- Platform: meet
- Source: Read.ai 01KYT7JK0G73CKBK5VAKCD64XF (via Ben Ornstein’s account, imported 2026-08-06)
Transcript
Blake Emal: Meet you, yeah. Michael Houck: You guys, you guys have never been on a call before, have you? That’s crazy. Oh, yeah, that’s true, crazy, crazy. Um, okay, I’m having this finish, guys. I’m an engineer now, it’s official. Um, I’ve totally, Rafael, you’ll be excited to know this, I think. Um, I automated the back-end processes for our SEO agency, Lore. So, like, you guys are working on finishing the site, but like, in terms of how we draft the articles, how we optimize client sites, and how we get generate backlinks with press releases, I just automated all that. Rafael Bruno Siqueira Da Silva: Oh, that’s really good. Yeah, it’ll really help us speed things up. Michael Houck: Yeah, so I think. We’ll see. I’m going to test it out. We’re going to test it out with Jolt first and see if it’s good. But if it’s good, then we’ll roll it out to clients and it’ll really good to have that. So I’m excited about it. We’re waiting on Ben, I think, right? We invited Ben. Blake Emal: Is other Blake coming too? Michael Houck: No, but maybe he should actually. Let me, I just added him to the invite and I’ll DM him too. Well, hey, Ben. Hey. Okay, thanks. It looks like that came through, I think. So that should fix the issue. It was just confused because of the Mac mini situation. Ben Ornstein: Yeah, I just didn’t have that ord on that ownership. Ben Ornstein: So now it should. Michael Houck: Okay. Michael Houck: So let’s see, let’s see, let’s see. Yeah, okay, let’s just talk about product direction here for Casper. So, Ben and Rafael, do you guys want to give me some context? I think mostly probably Ben, give me context on the state of the product today. What we’ve built. I mean, I know what we’ve built for the most part, but just like what’s like leading-edge stuff that we’ve been planning to do next, and why has that been deemed to be valuable? And then we can talk through how working with a Hermes agent might change that, or if it should at all. So, would love just Ben if you give us lay of the land to get started. Ben Ornstein: For sure. So, we’ve got a few different agents that kind of make up. Up a whole team. If we don’t even consider, like, well, we’ll get into the fact that not all the agents are currently live because the way, I don’t know if it was the way that I built them. Likely that is it. They were too expensive. They were just burning a lot of tokens. I’m sure it can be optimized. That was probably my fault. But regardless, at full capacity, assuming all the agents are live that are fully built. The team is Casper, the ghostwriter, Eddie, the editor, who is trained on, you know, everything that kind of Blake gave me basically, plus a lot more over time, obviously. He kind of edits Casper’s content. They’re looking at there, we’ve got Trent Spotter, who is that he looks for his biggest accomplishment when he was live was he would find. Relevant posts for the clients to engage with. He’d pull like 25 at a time. These posts would get thrown into a client thread in a Slack channel. And then Casper would automatically start drafting or skipping posts if they weren’t relevant. Trenton was pulling, based on relevant things for each client. And then Casper would draft, Eddie would edit, Casper redraft. Et cetera, until they hit like an approved engagement. And so that was working extremely well. And there were little bugs here and there, but it was honestly pretty impressive when it was at full steam. And I think the comments were high quality. So those were like the three core agents for a while. And then we added Connie Tacht, who is going to be like the sort of like, account manager agent. Hey, these were all your names. I liked all these names. Michael Houck: I know, they’re so funny. They’re hilarious. Ben Ornstein: And so her job was kind of to like understand the lay of the land, be like an agent version of like myself and Blake Stewart, basically. And so was, I didn’t get too deep with Connie, but but when I was using her, she was responsible for helping with like onboarding, surfacing the relevant documents. And then a fifth agent that we were going to build as well, but didn’t really red line. We didn’t get into. Michael Houck: Yeah, what was he supposed to be again? I forget. Ben Ornstein: He was supposed to be the client advocate agent. So, like, a very quick example for everyone here. For Ben Gusberg, our client, who a lot of you know pretty well. If we wouldn’t use a actually, it doesn’t even matter. We wouldn’t talk about a competitor’s product, but Casper, for example, unless he specifically in his infra was told do not comment on this person, this CEO, and then 20 of those, unless we listed that for every client. Basically, Casper would be drafting content if Trent found something relevant. If it was something in the clients, okay, I’ve made this way too confusing. Rafael Bruno Siqueira Da Silva: I think you can use the example of Burger King and McDonald’s. Ben Ornstein: That was That was the one when we talked about this. Okay, Burger King would never comment. Thank you, Rafael. Burger King would never write content about. Right? They wouldn’t want to do that. But unless Casper was specifically told that, he’s just going to write the content anyway. So the client advocate agent red line was supposed to come in there and be like, hold up a second, you would never, the client would never write this. And so that was be his goal at a high level for each client based on a whole lot of information. So I think that’s a decent lay of the land. I didn’t