speaker-0: Welcome back. This is ⁓ our very first inaugural in-person episode of Blood Sweat and Tokens. my name is Taylor McDonald. with me as always is Sean C. Davis, right. ⁓ so hey Sean, where are we today?
speaker-1: C Davis. We are at the Marketer Collaborative in Cincinnati. This is a co-working space that I typically work out of. We're in this fancy little podcast studio, which is probably why our sound might sound a little bit better to y'all.
speaker-0: Hopefully. I you know the the set the setup is always kinda hairy whenever you're walking into something new. But I am interested to see how this turns out and This is a really cool place. ⁓ so yeah, you should come check it out if you're in Cincinnati, the Marketer Collective. Or collaborative, sorry. ⁓ okay, so Sean, we're recording this ⁓ actually the same day as we recorded episode number nine. We're gonna kinda try to stack this up because we got some holidays coming up and some family vacations and stuff like that. So in a lot of ways, ⁓ today's conversation is a little bit ⁓ different in structure than earlier, meaning we don't have a demo to go through. That's what we kinda talked about this morning. ⁓ but I was thinking maybe it might be helpful for you to talk us through maybe let's start at the beginning and recap, you know, what we're building. ⁓ I feel like some of the d y we went through a little bit of this this morning, but it was a very ⁓ abstract kind of language and stuff. So let's just be really clear on what we're building. ⁓ we've seen a couple demos in the past. ⁓ we're You know, we'll start there and then I was hoping maybe you could walk us through some of the steps you took to get to the demo to where it is today. you know, are you using Claude Code? Do you prefer the command line interface? Do you prefer the desktop app? Are you using something else? Have you tried any open source models or anything like that? Maybe you could talk to us a little bit about that and then we could kinda back into you know, some of the the higher level, you know, more the more abstract concepts and and why we're talking about those this morning.
speaker-1: Yeah, gotcha. Okay. Yeah, and if I yeah, if I go too deep or whatever, pull me out. But I'll say that so right basically I'm thinking about it from the experience standpoint, and someone comes to us and they're like, Hey, I want to use this tool, I've got this site or not, and we're gonna pull it in. What are all of the piece you you you wanna look at like what are the moving pieces and how do they fit together kind of a thing. Yeah.
speaker-0: But let's start with a description. So, the original idea of our podcast here was to build product, right, using AI tools and AI tooling ⁓ to kind of augment the way that we would traditionally do stuff. So, one of the things that is a major point of contention for both of us over the years, we've been building marketing sites and ⁓ stuff of that nature is dealing with content management systems, so CMSs. ⁓ the big ⁓ kind of bear in the room that we've talked about from the beginning ⁓ for the last nine episodes has been content. Contentful has gone through a lot of ups and downs, and they've gone through a lot of kind of ⁓ pricing volatility and things like that. And they most recently got acquired by Salesforce about a month ago, which was kind of messy. ⁓ or we expect it will be messy in the in the long run, but it certainly justifies and we can understand now why some of that pricing volatility existed, because they were, you know, being courted, we'll say. Yes. So through that experience ⁓ over the last couple of years, you know, we've seen companies like Contentful. Precipitously reduce entitlements, ⁓ increase prices, and generally just kind of close up what is otherwise used to be kind of an open or like a forwardly open type of product. You know, Netlify is one of these products too. You guys offer free tier access, right? I can use Netlify today and I don't have to pay a dime, but at some point, at some threshold, you know, you gotta pay the Piper, right? So maybe we start there. Talk to me a little bit about how Netlify approaches that. Like what is that threshold between between the free tier and the, you know, premium tier and how do you guys when when to the extent that you can talk about it, when you're in those rooms talking about pricing and and all that stuff, like how do you make that determination?
