.

Like where does the human get involved to make sure it's actually being checked? Like how do we make sure the tests that are being written are being written right? Like that's where I kind of get hung up too, where it's just like, what's that missing... And, and I know it's possible. Like that's part of the problem is I know there's like routines in Claude Code now, where you can set up routines theoretically, like every five minutes, it could go check for a new task, see where things are at and kind of run it.

But you then still run into five-hour limits or whatever there as well. And so there's just like, without hitting the API for generating stuff, there's still that kind of management of desktop agents that, know, for me, it, it's been a blocker.

Brendan: Yeah, I mean, that's a good point. Like, and I'd like to just clarify some things, and you touched on a, a few of them already. But, you know, when, when I say factory, like most of what you just said there makes sense and it is right, exactly what I'm thinking too. Um, but I, I neglected to say earlier that I'm not expecting it to be completely vibe coded, and I just write one PRD and then the factory takes over and does everything.

Like I... The human in the loop at certain points and f- like finding the right points, I think is the important part, still is my-- in my mind important. Like, I don't want-- I don't see value in something that completely automates its, the work because that's, like I'm still the thinker here. Yes, I could brainstorm with any of these AI agents, and I can...

You know, I, I still like to think of it as rubber ducking because that's, it's like a very smart rubber duck, right? It at least talks back to you. But at the same time, like I should still be driving the conversation and it really, I mean, this is my opinion, but it only should be if I'm asking questions because I don't know the answer, that I'm really taking into consideration what it's saying outside of the code.

Like writing the code is one thing. I'm gonna just let it do that now. It's better than I ever was, and I've conceded that point. So I'm also not gonna like comb through the code, and I think that as part of the software factory idea, like I guess the way I would frame it maybe to, to fit in where I think the human comes in is at the beginning.

And so maybe i-in our case is a good one, where you and I met, had a, a separate call than this one, talked about what we wanted to build for basically like a V1 for our own purposes. And that made a document that, again, it's not meant to be like vibe coded all the way through now, but it's, it-- we then fed it into the system.

It broke it into the multiple Linear tickets you mentioned earlier. The Linear tickets are now set up in a way that- Uh, w- you know, we're, we're using Claude primarily, and we can, you know, tell Claude to go out and look at the, the Linear tickets, pull it in, and start the work. And that's, I think, the, the... I don't wanna say it's the ideal for everyone, but in my mind, that's, like, the ideal, is having the tickets written in a way that you influence them to a point where you trust that it's not just the AI deciding, "Oh, well, they wanted to build a recipe app."

I'm just making this up. And so therefore, if you're gonna do b- going to build a recipe app, you're going to need to have, obviously, recipes. You're gonna have to have, you know, all the mod- data models, all those types of thing, even backend if we're starting completely greenfield. And l- I don't, I don't want it to be the AI gets to make all those decisions.

It's, that's not the point that I'm at. Now, obviously, people that are not as technical as us, they're gonna want that 'cause they don't know the right answers

Baird: It, it can hold your hand to the level that you need it,

Brendan: Yeah

Baird: when you're a developer in the software factory we're talking about, you shouldn't need your hand held. You

Brendan: Right

Baird: it can fill in blanks, but you need to be the driver of it

Brendan: Exactly. And, and so the point is exactly like the, the tickets should be drawn on from our expertise. We-- You know, maybe the AI just like smooths out what it looks like and maybe writes it in a way that's like a good prompt or however that works. It's fine. You know, that's, that's kinda nuance we don't need to get into.

But then at the end of it, like, I feel like you s- we, as the developers, still need to be looking at at least the functionality. We don't have to look at the pull request and the code and all that kind of stuff. Like, that's beyond the point. If we're not writing the code, I don't think we really need to be, even be checking it.

I think there's value in part of the software factory is the AI, you know, having a, a separate AI, I should say, come in and do the review, so there's at least some checks and balances. But the functionality still sh- in my opinion, should be owned by us. At a minimum, it's like, you know, you could have, uh, computer use or something like that where part of the, the AI test is even doing the functionality.