explain everything extremely well, but how’s that? Michael Houck: Yeah, so what were we looking to do like next? Like looking forward, what was assuming we got Read and Connie kind of fully up to speed, what were the, what were we missing still? What were we looking to do next? So I have a good handle on that before we dive into what we want to actually do next. Ben Ornstein: I think that actually rounded out like the Casper and friends team. I don’t think there was any other agents we were considering. I think the next, the biggest step. Was going to be like, okay, how do we optimize everything? Like, it’s kind of patched together and kind of working, but it’s way too expensive. There were still way too many bugs. And those were, I think, like, that was the most important fix next. Michael Houck: Yeah, I agree. The cost thing was, what caused us to initially revert to just Casper and Eddie because we. I think our max one day, we spent like $350, right? That was, yeah, okay. Yeah, that was, that was, that was not a good day. No, but um, okay, so that's fine. Um, so then, okay, let me give you guys now a little bit of context on what I've done with Hermes over the last, I don't know, two or three days here. Um, so what we did was I had Hermes go through and just for context, the reason are all familiar with Hermes, like why we would use that over OpenClaw? Okay, I don't need to do that. Okay. But I had it go through the architecture that up that Ben just talked about and basically identify like what like it just took like a like a QA sweep through everything. And I didn't want to mess up what was live, Ben, because like, you know. I don't have, I think, like, full authorship over that, or at least I didn't until today. So, I didn't want to mess with that. And also, it's like OpenClaw versus a Hermes agent. So, what I did was I had Hermes basically not start from scratch, but like start from zero and use the OpenClaw setup and MD files as scaffolding to basically fill in the gaps and also reorient some things. And so there is no team on this side, on like the Hermes side, but it's like Hermes like interacting with Hermes, like Hermescept. And so I don't know if what it has done so far is going to be better, but I do think that because I'm running it on Fable for comments, that the comments that will actually be. Be pretty high quality and specific enough to people's voices to make them not be met. Because honestly, the biggest reason that we have people churn is because of growth. And the biggest reason they don't see growth is because they don't do comments. And so we want to prioritize that. And so, I figured, you know, comments are high visibility and also high importance per word. So, I figured spend a little bit more on Fable. And with the way that Hermes got in there and architected this new version, the estimated cost will be like under $100 per client per month, even if they’re doing 100 comments a week. So, from a margins perspective, it’s actually quite attractive. And with the higher quality from Fave, it’s like interesting. I was able to increase margins and decrease margins and increase quality at the same time. So, I think that was just a case of like it just took an architectural pass. Like, Ben, I think when we were kind of wading through the water here initially, we basically just like were like, oh, that’s a problem. Okay, let’s solve that problem, right? Like, we band-aided on top of band-aids. And so I think it just needed a fresh AI set of eyes, at least. Um, and so I don’t know if it’s going to lead to like better content and tooling, uh, Blake, for you or when we uh bring the clients over to you from everyone else, but um, I do think that it’s definitely going to be a better experience for common. So, we’re test that, Ben, with you. Um, I’m going to get that finalized today. But then, Blake, I think that the important kind of next thing here is for you, which is like, we’re going to transfer a lot of these clients over to you in the next, I don’t know, a couple days, right? Like, really, Monday, we’ll start transitioning folks. So you’ve mentioned that your internal like clawed setup that you have is comparable both in terms of quality of output and better in terms of usability for you. Than talking to Casper Nettie and Slack. So, can you articulate what about what you have is better in terms of how you use it and how you set it up? And also, then kind of what your recommendations would be for what we do with kind of the next iteration of the Casper idea? Blake Emal: Yeah, that makes it sound like my. Setup is impressive in any way, because it’s not. It’s not that crazy. It’s just that I’ve been using it for so long that it just has more of my style built into it. But it’s also what what I’m used to. So I kind of like having it in a separate place where there’s no way that the AI is gonna possibly do anything in any other channel, which hasn’t really happened with Casper. And I don’t think that would happen, but there’s I like that part of it. More so than anything, it’s just the fact that there are foundational things that Casper hasn’t done well that need to be fixed. And then there are advanced things that I would like it to do down the road that would make this much more possible for me to have more clients going through me. So, you know, for the foundational things that like Casper would really need to do, number one would be stay on voice for the client. It’s, and this kind of blends in with some other stuff, but because we have so many clients going through this one agent, I’m talking mostly just about Casper, not about any of the other ones. It forgets all the