speaker-1: Yes, and I mean, for the record, it's not my call at at Netlify, of course, but I think that it's y you can you can see you tend to see trends, I think, especially in these VC backed companies. And, you know, it was th we were giving way more away for free ten years ago or so, because at that time what the VCs wanted was we want a bunch of users. Now we want users, but we want more of a path to profitability. And somewhere in the middle there it was we want big contracts. So we kind like going, you know, changing the pricing model to, you know, to really ultimately you've got to serve the shareholders, whoever they may be. And ⁓ so I think that under the surface has a lot of effect on how you approach pricing, but I think if you look at it ⁓ from a it's ⁓ you know a foundational level because the pricing model is also really tightly coupled to the sales model as well. And so if you're leaning a little bit more enterprise heavy, you're gonna be like, yeah, it's gonna, you know, what we're gonna give away for free might feel a little bit different. But anyways, long story short. ⁓ You I think in general the trend that I see with products is you want y a user to be able to try your product to the point where they get excited about it. And so then when it comes time to pay for it, they're making they're making a value decision at that point. And if you if you have that process that or that decision point, if it happens too early in the process. And you don't, regardless of what the product is, you don't give users a time a chance to ⁓ establish what the answer to that value question is, then their answer's gonna be no, probably. And I'm not going I'm not gonna pay for it because I don't actually know what it is and what it can do for me. And so I think they still need to take a leap. You got you you do and then there's a you know, and then there's the matter of what's the
speaker-0: Hook ⁓
speaker-1: Okay, I think it's valuable, but what's the jump? You know, am I you've we we're talking about this actually with Riverside, right? It's like, well, you're incentivized with tools like Riverside to pay annually. So I'm gonna add another person. Well it's not just twenty dollars a month, it's like, well, it's a hundred dollars and then I gotta use this product for a year. So that's a different sort of value calculation than ⁓ it's Nine dollars a month and I can use it for a little bit longer and see if it gets me there. So there's like all these levers that you have to move just to see what works right for your particular product.
speaker-0: So I know we started way off topic on on that one. ⁓ getting back to your workflow. ⁓ tell me, yeah, so so at a foundational level, we wanted to explore concepts of rebuilding tools like Confent tool using AI technology, agentec workflows, and things like that. And the reason that we wanted to do that, obviously, is because now that this, you know, technology is available, we can take advantage of this. The it really kind of upends the buy versus build parish paradigm. ⁓ so I'm curious. You've obviously decided like we're gonna roll this product and and and you know play with it. Talk to me a little bit about how you get from that concept to what you've got today.
speaker-1: Yes, okay, okay. This I think This will make it ⁓ definitely more real. And I I will say, I don't know if we've talked about this before, but I have embraced being the crazy person in my neighborhood and it it actually really comes in handy. I did another thing this week on a related little tool that I was trying to build and I actually scheduled a meeting with no one, but I went into a different room, put the headphones on and started talking to my computer. So I've done this in a few different ways. But I will say that the typical pattern here is, okay, let's take this one. I talk to you, Taylor, and you're like, all right, this is what we think we want to build. And then I will say, all right, cool. I have an idea on this. And typically what I've been doing lately is I go on a walk, put the headphones in, and I go into voice mode in either ChatGPT or Claude. And the voice mode cannot go deep, technically speaking. It's I keep saying, like, can you research this and then let's talk specifics? And they're always like, Yeah, I think you're right. What it does, what I've found this out, I I don't I can't remember if we've talked about this or not. it's like just by having to explain what I'm thinking, it gets me further. Yes, that's what it
speaker-0: There's no tools, it is the tool. Yeah, it's rubber ducking. It is. That concept of ⁓ you have a complicated problem or a challenge, ⁓ just explaining it. And the idea that the reason they call it rubber ducking is because you're supposed to have like a rubber duck on your desk and you look at it and it personifies the the interaction and you kind of explain, Okay, I need to do this, this, this and this and simply by virtue of talking through that you can usually be like, ⁓ that don't work. I got I need to come up with a different strategy.
speaker-1: Exactly, exactly. So I I tend to I do that with the my rubber duck is the AI agent in voice mode. And then I g I get back home or back to the office and I kill that and I kill the duck and go back to I I've even had it a few times it's like yeah I can't I can't do that for you. You should go back to writing mode or chat mode or
speaker-0: Killed the doc.
speaker-1: It is. Like, okay. So go back to chat mode. I'm like, all right, now give me the handoff doc. And at that point, it's basically just a summary of what we've talked about. And then I take that handoff doc and I move over to generally I I switch between Claud Code and Codex, but I lately I tend to use Claude Code for the and and the CLI version, Cloud Code CLI to write the code and then I like using Codecs to test the code. Well one I like having different LLMs created by different providers and putting them against each other because I think they're just trained in different ways so they look for different things. So even if I can take I can take another claude instance that has no context and be like, go audit this and it's gonna come back with some stuff that the other agent
speaker-0: Yeah.
speaker-1: the other Claude agent didn't come up with, but then Codex trained totally differently is gonna come up with a whole different set of things.
