Solving Cold Start GTM with Claude Code (and Zero Headcount)

Speaker 0: Alright. Tell me when we're Speaker 1: good. We good? Speaker 0: Good. Alright. I'm going to talk Speaker 1: today about Ruth, which is Foundation's program manager. So Foundation is an invite only community of founders here in Seattle and to San Francisco. First of all, I say, who here has used Root before? Woo hoo. Okay. Speaker 1: Yeah. So this is a fun time to actually do this presentation because Root actually turned 1 year old this March. March 30 is the first line of the first commit. So Root was built, I tried with 3.5 and I want to blow my brains out. So 3.7 was the first time I was like holy crap I wanted to write all of this from scratch. Speaker 1: So what what is Root? What's Foundations? Foundations now has 300 members as a community. And what's interesting about communities is like don't just think about this as a founder community loyal customers. But what sucks about communities, and it's great, is they're made of humans. Speaker 1: And you know what? Humans don't scale. And historically what's been so hard about scaling a community is that because it's made of humans, how do you actually scale a community? You can't throw in more program managers. I'm sure you've been in these like dead slacks before, these dead like forums. Speaker 1: Like how do you keep it moving? So the thought here was can we actually pull this off with an agent? And part of how this came about was by necessity. So we we ran this path, and Ruth has saved about 59 hours to remember conversations, automated follow ups, office hour bookings, and a whole bunch of other stuff over the past 30 days. And that's only growing like crazy. Speaker 1: The amount of activity, if you look at it, it's like growing 20% month over month within the community. But really why I created this in the first place was office hour bookings. So when it started, we were, you know, a community of 20 people, then 50, then a 100. And said, oh, we're gonna start hosting office hours. And I didn't realize it was sort of like an n plus 1 problem, which is like for every office hour, there were slots within the office hours, there were people who were gonna reschedule, and you have to move shit around. Speaker 1: When I used to run Techstars, there was someone doing it Speaker 2: for me. Speaker 1: And, sorry, Jaren. I didn't quite realize how much work you actually did. So so I was like holy shit. We need to build an agent for that. And that's actually how we started. Speaker 1: We layered on from that into basically creating a knowledge corpus, which is, which we'll talk about in a little bit. But Root is our agent to manage our community. He is super proactive. And if you are inside of this community, you will be shocked because you do get that human experience if you're not just talking to the ether and you're like shit is someone gonna respond? He's immediately like here's everybody. Speaker 1: So here's Root's architecture that was generated from his agent's MD file. So there are some errors. I basically had Claude take the agent MD file to create, a mermaid file which I then fed into Gemini, which I then turned into making it look like it was written this way, which is kind of fun. But to tell you Speaker 3: how Root Speaker 1: works, he has a he has a group of tools that are only accessible depending on your member type. He has personality. He's always rooting for founders that comes in. And then he has the history of the the conversation. He can reply proactively. Speaker 1: He can talk in DMs. He can be part of group chat. But then he also has all of your history with him. So what we do is after every conversation you have, there's a whole other query that creates a compacted memory for him. So we don't have Speaker 4: all of it. And we Speaker 1: try to be thoughtful about what are the things I should think about, some of it persists, some of it we don't want to as much. It all goes into an agent runtime, which runs as sort of a traditional run agent loop where it's like what tools do I have? He'll try to run them in parallel if he can. He'll run them in a row and he'll run until he feels that he has actually been able to solve everything he can. 1 thing I'll say that's amazing is I use tool, which I'm on Speaker 5: the board of, so I'm sort Speaker 1: of pitching it. OBIQ for tracing, but what's so fun is that is all the telemetry, and I built a helper that I can use inside of claw recursor where I essentially just put in the the OBIQ ID, and I'm like, it gave me the wrong output. It should have done