time, or it was like which client I’m talking about, even if I trigger it every time. So I’ll like try it in Slack, I’ll trigger it for what I need, I’ll prompt it pretty in a pretty long, in-depth way, and it still forgets a ton of stuff that I’ve given it before. And said, like, hey, save this to your project knowledge. This is not just a preference, this is like an actual guideline that needs to be followed, especially for Gusberg as an example. Like, there are a ton of things that I’ve nitpicked and said, this can never happen again. We don’t say this word, we don’t use these things, don’t do these formats, and then it’ll just forget it. And I don’t know if that’s because there’s just so many other things that it has going on that it just comes back to that and just forgets, or like, very strict hook preferences that I’ve said, this is this should be in your project knowledge. You should never forget this or do anything different. And then, like, literally the next day, it’ll do something different. So, that’s kind of for me, foundational table stake stuff that has to be black and white. It can’t be gray area. If I don’t know what the best way to, I guess, when it’s in Slack, it also makes it seem like you can chat with it about anything, and it doesn’t make anything so official. So, like, if I tell it, remember this thing forever, I really mean like, this is a black and white. You have to do this for the rest of time. But I don’t know what it is about it that seems less formal. And so, it’s like, I’m not submitting a rule or saving it anywhere. It’s just kind of like something I say, and it, I don’t know what happens with it. It just ends up not remembering. And then another part foundationally is just that the improved baseline of quality of content needs to be better. Not every client has the same voice, but there should be certain things that are just an established baseline and it will never sound like AIisms. There won’t be slop by nature. And then we can add the layer of voice on top of it again. But it seems like that it’s just a little bit better in my cloud projects that I’ll use as opposed to Casper. So those are foundational things that have to be true. And then I have a list of advanced stuff that. Eventually I think needs to be true for me as well. It’s just not, it doesn’t have to be built until we actually have those things figured out. Michael Houck: What about what you mentioned to me? I think it was last week, or maybe it was a week before we talked, where you were talking about like the user, the way you use it, right? Like you’re talking about the Slack interface versus like, and you just mentioned it to be a separate space. Like, talk to me about that, because that might be something that we set up. Blake Emal: Yeah, if it’s just a separate space that still ends up looking the same, then there’s no point. Like Slack is convenient. I’ve given it a go. I’ve used this, like the Casper side more than I’ve tried other stuff on my own that’s not just clawed. So I don’t think it’s like a broken interface necessarily. The hard part is that if I prompt it and say, I need a batch of three long form posts, in theory, in my mind, that would open up a thread that has three messages. One message per post. And instead, what I get is 56 messages that’s like, all right, I’m generating it. All right, I ran into this roadblock. Then it’ll post the initial draft, but then Eddie will butt in with an edit, but then the edit wasn’t good. And Casper rewrites it or like, and it literally is like 60 messages before it shuts up and I can do anything about it. So, like, that part of the interface. Blake Emal: I don’t know. Blake Emal: There’s something nice about how it segments things and it doesn’t just give me a huge output that I have to read through in one message. But at the same time, it’s got to be way better at just understanding that I don’t care about its thought process at all. Cause I just need to see the content it’s going to give me and then be able to tell it that sucked or that’s good basically. So I think, like, the thread stuff, we’ve got to make that way briefer, and it just has to understand what I actually care about and give me that. And if it’s going to, if we’re going to build on an editor agent or whatever, I feel like that all should just run without me having to see the whole thing because it just makes it more work for me to have to parse through everything. Sometimes, what I see is also like 13 messages in, there was an edit. And I liked that part more than the final output that it gave me. And I had, in order to know that, I have to read like thousands of words. So it really doesn’t save me that much time. Does that make sense? Whereas with Claude, like, yeah, sometimes it gives me a bad post, but I just tell it that sucked, give me a different one. And it’s just a little bit easier for me to bully it and get something back quickly. So there’s probably something to the interface to just make it simpler, but Blake Stewart: to jump in there too, I think it’s kind of creating clutter too. Like, if you are working on other operational stuff too, and you’re just trying to read your threads, you have to scroll through hundreds of messages that are distracting and kind of, you know, you’ve got other stuff going on, other people talking to you. So I think it could cause some kind of disorganization on that front, too, being in Slack. Blake Emal: Yeah. Michael Houck: Yeah. I think it’s a good point to say that you don’t really care about its thought process. You just care about. The output that it’s going to give you, and so it’s like, you know, whether we do that in Slack