speaker-0: Do you use a plan? Do you separate the writing of code from the planning?
speaker-1: Yes, well usually contact speaking, yes. So what I've got that handoff doc. And I feel like I totally changed ⁓ my flow when I came across you you know Matt Pocock. Matt Pocock's he's an educator, developer educator, and he's he's spectacular. He's built a number of different courses, but he recently had this skill called grill me go viral. And and it is more I was doing some version of that, but watching that video from Matt was like, ⁓ I think he's articulating it in such a better way. And I kind of combine my history with his skill and I essentially do this. I say, ⁓ grill me, ask me questions until you are ninety nine percent confident that you can build the entire MVP without asking me anything. Wow. Don't care how long it takes. I've done this with Claude and with Codex. Codex is like, cool, here's a hundred and fifty questions. Go talk. Claude is more like, here's four at a time. Here's four more at a time.
speaker-0: Are you familiar with Andre Karpathy's auto research loop? ⁓ it's similar. It's basically like a harness that encourages the agent to come up with a testable outcome and then iterate and iterate and iterate and iterate. I it'll go for hours until it achieves that outcome. ⁓ but it constantly is like refining its con context and awareness.
speaker-1: No.
speaker-0: of the problem statement as it as it executes. ⁓ It's kind of like a more sophisticated version of Ralph Wegum loops, you know, which is sorta reminds me of what you're talking about there. Like the I I forget how you phrased it. You said ⁓ do this until you're 99%. ⁓ grill. That sounds delicious. okay.
speaker-1: Yeah, that's that's what ⁓ Matt called his skill.
speaker-0: Okay, so and then in terms of the different models that you would use, so you're using clawed code to write and you're grilling your you're having it grill you to plan. Yeah. But what are you are you planning with like the biggest, beefiest model? Have you tried Fable yet? ⁓ is this more of a
speaker-1: ⁓ usually lately it's been opus for the planning. Just w I mean, it was just what, yesterday that Fable opened up again. Not this last time around. I did the first time I had it do some research at a pr did pretty good. Yeah.
speaker-0: Have you done anything with it yet? You're gonna love this. So last night, just for the heck of it, I thought, what the heck? You know, I'll I'll have Fable see if it can port my website from Contentful to Dado. Yeah. It nailed it. Dude, it nailed it. So it largely did the planning and it's it and like I went ahead and approved it and then a couple minutes later I'm like, ⁓ wait, I don't want Fable to do the work. So I moved it to Sonnet, I think, after the plan had been kind of crafted. But it used the Contentful CLI to export all of my data and all my assets. Nice. And I I don't even think it exported it. I don't know how it did the assets, but somehow it ported them all over and and imported them into data. A hundred perc I mean it was crazy. And I yeah, it it built all the models w almost one for one. It made a couple of adjustments where it made sense for for data over contentful. But yeah, it was pretty impressive.
speaker-1: Yeah, that's what, weeks, months of work in the past?
speaker-0: It is kind of impossible to quantify. But you know, here's the thing. It's like it I had to answer all the questions twelve times out. I mean, it grilled me at the start, like through the planning process, but you know, it was unbelievably fast. And it we have the most long running saga about this stupid website. You know, we build websites for ⁓ for a living, for a business and and building your own is often the hardest thing to accomplish. Because you know you're constantly deferring that work and other things blow up and you had a bunch of people with opinions in the room and you you have this emotional connection to the brand. It's it's been a a real fun job. But I'm I'm gonna push it through. Anyway, more on that later. okay. ⁓ so tell me what do you use for design? So like are you just YOLO in it? Are you pointing at a tailwind repo or what's like how how are you
speaker-1: You know what I've been doing lately? I really like in fact I I was just doing this earlier today. Is I have I mean to be determined how well this works, but it's usually some form of this. We've got Claude cooking on the the MVP, and I usually say don't worry about the design yet. Just let's make it work. Because I I think depending how beefy that MVP is. It can get pretty distracted along the way. In fact, this is what I ran into when part of the reason I don't have a demo this afternoon is after we stopped the stopped recording this morning, I fed some of the updates in there and I was like, ⁓ something feels weird. I think your mental model is like totally wrong on where we were originally going with this re architecture. And I was like, So either either I spoke it wrong originally or you're distracted. And I was like, let's let's take a step back. And it went back and it read the whole planning doc that we got from the original re-grill me session. And it was like it was what I wanted was their verbatim. And it was like, ⁓ yeah, ⁓ my mistake. You're absolutely right. So we got so distracted that I'm like, you gotta narrow the focus as much as possible. I a a classic Claude phrase is thin slice, you know, give it that thin slice of stay focused, build the narrow path. So it's doing that. And either while it's doing that or when it's finished with just an early version of the build, I'm like, okay, cool. Now I want a design handoff doc. And I will what I will take that into Claude Design. And I generally I'm not really planning to use any of the output. ⁓ at least not explicitly, because I just don't wanna I don't wanna iterate forever in the design tool, but I think it's Because the design tool doesn't care at all about the functionality, you can get it really focused to like iterate really fast. And what I tend to do is say, Hey, I want five completely different designs that serve this MVP. And then just because I have so much history building designing websites, I can be like, Yes, take this piece and this piece and this piece and we'll come up with something that is unique. ⁓ That's a challenge too. So were there is a ⁓ I should an article I should have sent you yesterday where I think it was GLM five two is now it's it's that plus fable are the top when it comes to design uniqueness. Yeah. Because yeah, a lot of these designs end up being homogenous. So I think you really have to have a you gotta have an eye and an act for it if you're gonna make something truly unique.