this. And it looks through every prompt, every input output, everything that happened. And it's like, I just updated the prompts. Speaker 1: Try again. So highly recommend adding something like that to your workflow. So there's a whole bunch of recovery. Other thing that happens, it has access to the tools I'll talk about in a second. And then it combines everything into a final Kelly, and you can interact with Root via Slack and via email. Speaker 1: So it's not just a chatbot, like they're scheduling office hours. There's things like, holy crap, I don't have access to the building anymore. What do I do? There's also some really fun ones I'll share with you. So for Root's surface area, he has this is actually outdated. Speaker 1: I added another tool today. Root has 66 tools that he has access to, which you know functionally lets him answer just about anything that the community needs. How about do I have time check? Speaker 6: 2 minutes. Speaker 1: Good enough. Perfect. So that's that's all good and Kelly. We get into questions. Like, this thing's service area is huge. Speaker 1: There's a whole admin interface. It's really freaking nuts. I kinda lost my mind. The general thought was, like, I did not wanna do a single piece of work that a program manager had to do. So every time, just like any good engineer would do, you know, it's like if I've done something twice, I've done it too many times, It was like I'm gonna make Speaker 4: the agent do it, I'm gonna make the agent Speaker 1: do it. And over a year, we finally got to the place where I can pretty much handle everything without me doing it. So let's look at a couple examples that I've preloaded and then we'll have some fun. So let's look at this 1 here. So this is a question of someone being like, hey, I'm trying Joe close my first enterprise customer and I need help getting SOC 2 done quickly. Speaker 1: Please help me. Speaker 0: So yeah. But actually that that is part of the fun. Let's say, shouldn't I just use Kelly? Let's, see Speaker 1: what happens here. But so you see it's actually pulling from, a mixture of our knowledge base, conversations people have with community, things that we pulled out. I mean, so this is stuff that usually you Speaker 4: put that in the community Speaker 1: thing and, like, no 1 responded to the community of dead. Immediately takes a year and a half plus of slack that the community has talked about and primes the conversation and then everything kind of goes from there. So here we go, congrats. Regarding Kelly, the community consensus is a bit mixed. Oh cool. Speaker 1: Yeah they're they're super fast but there apparently is a risk that they've been survived as insufficient. So like this takes from literally the conversations that are happening in Slack so that every time he usually jumps in but it's a great way to prime the Conversation. Gambier base that he has access to. So this is a great example of, like, finding people who are relevant to help, finding strategies from our knowledge base that all all links into things you can have to have access to. And I can't say enough about how like I was murdering myself trying to make sure everyone got a good answer, and all founders are way too busy. Speaker 1: And imagine what you're doing with your customers, you want your customers to be giving each other advice, but if you aren't priming it with something that has good answers immediately, you're losing. But then also if you're taking out the community experience, you're also losing. So instead this is an infinitely scalable community member who's always rooting for you and is super freaking pumped all the time. Let's look at another example here. So this is a fun example of tools all being used together. Speaker 1: So I want to pick someone's, someone from K Skill's brain about document extraction over lunch. Are any of you in the office right now? So what this is doing is it has access to cameras and access to Zoom and when you badge in, it can set reminders and so here it lets it'll let me know before lunch which it is decided is actually slightly different time on different days but it's always before noon. But that 1 to let me know who from KCL is actually in the office. And I can even use the memory to be like, oh, by the way, they're in the back of the 2nd Floor if I want. Speaker 1: So only look for those people. And then, you know, you take the whole knowledge base, which is I'm trying to raise money. Who's raised from Excel? Okay. Well, what about Madrona? Speaker 1: Let's see here, Ruth. Like, I'm sure there's some people from the community who