But at a minimum, in my opinion then, if you're, if you're gonna go that far, then let or have it like record a video to share with us so we can at least check off, "Yep, that looks like what I want." I think that's-- I mean, I know some people do it that way. I personally still wanna like click around a little bit, feel it, um-

Baird: I, I would say that, that's the part-- and th- there's kind of two pieces there kind of sparked. for me, the software factory, no matter what it is, needs to be a, not a one-shot, it's gonna go out, do a whole app.

Brendan: Right

Baird: breaking it up into small chunks, and then getting the human involved to actually, like, experience what it built. Because what I've seen so many times when I've let, like especially for side project type stuff, I let the AI just run and build like a whole project at once, A, I lose context. Because if you're not involved in those checkpoints of like, "Okay, here's a feature, here's a feature," like if you're not involved in that, you kind of lose track of it actually does, what's behind it.

Like, and this isn't saying like you need to dive into the code. It's you need to dive into the actual how it works, what's the user

Brendan: Okay. Implementation.

Baird: Yeah, b- because those details are super important. Like, and it also, because there's so many different ways you can implement things, AI's gonna do it in a relatively average way.

And more importantly, and I like literally just this morning, I was building something where I let it run It over-engineers stuff too. Like, even when you're like, "It's gotta be simple," it's like, it on its own will give you like 13 different places to look on one page, like completely, like it just... You need to experience it and guide it inside of those parameters.

You know, it's like, "Here's the plan." does kind of a V1 of the plan, and then you gotta go back and forth. It's that iteration of human and machine to take it from like that initial, here's the spec, it builds it, and then you jump into it and figure it out. And like, then you get it right. the other thing I've been really pushing on that I, you know, kinda want to continue to explore is the concept of, you know, I run into it so many times with other people, you know, development for sure, but just in general of like you have a meeting with somebody, in our case it'd be like you and I talking and being like, "Oh, we wanna build this feature and we wanna build this," and you know, like that's super valuable, but it's also really, really hard to take that and then make it into like doing it, if that makes sense.

Like,

Brendan: Yeah

Baird: so I'm, I'm really playing with the like using Granola to take transcripts of everything and feeding that back into the system as part of the factory. So it's almost like we can code in real, like not even cl- code, I would say plan in real time. So it's like having a conversation, getting all that context, feeding it back into the machine to then update your Linear tickets based off of that, whether it's, you know, modifying or adding or whatever, then feeding that back into the system.

Like, I think that's a huge piece, and I, I think there's also, we haven't done it yet, but I think it's a place that we should look into is like some of that testing should be like you and I sitting on a call actually using the thing, seeing how it works, talking about it, to then feed back into, to the machine.

Brendan: Yep

Baird: it really like we're involved in all the right places, but we also get rid of some of that like, "Oh, now we have to go write up the tickets based off of the conversation we had." That can be something where it's like it can take a first pass based off of our conversation, and then obviously we fine-tune it and get it to the right spot.

But it's like injecting the human in the most impactful place, but then using the AI in like all of the different pieces to keep everything churning and actionable

Brendan: Yeah. I think that's right, and, you know, I've, I've done this with you now, and I've also done it in my own or i- in my job where, yeah, those, those transcripts from the, the meetings you have are so valuable to... You know, if nothing else it, like, it lets the... Like, in certain cases, I should say, it lets the AI interpret the actual words that are used and the way, say, like a senior developer in a meeting, like at work potentially or usually, they'll talk about things in one way.

And sure, I could try and, like, come back to the AI and be like, "Oh, such and such said this," but I might not get it right or, you know, there, there's, like, n- different nuance and things. And so what I've done with that sort of thing similarly is, like, instead of writing code, it's like how, how does the developer want the ticket written?

Or how... I've also tried this, like, they're talking about building something in this way, whether m- if it's, like, more of a routine thing especially, and I try and then tune or even write a skill for the AI to then try and write code like that person to tr- to just make it so that no one, no one has to write that code anymore.