and just don’t show the thought process or whether we do it in some other space where like the thought process is like hidden by default but expandable, something like that. Blake Emal: and honestly, like, I care the thought process stuff has been good to watch for a while to see if I actually agree with it. And sometimes I really don’t agree with its thought process because it’s just like, oh, this should be longer for literally no reason. So, I guess it’s been kind of interesting to see that part of it. But as long as the thought process is good, I really don’t want to see it at that point. The other thing is, I’ve been working out of basically one channel for everything. And I don’t know if that makes a difference architecturally with it. But if I just had a Casper per client and it was very, very strict black and white within those confines in that channel. I don’t know if that would help. I don’t know if it would make it more expensive or whatever, but it might help me too to be able to set up like automations or stuff in there that I want. Because ultimately, a lot of the advanced stuff is it being able to anticipate what is needed per client every week and give me the rough draft of all that stuff, whether it’s ideation, analysis, prompts, or actually creating content. Eventually, the advanced stuff is like, I’m good at all of the baseline things. And now within each channel, I know what each client gets. Like this one gets three posts per month or per week for LinkedIn. I’m going to give you all the prompts and stuff that you need to get voice notes. And then, if we don’t get voice notes, I’m going to generate those three posts for you without me having to prompt it. So that’s like ultimately what’s going to be the most awesome is to have all of those things proactively covered. And then I just go through and edit mostly. Michael Houck: Yeah, I think there’s definitely something to the idea of having like different instances of the agent per client. And I think we probably, I mean, we definitely have the resources. It’s not more expensive to do that inherently. I don’t know. Much it would increase performance, but I can find out. And then if it does, we could just set up the architecture that way. But what are other advanced things that might be useful for us to bake in or test? Blake Emal: Yeah, one thing that I run through Claude co-work is I analyze, I like feed it the weekly doc of, well, I have it go create a plain text document. Of all the last week’s posts and the engagement that each got and analyze patterns against like the past three, four weeks or whatever. So every week I kind of do that and just use that to update the project knowledge of what’s working, what’s not working, and analyze patterns. So for it to automatically do that would be super valuable. That way, if we see something’s clearly not working over time, I don’t have to just go through every single account all the time and figure that stuff out. It’s actually surfacing those insights for me. And I can avoid things that are clearly not working for us and double down on what has. So, like, if we find a great hook, let’s just find a way to rework that. Let’s, let’s, it’ll analyze those patterns. So, that’s one automation. Pretty much all of them are automations. So, analyzing patterns on a weekly basis and then just doing stuff with it, not requiring me to do anything else with it. Number two would be like the prompt. Portion. So, that would be the first thing that happens in my week is sending clients prompts. Not every client gets them, but for the ones that do, and probably all of them should, even if they’re not going to respond, it should proactively propose weekly prompts that are not what they’ve already been given before. It should be able to read all of Slack every single message and never ask the same question twice and do good prompting, which it doesn’t give good prompts when I ask for it right now. It gives prompts that are like not that open-ended. But also really, really specific and hard to answer. And so, like, there’s some fine-tuning there to give good prompts that are simple to read, but actually generate a longer form open-ended response. So, on a weekly basis, it should generate all that stuff without me even having to prompt it. And it’d trigger every Monday and just tell me, like, hey, push this button and send it to the channel or just do it manually. Number three would be proactively generating content based on inputs from Slack. Clients do respond to those prompts, it automatically does that. I think it’s kind of done some of that stuff before to push to Notion a little bit, but having that really built out would be awesome. I guess another portion of this, if we pause in that, some of these automations end up pushing content to Notion that clients can see, even though most clients don’t really look in there, but it drops stuff in the backlog that isn’t really ready and is clearly AI. So that’s a consideration. Michael Houck: Maybe there’s a like we don’t want to do that, like, yeah. Blake Emal: I mean, I don’t really like putting content like the other day, I put some backlog stuff in there for Robling that I didn’t want them to see. Kind of tweaking and working on it, and he went through it. And I was like, that stuff wasn’t even remote, that was just an outline, you know. I don’t want them seeing that stuff. Michael Houck: And that’s not like an outlier experience. Like, we’ve had that happen 10-15 times with various clients because they’re actively going in there and looking. Blake Emal: So I don’t know if that means one thing I did for my business, PageTear, is I have like