have raised Madrona. It could be a good intro. And you should be able Speaker 4: to help me out Speaker 1: with that. And lo and behold, in 3, 2, 1, he certainly Speaker 3: well, my demo, trying to Speaker 1: get the timing. He's definitely not going to error out. I'll blame the Wi Fi here before I actually blame Root. There Speaker 7: you go. It's probably Speaker 1: a formatting error. So I I I use Flash 3 for this. Which with a lot of stuff I'm trying to do, there's a lot of errors. So there ends up being a ton of retries. Make sure every all structured output with Zod gets validated before I throw it in to continue the agent loop. Speaker 1: So I try to be really rigorous about that. So, and the TLDR is like scaling community kind of sucks because scaling people is really hard. But a powerful community is an amazing asset. I think it's something that people have really not thought about for go to market because it just doesn't scale. And we would tell you that you can actually scale this with agents now and it's really fun and cool. Speaker 7: Awesome. What's Q and A? Yeah. Speaker 8: Yeah. How do you prevent Speaker 1: some stuff that's going there that eliminate stuff from the Conversation? Like, you have, Speaker 4: like, some kind of redactors? Speaker 1: Joe so so essentially, like, he has a certain set of of, of tools. And I know what is the full service area of those tools, and the thing I'm smart about is there's admin level tools and other tools. So he he does not have access to things that the user does not, so it's all, like, indeterministic code about what could go into the tools. And then I actually don't even make the agent aware of tools that it shouldn't have access to. So you it just it doesn't even show up in the prompt. Speaker 1: So it wouldn't even know how to to go about that. But what I also try to do is there's no shared memory. I tried messing around a bit with like in a public Conversation, don't reveal personal preferences. That actually did not work well at fucking all. Like, dude was high. Speaker 1: So I brought that back, but I would love to see with like the next iteration of like a fast model like Flash if I can bring that back. Because there are some hilarious things where like some people like what what does he call you? The great like Yeah. Speaker 7: I changed my name to the amazing Dan Moore. Yeah. Speaker 1: So that's what always happens when it's great in public channels but like there are certain other things where you just like you don't want gonna be like respond at the end because we'll try to be helpful. It's like and if you need any other help, you know, with your like, you know, I don't Speaker 7: Root taxes. Speaker 1: Yeah. Like, you know, you know, you don't want that in there. So that that that I don't have a birthdays or 2. Speaker 7: And I've I've tried to get access to tools, and it doesn't work. I can confirm that. Speaker 9: Get root taxes. Yep. Speaker 3: No. No. No. So so, Speaker 1: he he's in the public channels. You can app reply and get full context, inner thread, outer thread. So he's actually really smart at, like, helping you brainstorm and do things beyond that. So full context to that plus all the tooling. And he's also proactive. Speaker 1: So I we we will and this is costly, but it really doesn't fucking matter. Is we will process everything and then if it is a proactive response, there's an extra call before reply that actually scores how helpful to I think this response is. That's done by a different model than Flash. So I actually have, like, an obvious model that's, like, is this helpful? And if it goes over a threshold, we will proactively reply. Speaker 1: Otherwise, I'll just log. Mhmm. That's awesome. Speaker 4: Did you Speaker 5: Joe perform much of Speaker 8: the new, like, Slack, API or app building tools for these kind of agents? Like, I saw Speaker 1: your little tool, like, for Yeah. Joe thing Speaker 8: which is awesome, but I think they have some, like, panel stuff or Speaker 1: Yeah. I I, like, I I want this so it works. A lot of the same tools are also available in email. So because this isn't like a Slack specific business tool, I didn't wanna get too ingrained. But there are certain areas where I will use, like, Slack expand and and compress because the the formatters are different for each channel, but I'm purposefully not trying to get too deep. Speaker 1: But I will say the API has gotten so much better like the image thing. That was a nightmare to actually pull off before, but now it actually works quite Kelly. Speaker 8: Follow-up, how did you