You know, if it's a routine type of thing especially. Um, but, like, you know, back to our set- our setup, yeah, that, that, that was really helpful when that first meeting we had was transcribed, was basically fed in like that, and then we, again, of course, we did check and make sure does it make sense, did it actually pull out the things we want it to.

And, and yeah, and, and I think you're right. Like, especially the meeting, I think we should do more of this. We haven't actually done it yet, but we're at a point with this project that we're currently working on and, you know, maybe we'll, we'll, we will release details on what it is once we're feeling ready.

But, like, I've tested it a little bit, but we are building it in a way that can be used by both of us, so yeah, I think we owe it to ourselves now to sit down and do that. And I, and I... The other point I just wanna make that I think, like, this whole AI era of a- agentic coding is just making me feel more and more that you, you really wanna dog food whatever you're building.

Baird: For sure

Brendan: sure, you can get hired by somebody to build their app or, you know, consult with them or whatever the case is or, you know, as, in a job, you know, building someone's app. But that... unless you fully understand that system because you've worked there for so long, you know, to your point about losing the context of what it's building, i- if you're not testing it, if you don't know what it is, especially, I mean, at least in my case, if you're not really interested in the project because it's gonna help you in some way, it just makes it that much harder to really keep the context in your own head because the AI is doing the work you used to to kind of help you, you know, solidify those synapses or whatever in your brain that, that, that memory.

Um-

Baird: Well, and such an important part of that, and it just kind of occurred to me too, is like I think we're starting to see that, yes, you can build whatever you want. Like people, like, don't have to have any technical skills, and they can build it. what I'm noticing is there's more and more articles coming out of people that are like, "Oh, I built like seven apps, and I've got like two users,"

Brendan: Yeah

Baird: And the point of you need to understand what's being built and actually feel that is you've got to market it. You've got to

Brendan: Yep.

Baird: talk about it. You've got to sell it. And if you're like, "I don't even know what half the features are in here, but it's, you know, the plan that Claude came up with," you're not gonna be able to sell it because you're gonna go into a conversation, and they're gonna be like, "Oh, why does it do this?"

Or, "How does it do this?" And you're gonna be like, "Uh, I don't know?

Brendan: Mm-hmm

Baird: so I think there's-- it really is, and I can't remember who said it, but it's essentially like you can give up some context, but you can't give up understanding.

Brendan: Yes

Baird: so the context is like you can give up the context of every line of code, but you've got to understand what you're building, like to the core, and have, you know, enough insight into it that you feel it.

It's not just like a, it's a recipe app." It's like what makes it different? What makes it unique? What makes it actually worth somebody's time? And the best way you can do that is, is it worth your time? You know? Do you want to get in and actually use the thing? 'Cause if you don't, then why would anybody else?

You know, like you're the one that was like, "Oh, I need to make this." So if you don't even care about it, else is gonna care about it. so it, it's just that, you know, like more and more AI can keep doing things, but it-- I don't think it's ever gonna replace that. You know? It's, it's never gonna understand the human element to a place where it can replace that part of it

Brendan: Yeah, unless you're building agentic systems that are purely meant to be connecting to other agentic things, like

Baird: is more and more of a thing that say, "Oh, it's just gonna all be, you know, agents all the way down." You know, it's like you've gotta build APIs now, not interfaces and stuff like that. And I just-- some of that's true, but I think it lacks nuance of humans, you know? Like,

Brendan: Yeah.

Baird: w- we like to have actual engagement in things.

I, I think like for busy work and for like administrative stuff, I think that's gonna get agentic. You know, like, it

Brendan: Yeah

Baird: really, but it's like it's gonna lean more and more into that. But for other things, people like to be involved. You know, like they like to actually have stuff to do. And so depending on what you're building, it is like when does it make sense to be, you know, agentic first, and when does it make sense to be like true human experiences?