speaker-0: I completely agree. sorry, have you played with Claude Design yet?
speaker-1: Yeah, so that's what I was saying. Is like while I have Claude code rolling with the functionality, I'm over in Claude Design for the visuals. And then eventually I I bring s I'll just take screenshots and be like All right, let's work from there.
speaker-0: ⁓ okay. You played with Figma Make.
speaker-1: No, ⁓ not since it was announced, which has gotta be almost a year ago.
speaker-0: ⁓ yeah, it's yes, it's been it's been a minute. Have you played with Reloom or ⁓ UX pilots? And then Google has one too. I forget what it's called. ⁓ Stitch maybe? Does that sound right? Google Stitch? anyway. Yeah, design is kind of this weird, elusive thing, and I I really have been impressed with Claude Design.
speaker-1: No, I don't know either of those.
speaker-0: ⁓ over the others like UX pilot, ⁓ I want to love and I it is unbelievably consistent with the designs that it cranks out. It's like, man, can we get any Yeah, it's just like follows it it's like they have one template, you know, and everything that comes out of it looks just like that same thing. maybe that's I was probably not using it to its full capacity, but I ⁓ reloom is R E L U ⁓ E
speaker-1: Yeah, what maybe I need to introduce some of that, ⁓ know, up the entropy. I haven't tried that since we talked about
speaker-0: Well so that's ⁓ you know, ⁓ we were t we talked about that a little while ago and I I have found that to be a fascinating concept and I'm actually I'm gonna meet some of those guys here in another couple of weeks and I'm really hoping This guy that I talked to about the entropy thing comes to the conversation 'cause I wanna pick his brain. But ⁓ as I mentioned this morning I've been playing aro a l a lot with comfy UI and comfy UI has got a seed value. So basically when you pass it into the sampler that does most of the work. Which by the way, do you know how the how stable diffusion works?
speaker-1: No.
speaker-0: It is so cool. And if you think about it, it makes a ton of sense. So it will take I mean it it'll take like an image or something, you know. Let's say that you wanted to make an adjustment to an image of yourself on a beach and you're like standing there and there's some guy in the background and you want that guy gone. My understanding is that you can use stable diffusion to make that adjustment or any of these systems. And what they do is they basically take that full res image and add noise to it. And ⁓ so I think that gives the AI a fuzzy canvas to work with. Like I can see you, but there's all these little like staticky things over you. Right. And that gives me a little bit more access to like, ⁓ I'm gonna smudge this part in the background and kind of fuzz that out and then from there it starts denoising. It starts removing noise. Yeah, it enhances, right? And by it's doing like different I think it's just doing like
speaker-1: Hand second.
speaker-0: you know, we're gonna do a pass and take noise out. We're gonna do another pass and take noise out. We're gonna do another pass and take noise out. And through that process somehow it it artificially can, you know, repair and replace certain portions of that image. Which I thought was fascinating. And so w anyway, comfy UI, one not related to the noise piece. I just found that interesting. ⁓ you create a seed. So it's a random value for the sampler. The sampler is what's adding the noise and removing the noise and all like doing all that stuff. And and my understanding is that seed is like it's like the entropy that's being fed into the sampler and without that you're gonna get the same thing every single time. Or if you don't change that value at all, you always get the same image over and over again. I'm sure I bastardized that entire explanation of the way this works, but I'm learning. So please be patient with me. I will ⁓ you know, be an expert one day. That said, ⁓ okay, back to the so we're t we're still talking about your workflow ⁓ and we're getting a little bit close on time. I just want to make sure we don't go too far into the or too you know, too up close to the hour. ⁓ your workflow, we're using Claude design. You mentioned ⁓ you're you're planning first primarily in voice mode. Using Chat GPT for that, or do you like Claude's voice mode?