pick your placeholder, Lenny? Would you like not ruminating, gesticulating, seasoning? Speaker 1: Claude picked it. Speaker 7: 1 1 more question. Still Han,PhD. Yep. Speaker 9: First of all, Speaker 0: awesome product. Really, really cool. How it works. The second, isn't it that because people start talking more to this bot, they have published less information outside of the DMs and connects people directly. So now the information is inside this direct messages, not out of public and the, like, the overall strategy of getting more, information in the database failed because of that. Speaker 1: So so so group will have access to that to be able to use that information. But I think what's really important to to know here is, like, managing community is a lot more like the realities of, like, managing any ticketing framework, which is, like, you think, like, you talk to your IT department, like, I do so much important shit, I help people, it's really just fucking resetting passwords. So, like, that that is what I Kelly learned here is, like, that's not the case and the value proposition of route being able to begin a conversation, it actually takes a lot of the human pressure away of like, oh man, no one's responded. This better be really fucking good. And like it actually just creates like a Joe, just dumped all this stuff up. Speaker 1: And you're like, oh that's a good follow-up. It's so much easier for you to jump in. Now time will Kelly, and this is an Entertainment, But the vast majority of communities at this scale, like, die and go stale, and we're experiencing the opposite. Joe, I want to believe Speaker 7: we're on the subject. Speaker 5: Awesome. Cool. Speaker 7: Thank you, Javier. Yes. This is AI generated music. Speaker 1: You you don't get to pick your walk in music. Speaker 7: Joe. You don't get to pick it. Maybe maybe in the future. We'll we'll have more at our last presentation about the music. Speaker 1: Yeah. I'm getting ideas there. Speaker 9: Does it personally personally generated? Speaker 3: Are you Speaker 7: able to share your screen? Let me kick him out. Here we go. Here we go, Jimmy. Speaker 5: I think this music should be Speaker 7: in all of, hey, I think, our events. Speaker 3: I like it. I'm soon. I'm just Alright. Speaker 7: You want a microphone too? Speaker 5: That's just for the recording. Speaker 7: That was just for the recording. Speaker 0: Yeah. Yeah. Alright. Speaker 5: Thanks everybody. So I'm Jimmy. I'm a product manager by day, Speaker 3: and I, started building Speaker 5: my on my hikes and weekends. I got really into watches last year, and I'm cheap, so I didn't wanna pay for them like new. And then when I started looking at listings, I found that prices for the same watch would be, like, from 4,000 to $6,000. That doesn't make any sense. My background, I worked in pricing Expedia for 4 years. Speaker 5: I worked on your hands on retail. So I got really obsessed with building a tool that for any given watch, for any number of dimensions, you can basically get a fair market value price. The guy that sold me this watch is using the tool every day. He said it's very accurate. So it's very promising. Speaker 5: So that was cool. But then, I have a bit of an aversion to marketing. I'm a, like, builder. If you know the history of marketing, you'll probably end up. But Speaker 1: my dad is a real Speaker 7: good marketer, so I guess it took, Speaker 5: you know, I had to pick up the family trade Speaker 3: at some point. Speaker 5: And I am 1 person. I didn't wanna spend a lot of money. So I turned my good buddy, flawed, and thought about, okay, what are the channels I wanna target? So, Reddit is a huge community for watching. You can find some watches there. Speaker 5: People talk about it. They get really obsessed. Instagram, obviously, Han,PhD then our good old friend, SEO, SEM, I think is still, is a free channel. So why not? So I built the, skill here to simulate it to basically start creating landing pages. Speaker 5: And I can either tell it what kind of landing page I wanna do or it will do a web search. So I was lazy on this 1. It's probably a little hard to see. But it did a search and found, hey. Let's talk about Rolex Submariner versus the Teeter Black Bay. Speaker 5: Then it will look at my database, basically figure out what is the market value for that for these watches, and then it will create basically based on a template. So this is an example of 1 of them. I also spent a lot of time on copy. So I have an editorial MD, and then I'm using Tropes. MD is a really good file for making it not sound