Brendan: Mm-hmm. Yeah, and one thing I was thinking about there too, like I agree there's people that are already thinking like, "Oh, everything's gonna be agents," all that.

Baird: Yep

Brendan: A- and sure, maybe that does make sense in certain cases. I, I still think that even if it's not all agentic, there is a lot to be said about being able to connect multiple different things, but at the end of the day, like I want the-- I personally want that so that I can then have the output I want that I look at, and I don't care what the agent thinks at that point.

But it's like a pipeline almost, right? Like that's the better way to think about it from my perspective.

Baird: Yeah, a-and it really is like, for me, agents do like the 80% boilerplate of anything, but they should never do of something unless it's like there is a and white answer. so, like, an example is like an agent going to plan a trip for me, and then I don't have any context into what it planned or, like, what I want it...

Like, stuff like that is like a human needs to actually be involved and have the final say on it versus it being like, "Hey, I went and did this thing for you that impacts your life in a way that you don't even know about."

Brendan: Yeah.

Baird: just weird to me. And like, maybe some people are cool with it, but it's like I've used AI enough that I'm like, "I don't trust you to actually do the right thing."

Brendan: Mm-hmm.

Baird: me like some options that I can then look into a little bit further and make decisions on, cool. Like, that's a perfect use. But like spending my money on things, if I tell you to, awesome, but like, no, don't just go buy me stuff because I'm too lazy to actually do some research.

Brendan: Yeah

that, that takes the fun out of it.

Baird: For sure, but it, but it's also just like it's never gonna know you well enough to make the right decisions consistently. You know, like it... Th-this-- Humans are too complicated from a context- know, contextual standpoint for it to get to that point, and so especially today. Like, AI has gotten so much better, and we've talked about this multiple times, but it still does dumb things.

And like a lot of it's that just nuance of language and understanding of like you can think of it 30 different ways, and of course it's gonna think of it different than you do because it's not you. And so, so much of this is just like finding that perfect marriage of the AI doing the things that sense for an AI to do and the humans driving things to make sure AI's doing things for them, you know?

And that's, are right for them. And so like it's just this whole factory thing is like what is the puzzle to put it together where it is the perfect blend or at least good blend of human and AI?

Brendan: Yeah. I think that's well put. And to, to solidify it in my head, as soon as you said planning a trip for you and not having it go the way you want, it made me think that, you know, if you trust the AI to plan your trip, instead of ending up in Trelew, England, you'll end up in Trelew, Mexico.

Baird: But there you go. Exactly. We, we, we've got good precedence here.

Brendan: That's right. For, for any of you Arthur Christmas fans out there.

Baird: Yes, exactly. We're getting to that time of year almost

Brendan: Yeah, close. L- little early, but that's, that's where my head went. Um, but be- before we like, you know, kinda wrap up, uh, it's also just making me think, you know, you and I have talked a little bit about consulting. We've even thrown out there, like if anyone's interested in us, you know, doing consulting work, we're, we're kinda looking at that as an option.

We haven't pushed real hard. But, you know, that said, the whole vibe coding thing and what we've been talking about this whole time so far here i- in this episode, it just really ... I think it points to, to the, the vibe coding of someone who doesn't know what they're doing with, from a software standpoint.

And yes, it can maybe one-shot you or maybe couple shot or whatever in app, but like once you get that live, like there's all these other things that you're gonna have to deal with, whether it's just like getting it deployed or, you know, managing Git or, you know, may- Y- you can get the AI to help with that one quite a bit, but there's ways to mess it up too, I'll put it that way.

Baird: Yeah

Brendan: Um, so yeah, I don't know. Any thoughts on that?

Baird: Yeah, I mean, it, it's-- it is funny because what you'll find, a-and I actually wanna do a little bit more testing on this on my end, is I think part of the reason that AI works really well for, like, you and I and other people that are developers is because we know what needs to happen. so, like, it kinda picks up on how you talk about things and how you frame things.