speaker-1: Mm, you know what? I I feel like my habit is to go to Claude just 'cause that's it's my default. I think that Chat GPT voice mode is better. ⁓ enjoy it more.
speaker-0: I hundred percent agree. Man, I can't stand Claude's voice mode. And I don't know why. That's fine. And also why can't it's AI and it's so smart. Why can't it give me a voice that I don't want to punch? Like through the
speaker-1: It also sometimes when the network is bad it slows down and so it changes the voice.
speaker-0: Yeah, dude. What is up with that? Sometimes it'll be like Bob shows up. It's like, dude, who is this? You don't sound like she sounded yesterday. ⁓ anyway. ⁓ all right. That's great, man. Any other ⁓ so sorry, so you're writing cloth
speaker-1: ⁓ yeah.
speaker-0: code, you're using Claude Code CLI and you were talking earlier today about building some conductor thing and all that, which is really cool. And then lastly you said you test with codecs. What do you mean by that? You mean you're actually writing unit test or you're just using it to review the output from Claude?
speaker-1: Yeah, this okay, I keep going back and forth on this one quite a bit, and I feel like I'm gonna give you a different answer in a couple of weeks if you ask me that same question. But today my strategy is generally I when I'm especially when I'm working synchronously, I say I want Claude to set up unit tests. And like if you if you can do something that takes takes less than thirty seconds, do it. So if you want to run unit tests to validate your work, fine. If you wanna run a build to validate your work, fine. Otherwise, let's rely on a combination of me human and our production workflow. Which isn't always Netlify because some t you know, we've g we're Building some things like the and the AI engine is local to this Mac Mini for now. ⁓ so some of that's just gotta be like, I'm just gonna test it. Especially in these early days though. I I don't know. Lately it seems like Claude's like, okay, cool. I'm gonna install playwright and then I'm gonna run all this in a browser and I'm gonna record the whole session and I'm gonna look at it. I'm just like, just push the code up. It's gonna deploy to Netlify. I'll take a screenshot if something looks weird. It's fine. Now I think In a couple of weeks, as what I what I think is is gonna happen with this tool is we're gonna get to a point where we're going to start introducing some concept of loops and automated QI QA. And when we get there, I think we're gonna need some sort of like visual component to it. But for now I'm like, ⁓ I don't it takes so long.
speaker-0: QA is the last frontier of technology. Like, nobody has solved this yet. Even I mean, like, the best QA that I've ever had is when a human being that is like methodical and really good at their job can sit down. I don't know how they do it, but they find the craziest stuff in the most bizarre places and it's super helpful. I've never seen a software package solve this. I mean, we have Playwright, we got Cypress, we got, you know, all of And yet it I don't know, Selenium's been around for twenty years and you still can't figure out how to use the damn thing in a manner that is consistent and doesn't it's so brittle. Like how is this so brittle? That card on the homepage never moved for me, and yet Playwright can't find the damn thing every third time it runs the test. It's infuriating.
speaker-1: Yes, a hundred percent. We did the same thing when I was when I was working with you all for the the the site migration from Gatsby to Next. ⁓ yeah. One of the things I did was take scre I I had AI run and take screenshots of production and then move the code, take screenshots of the new yeah, yep, staging, compare them.
speaker-0: Staging.
speaker-1: And it was something like if there was a you know, it was like half a percent difference flag it.
speaker-0: Dude, Rudo just built one of those too for a client. It was like an onion skin kind of thing, you know, with like a slider where it had like the old version of the website adjacent. And I guess it was the new version was overlaid on top of it. But there was a there was like a onion you know, the onion thing, right? Like basically a little slider in the middle and you could pull left or right to expose the hidden part on the either image or whatever. I thought that was really cool.
speaker-1: Yeah. That's it.