like a guy. Speaker 5: So this is, industry. People love celebrity culture and Joe. So just showing basically all these pages and they're they're now starting to generate clicks. I got, like, 10 clicks a day, but again, it's free traffic. So I'll take it and I'm finally starting to see an uptick. Speaker 5: So this is like a slow burn of it. And then for Instagram, I already ran this 1 here. But I can basically pull in and say, hey, for the for a given landing page, create, a new Instagram carousel or reel. Or just, basically, if I wanna do 1 off, I've started doing those as well. Joe, like, this 1 for the Oscars. Speaker 5: This is an example of 1 of the carousels that's worked well where I'm comparing top 5 watches to the S and P 500. It's a change Joe if I need to start getting the memes. The memes too. And this is starting to pick up, so we hit 300 followers in, like, 3 weeks. Speaker 7: You know, Speaker 5: I'm just part time, so it's not bad. And then there's Reddit. Right? So Reddit is like a whole beast. And people on Reddit hate to be sold too. Speaker 5: So you have to be really mindful about how Speaker 7: you approach them. Speaker 5: So I built the Kelly. I started just going on, watch subreddits and being like, if somebody was talking about pricing, I would use my tool, create a report, share it, sort of get traffic. That's really time consuming and lazy. So I built this Kelly that will basically scan all the stuff to watch subreddits, and then it will prioritize, basically, Speaker 3: top 1. Speaker 5: Someone's asking, like, what should I pay? So that's perfect for us. So then it looks at my database, tells me what what it's worth. I'm still going in and I think this is important to still, like, do the editorial review. I've thought about automating this, but, again, it's Reddit. Speaker 5: And that that has been, like, a good trickle of traffic. And then finally, I have the skill for analytics. So every day, I am running this in the morning and it connects up to Google Analytics and, Amplitude for session data. And the data itself is really valuable, but at the bottom, I have basically, like, key insights of what's going on. So I can use this via back in the cloud to improve the product. Speaker 5: I've gotten, like, a 2 Conversation boost just by iterating on it, or I can change, like, where do I wanna target my my marketing spend or my Speaker 3: my small marketing spend. Speaker 5: So all this is really great, but what I found is, like, it's, Claude produces, like, better than mid content, but it's not gonna be great. So our 2 best pieces of content were more organic. We have 1 Speaker 3: right here. Speaker 5: So this reel got to 5,000 views. I don't know if it's PSL's Internet. Speaker 7: I guess. Speaker 5: This is a total wolf. He runs Mercedes f 1 team. There's a huge overlap with the f 1 community of watches. And he's wearing 2 watches. He's being ridiculous. Speaker 5: He needs a big personality, so this looks like this really drew a lot of traffic. I then boosted it to get a little bit more Speaker 3: viewers and then started Speaker 5: to, like, introduce just all with, Claude. So this is using FFmpeg for video and Puppeteer for the sequencing for both images and video. This did really well. But my biggest piece of content was this. So my fiance told me I can't buy even watches. Speaker 5: I was very sad. And then I but I decided to make a post on Reddit that was, like, not quite a shit post, but it was engagement bait for sure. So I have 2 watches that the watch community dates. And then 1 of them I put upside down. I didn't do it on purpose. Speaker 5: But this was really popular. So I got 700 uploads, 500,000 views. It only led to 20 sign ups. And this is kind of an interesting thing about, like, marketing the community. Right? Speaker 5: Like, I had to kind of bury, like, hey, I'm making this tool because I didn't want to feel like I'm selling. But I'm still seeing matricle traffic. So I think the the big learnings is for data analysis, super useful, for product iterations and for creating the you know, the algorithm demands content, so you have to be creating content every day. But for truly getting hits, you do still need to have the editorial and be thinking about, like, ability. Awesome. Speaker 5: That's it. So are we at the Speaker 8: stage where, like, when you when you think about your landing pages, do you wanna make them agent friendly or agent optimized? Both for where if I walk into my Google Home and I say I want the next submarine or I expect it to go out through a web crawl, Do you are you