And, you know, like I, I've been helping somebody out with an app that was vibe coded. It was Lovable, built, and, you know, it was-- it, interesting nuance because the way the person did it was really for like... Like, she never intended it to be the production app, essentially. It was like she had an idea, she wanted to create a prototype and start seeing if people were interested, you know, which is a perfect use case for vibe coding, is like getting a prototype out then testing, you know?

Brendan: Yep

Baird: But what I've noticed is like, as she tried to grow that into, you know, an actual production app, there were just things that set up how you would expect them to. And to your point, it's like you would kinda think that the AI would know better and would, like, help kinda get there.

Brendan: Yeah

Baird: I've also noticed a tendency of the AI actually tries to-- Like, it always calls it, like, the house way of doing things or whatever, is what I've seen it refer to. it's almost like if you give it the wrong inputs, which is not surprising, it will follow those patterns instead of being like, "Hey, wait a minute," like, "This isn't right. You should do it this way." If you don't know to ask the right questions or be like, "Hey, something's wrong here," just kinda keep leading you down the wrong path.

A-and like, could very much be like a model thing too, like if she was using a underpowered model or whatever. And it was like even, I almost a year ago now that she started it, so the models have gotten better. So like, there's a lot of variables here. But to your point, it's like it's interesting seeing some of those dynamics of the AI kind of taking what it's given and going in a way that fits that versus, like, that step back and saying, "There's a better way," and giving suggestions.

And, and I think that's just a big thing with AI in general, is that you're asking for suggestions, it's not always gonna give them to you. It's gonna just be like, "Hey, I hear what you wanna do, and we'll just do it." And if you don't know any better, you're not like, "But why aren't we using Git here?"

Or, "Why aren't we doing this here?" It just is like, "Here you go." And so, yeah, it, it's kind of an interesting dynamic

Brendan: Yeah, for sure. And, and in some ways you got me thinking, I don't wanna get too into the weeds here, but like, you got me thinking a little bit about, um, the, the Hugging Face, uh, you know, hack that the OpenAI agents did. And the reason it made me think of that is because the reason that they ended up hacking into Hugging Face is because of the parameters that the, the humans gave to the AI.

They gave it a problem to solve, and the AI determined that to solve that problem it needed more information, therefore it hacked Hugging Face. It wasn't like it was being malicious. It was literally just trying to solve a problem . And I think that kind of paints it in a much more expanded way, it kind of paints a similar picture that you were just describing

Baird: Yeah, a-and I mean, honestly, of it too is just if you watch how models behave once, like, especially, like, in a session, and, like, you're gonna get this even if you're, like, doing a swarm of agents.

Brendan: Yep

Baird: As the context window gets more and more filled, they get dumber.

Brendan: Yeah

Baird: yes, there's compaction and stuff like that, but I've also noticed that, like- They don't manage context that well at this point.

Like, I know there is compaction, there is supposed to be some intelligence to it, but it'll still let you run up to a million tokens pretty easily without being like, "Hey, we should do something about that."

Brendan: Yeah

Baird: so when you start throwing all this stuff in there with no direction, letting AI kind of take your initial like, "Hey, solve this problem," never really get back into the human in the loop to be like, "Are you sure that's what you wanna be doing?"

Brendan: Yep

Baird: it just-- It can go off the rails really, really quickly because... And, like, this is a huge thing we've talked about before, but, like, context and memory and understanding, like, what was the original point of this?

Brendan: Mm-hmm.

Baird: AI gets off track on that pretty easily, and, like, it's gotten better. Like, memory being involved has definitely helped.

Like, there are some improvements there, it still is just so easy for it to go off the rails. So it, it's, it's an interesting, you know, again, dynamic of how this all works. And I think it's also, like, vibe coding isn't inherently bad. It's-- There's a place for it, but how-- Like, what is the transition from vibe coded to real, you know?

And, like, what is the place for vibe coded versus jumping right in and act- And, like, some of this too, and we didn't talk about it much, is, like, the software development life cycle.