speaker-0: I also love do I love those little UI tweaks and things that you can do. I am a New York Times subscriber almost exclusively because I love the visual t storytelling stuff that they do. I don't even care what the content is. I'm like fascinated. Every time I open it up, I'm like scroll through there and something jumps out and they do all this parallaxy scroll it's called scrolling t s sorry, scrolly telling. Have you ever heard of that term? No. Okay, we'll talk about all this stuff next time because I think I have struggled for years to like amass all the You know, it's like you see these techniques on websites and you're like, What is that? How do I search that? I don't know I you know, I I want that on my website. Or I want my client needs this. This is such a great way to convey that information. And heck half the time it is just a challenge to find what you're looking for. Yep. Now that's it. We're getting close on time and we spent a lot of time on your workflow. I really appreciate you taking me through all that. That's pretty cool. That's very similar to what I'm doing in a lot of ways. ⁓ we'll come back to token maxing next time. I think that's a good topic, but we're gonna skip that today. ⁓ let's talk about Momentum Conf. So what is Momentum Conf and tell me a little bit about how that's going to impact your October. I have never been. Okay.
speaker-1: I've actually I've never been there. Have you been there?
speaker-0: Yeah, to say that I almost sponsored it last year, but it was like so wacky, dude. Anyway, Momentum Momentum Conf is a tech focused, developer-centric kind of, but I think it also has some good business insights as well. Conference in Cincinnati, Ohio. And it's been around for several years. And I, you know, I think it may have predated the pandemic. Is that right? I know it's gotten a ton of transit ⁓ you know, a ton of traction amongst the community. It's organized by
speaker-1: Yeah, definitely.
speaker-0: I I don't know actually I is it I I don't think a particular company organizes it, do they? It's just It's a community thing. Okay, very cool. Yeah, yeah. So momentum brings software developers together with ⁓ to sharpen skills, make connections, and geek out with impunity. Okay. That's gonna be ⁓ fun. Okay. ⁓ the conference this year is October fifteenth, and again, this is held in Cincinnati, Ohio, where both Sean and I live. Sean, you got selected.
speaker-1: No it's a c it's a community. Yeah.
speaker-0: To speed, what are you gonna what are you gonna talk about?
speaker-1: So I oddly enough, this is a very similar pitch that I gave to this Cincy AI Week ⁓ thing that just like
speaker-0: The one that you showed up to actually like you've already got the material because you developed it for the last conference, you just didn't actually present it. Yeah. Because you got roped into just a rando panel. Which was clear. It was entertaining. ⁓
speaker-1: I didn't get the Yeah. ⁓ ⁓ yep, so I get to give this for real. So yeah, the yeah, back this and and I ended up recording a video, a long version of me giving this talk, but yeah, essentially the idea we were talking about Rails earlier this morning, ⁓ last episode. And just this idea of I was building everything for myself in the Rails days and then stopped because it was so complex to stitch everything together and now it's fun to build all these things again. So that's it's lightning talk. It's I mean we're gonna get what, five, ten minutes on stage or something like that. So that'll that'll be a quick one. But you also got
speaker-0: I did. So I got a full talk which is pretty cool. So mine ⁓ the title of mine is How I Digitized My Brain, Building a Queryable Knowledge System to Control the Case. So I'm looking forward to this and I I generally am terrified of public speaking, but I do have a lot to s talk about on this topic. What I'm referring to here is a while ago, early like January of this year, I built a system that piped all of my communications through Slack, number one. So like from a communication standpoint, it became this awesome platform. Like this granted, I I don't I'm not like a Slack fanboy, but we have it and it's available to us. And honestly, Telegram sucks out loud. So I like this much more than than Telegram. ⁓ but anyway, and it's also like I have the vestige of
speaker-1: Okay.
speaker-0: privacy inside Slack, even though I know it's probably not, you know. Yeah. It's just like everything. It's like, well, I hope that I hope we don't get breached. Yeah. ⁓ regardless, you know, we've all got two factor and stuff. So we got that going for us. Anyway, so ⁓ the idea is is using Slack as like this central conduit through which all information flows. It help it is is giving me an opportunity to like stitch in all of this additional agentic stuff that happens. So ⁓ I can grab a URL and plug it in there, right? And Slack will unfurl that
speaker-1: I hope so.