all these separate boards for all the clients in Notion, just like we have, but I had one internal request thing that had everything combined into one, but nobody else had access to it but me. So just like one humongous thing that had every client stuff, and it just pushed all of it in with relation things in Notion. And they can’t see that stuff. So I don’t know that that works for me, but that might be a little bit of a headache to have a separate thing. Michael Houck: No, I’d honestly prefer that. I think we tried some point up at one point, but didn’t do it right. So we probably should, once we, you know, kind of consolidate the workflow around you, we should just do that too. Blake Emal: So we can figure that out. But as long, like, I want content proactively being generated if we have context. And then for 90% of the content where we don’t get any inputs. We still need to generate content. So, if later into the week, nothing’s been happening in Slack and clients haven’t given anything, we should also have a trigger that just says, you know what? Create new content that we have not done before, that’s in our scope and in our pillars and everything, but just make sure it’s not redundant. Something like that that triggers so we can create more evergreen content every time we don’t have inputs, which is probably going to be every single time because that’s just how most clients operate. But we need that as well. And it should be better at like following those guidelines. And another thing that I’d like to see tonally, I guess, or just strategically is these agents being okay, being a little bit more adventurous with stuff and trying stuff out. Because most of the time, it’s just the same format. It’s a comfortable format for LinkedIn. It’s like 200 words, 100 words, whatever it is. It’s almost the same length every time. A couple arrows to break things up, but it’s pretty much the same exact looking post every time. And if I want something different, like a post format that really works is a listicle on LinkedIn. So just like a quick hook that’s jarring, and then a huge list of stuff, one line per item. That stuff works really well. If I wanted to get something like that, I have to send it examples and tell it exactly what I want every single time. And it like never remembers what a listicle looks like, And I want it to be better at having a huge bank of formats and stuff to be adventurous and just try stuff because I don’t want every post coming out looking exactly the same. Michael Houck: It’s funny because we have the viral examples. Library of posts in Notion. I think most of them are X rather than LinkedIn, but I think some of them are LinkedIn. But either way, maybe something that we do, like an action item from this, is we just have it go through and scrape a ton of these posts and ones that are outliers, right? Ones that did well, but not from necessarily big accounts and that aren’t like overly personal story related, but actually they did well because they’re formatted in like a repeatable, templatable way. And we just have, like, I don’t know, like, some little repo of MV files with like different formats in there or grouped in some way. I don’t know what it would look like, but something like that. UNKNOWN_SPEAKER: Yeah. Ben Ornstein: So Rina is actually in process of like pulling together that list of like creators that on LinkedIn that have, you know, have done well. She’s pulling like between 500 and 1,000. We could have it scrape through those URLs, maybe. Michael Houck: Yeah, I mean, we could. I feel like that has other uses too, but we could also use it for this. The big thing we want is like outliers, right? Like, similar to like YouTube thumbnails, like one of like one of 10.com. We want to find the ones that, you know, obviously it’s good to use ones from big accounts because those accounts know what does well and you can kind of trust their formatting. But also good to see, like, what were the real breakouts, right? What were the ones that shouldn’t have been big because they didn’t have a big initial audience and managed to go into the, they broke out of the core and they went into the casual and the new really, really successfully. So we can see. But okay, anything else we should sort of keep in mind as we go into this, Blake, on your end? Blake Emal: I mean, that probably suffices for now. There’s a lot in there. So happy to share some notes with you all. Rafael Bruno Siqueira Da Silva: But let me just check if I understood correctly. The idea of this product that we are building is that clients will also have access to it, right? So they will have like a Slack channel where they could interact with the bot. Michael Houck: In theory, yes. First, we want to nail it with our own internal clients. Like we want to nail it with our own internal process, like make it useful for Blake as Blake is writing for clients. Once we do that to a level we’re confident in, then we can say, Oh, hey, clients, try this out for yourselves in your Slack channel. And if that goes well, then we can say, Oh, hey, let’s sell this for as a subscription for X dollars a month to anyone on the internet, basically. Okay. Okay. I have to go to a sales call. So I’m going to do that, but I got a lot of good stuff from this. I took notes. So we’ll sync up and prioritize some of this stuff, Blake. And I’m going to try to build as fast as possible for you so that next week as we’re transitioning folks over, it’s not. And then going into the week after that, it’s not. We make that as easy as possible. Blake Emal: Cool. Michael Houck: All right, guys. I’ll talk to you. Blake Emal: See you guys. Thanks.