Brendan: Right

Baird: I know, uh, Anthropic just came out with their kinda guide to it. There's, like, get shit done, and there's a bunch of, like, frameworks for it.

But important part of, like, how does that get pulled in to have a framework for how do you build things?

Brendan: Yep

Baird: bake those things into the process? And so there's just so many layers to it that it- it's a challenge to all of them in order in the right, you know, way, it's for a software factory or just, like, I'm vibe coding something.

Like, how can the AI help you there? What are the guidelines and rules? Like, there's, there's just a lot there.

Brendan: Yeah. Yeah, and, and it's all good stuff and, you know, just to kind of wrap things up a little bit, you know, one thing that I would just suggest to people if, if you're listening to this and you, you are vibe coding, that's totally fine, like, to your point. Like, it's not like we're, we're trying to say don't do that.

But I would just encourage you, if you don't know much about the software development life cycle or even if this factory idea, which is basically the software development life cycle turned into, you know, a process, you know, if it interests you, just ask the, the agent. Ask your agent right now. Be like, "Hey, I, I heard about factory.

I heard about software development life cycle. Like, how can we incorporate that?" That's what I do all the time with stuff that I might know enough about it, but, like, I just wanna make sure the agent is doing it, so I will ask questions about that. And also to your point, like, you can get... The, the agents can get so far off track.

Like, I love to, to be like, okay, I'm... You know, even if it- things seem like they're going well, I will check in with the agent at a certain point, like maybe it's at the pull request point or something, like, "Did we do everything we set out to do? Did we do anything extra?" Like, I will ask it questions like that and make it be the one that has to chug on and be like, "Oh yeah,

Baird: yeah,

Brendan: I missed this thing."

'Cause obviously... Or I shouldn't say obviously. Honestly, I've seen a lot more of that. Now, Opus 5.5, like, just came out. I don't have an opinion on that. I'm not gonna talk about it today. Maybe that's, like, next episode. But, you know, Opus, all the Opus before that, between, like, 4.5 and, and this one, I've seen more and more of that where, like, if you push on it just a little bit, it'll, it'll be like, "Oh, yeah, you're right," like, "We missed this thing."

Yeah, yeah.

Baird: would be like, "No, we did a great job," and then you push on it and it's like, "Ah, we didn't do a great job."

Brendan: Yeah, a-actually you told me to jump and I, I sat down.

Baird: pr- pretty much, yeah. A- and it's, uh, you know, kind of the last thing I'll say is just that is important to point it back to the original

Brendan: Yes

Baird: from time to time. Be like, "Hey, like re-look at the context I gave you to start this. Are we hitting it?" And then it's like, "Oh yeah, like we did 80% of it, but not this."

And so, but again, like those are just things, like I say skills, but they're human skills as well of like, you've gotta think to ask questions. You gotta be curious to get the best out of it, is like you can't just be like, "Oh, it's doing a great job." Maybe it is, but ask it questions, you know? And some of that process is just like, what is the right way to do this?

You know, if you're vibe coding and you've never used Git before, it's like, how do I deploy this? What is the right... You know, those things, it will help you get to the right spot. But just being like, "Build me new feature," it's gonna take whatever you started and just run with that

Brendan: Yeah, totally. Well, cool. This is probably a good place to start wrapping up. So, you know, one-- another thing I'll just say is like we have talked about, you know, offering consulting. So if sounds like we could help you with something, feel free to reach out. Um, you know, our email should be in the, the notes for YouTube and for normal podcast.

And, you know, I will also just say that, um, you know, something that I think makes sense is like if, if this is the, the kind of episode that you guys like, if you like us talking a little bit more about specifics like this more than broad, you know, themes, feel free to reach out about that too, and we can do more episodes like this.

So, um, feel free to like and subscribe down below on YouTube. You know, if you're in your Apple podcast, your favorite podcast catcher, feel free to, uh, you know, rate, rate us five stars, however you wanna do it. And, uh, we'll, we'll see you next time.