speaker-0: Well, I've got a little webhook that will fire against a application and that will go ingest all the content from that URL and it will stitch it into a rag database. It does the input filter like the input output filtering to make sure that you know it's the content is safe or whatever. I think I might be using fire crawl for that. ⁓ so it sort of covers the security or the ⁓ prompt injection stuff that theoretically could happen, which also is a security risk. So just know like some of the stuff comes with a little bit of risk. ⁓ anyway, once that ⁓ RAG database has ingested all those little chunks of content from that article, then I can query it from somewhere else, right? So I built a MCP server for Claude and for any of the other clients that support that. ⁓ I built a ⁓ Well like a Chad GPT custom connector a while ago and so that can talk. And then at the company, as we're building up more and more platforms and products to support this kind of stuff, we built, for example, like our own LM studio clone where you can upload a bunch of sources, you can query it, you can generate, you know, audio outputs and stuff like that. And that too connects to all of this apparatus. So I can say, you know, what's that thing? And discovery, what was the conversation I had over here? And and so that's part of it. And then the other piece is that ⁓ every conversation that I have that is being transcribed. I'm scraping those off of like the Fathom API or the Zoom API and the ⁓ Pocket API, which you know my pocket is my little voice companion thingy. And so every morning I have a cron job that runs and just scrapes those transcripts in. Those also get dumped back into the RAC database. So what I've done is I've kind of created this sort of holistic brain of information. Now, the reason this is beneficial is because I'm a little bit of a space cadet and sometimes I forget a thing or two. And it has been really helpful for me if you and I will have a conversation on Monday and then I'll have 10 more conversations that day about different topics or whatever. And then on Wednesday I'm like, who did I I just talked to somebody about this thing. You know, I think we probably all suffer from that. Like in this age, it's information overload. And so anyway, I'm excited to kind of talk about the system and like the good parts of it, kind of the clunky parts of it. And it's also giving me a good motivation coming into October to like go back and look at it and think about it and like what's working, what's not. You know. So often we build these tools and they'll be like ninety percent done. And that ten percent is sort of annoying, but you'll just sort of live with it forever. And so ⁓ it's giving me a a newfound encouragement to go back and clean up that the cruffy the the crusty bits.
speaker-1: Mm, I like that. That's that'd be a good segue for for next episode too, because I I've been thinking a lot about that. what's the concept I was just I was just playing around with yesterday was I th I think that the one the whole like one shot first fifteen minutes of an app is like it's interesting. It's magic. It's also no one's going to actually keep using their app that they built in fifteen minutes. So I'm kind of thinking about it as like your first fifteen minutes or your first hour is magic, your first day is what you need to get to the MVP and your first month is what you need to make it actually usable. And so yeah, I I would love to talk. more about that and just like is there a strategy you have to to to get to knock out that last 10% or so
speaker-0: I don't know, but I will tell you this ten percent to your point, it's like a the it's like a inverted yeah growth curve or whatever. Like you get what do they say the last twenty percent of a pro it takes you finish the last eighty percent of a project and the last twenty percent of time available. Or what it what Parkinson's law where it's like the work expands to fill the available space?
speaker-1: The other way around. Yeah, but I think it was it takes eighty percent of the time to get to finish the last twenty percent or something.
speaker-0: Exactly. We're saying, yeah, you know what I mean. At least I'll probably put it's true. I don't know. ⁓ it's interesting because lately we've been getting a lot of vibe coded apps. Like customers are like, hey, I vibe coded this. Make it work. And you're like, what you did what? Like what like one of the guy ⁓ man, I feel bad because he might listen to this and I don't want to like throw him under the bus. But my favorite was somebody whose laptop has to be open in order for the app to rem to be updated to like
speaker-1: Something something like that.
speaker-0: He's g he's got this whole five coded system that will update the dashboard on this thing, provided his laptop is powered on and open and is working and it's like That is not scalable. Right. ⁓ but it's fine. I mean, you know, I it's really exciting actually. I think it's super cool. I love this stuff. And listen, I really appreciate you taking everybody through your workflow because the way you thought about that, the different pieces, the different segments. I mean, this is really important and valuable stuff and and it comes from years of of experience doing it by hand and now you've got these batteries included sort of solutions, but the process hasn't really changed that much. You still go through all those those steps to get to the the finish line.
speaker-1: Yes, yes, absolutely.
speaker-0: All right. Well, Sean, listen, I really appreciate the chat today. ⁓ thanks everybody for listening. This is the ⁓ Blood, Sweat, and Tokens podcast studio in downtown Cincinnati, and it was really fun to see in person. All right, thanks everybody. We'll see you next week.
speaker-1: Yes, likewise.