thinking about kind of like optimizing sort of agents for Speaker 1: recording and progress. Speaker 8: Going to your site? And then do you do any of that when it comes to acquisition of your of watches for you, Sal? Like, are you using tools and agents to, like, what do you find for optimizing for that? Speaker 5: Yeah. I've I've thought about it. I think that it's still too early for me. Joe, like, Speaker 8: I think Speaker 5: we just hit a 100 users, so I did hit my goal, which is pretty great. But, I I think that's still, like, a little bit more I think the traditional channel is still Speaker 8: the best place to get Speaker 5: kind of a hybrid. Because you need a lot of volume to get into the training data. Speaker 1: For the self improvement loop part, do you have any tips on long term goals that you actually want versus just make these numbers go up, Speaker 3: which could just create bad content? Speaker 5: That's a good question. So 1 of the things I didn't mention is every time I run 1 of these skills, I will say, like, you know, I'm sure you've all, like, dealt with with Claude and when it goes off the rails. And every time I'm like, what did you learn from this? And I feed it back into the scale and build it. I think probably the next step is to feed the data of, like, hey. Speaker 5: This is the performance. I do for Instagram and for Google Ads. I have the I haven't built MCP IDENTIFICATION, so I'm I'm just I have a clawed browser plugin to be, like, hey. Analyze my performance and then optimize that piece. But the the self optimization has been awesome. Speaker 5: I can create content much faster that is much higher quality now. It gets it right, like, 9 out of 10 times now. Speaker 7: 1. 2 more. Speaker 8: How are you, Speaker 1: I've been banned on, Reddit a couple Speaker 6: of hours. Speaker 7: How are you how are you driving traffic Speaker 5: to your site without, you know you have to mention your site at some point. Speaker 0: It is a careful dance. Speaker 5: Right? I Speaker 7: know. That's what I'm asking. Speaker 5: Yeah. So I in this post, I just said, hey. By the way, I'm tracking it on on this. On the other ones, I just say, hey. If or like, I wanna know how much this is. Speaker 5: The market value is this. Here's a report for you. And people seem to be okay with that. I have a competitor who's being much more aggressive. I think Speaker 0: he's gonna get a hand. Speaker 5: Mhmm. But it's a careful it's a and this is the thing about community, and I think it was good to blast off. But community is super important. I love this community, but there the thing I didn't mention is, like, 30 people told me I should leave my wife. Speaker 7: Are you using the Reddit API to post or are you just copy pasting whatever the fuck's with that? Speaker 5: Right now, I'm still editing it manually just because I am mindful of the Reddit community. Yeah. But I am I have, like, pretty elaborate scraping network both for Speaker 3: the site and for scraping. Speaker 4: It's gonna Speaker 7: take a while to get access Speaker 8: to the Reddit API. Speaker 5: Yeah. They closed it down. So I just have scrapers running on Oh, Speaker 7: I see. On top of it. Speaker 5: Yeah. Okay. Interesting. That they haven't seen in mind. Interesting. Speaker 5: I tried to sign up. It's like an old sign up page and it says, you know, you can use it. Speaker 7: When they start when I started with OpenAI. Speaker 5: I was like, I tried just a couple weeks ago and it was just asking Speaker 7: for anything. Yeah. You're right. It was before that. Interesting. Speaker 7: The blackout? Speaker 9: Yeah. The the API, it was around that. Speaker 7: Yeah. When they had custom mobile apps Yeah. I think. And then they shut that down and a whole bunch of moderators, revolted through a blackout. It was a and then it all blew over. Speaker 3: It's a big deal. Cool. Speaker 2: Yeah. Hi. I'm Kelly. I'm the cofounder of C people at Xcare. So basically, we work with, patients suffering with chronic care needs and dementia. Speaker 2: And our customers are mostly home care agencies in California. And the go to market that we built was after we, been through a really hard pivot, and this agent actually helped us be our first paid customer. He's a, partner in California. And the funny thing is, our customer are pretty anti AI because in the AI space, in the health Kelly, it's mostly about personal relationship, especially when working with, people who are vulnerable in general. So the funny story is after we disclosed that it's actually AI contacting him, he is totally shocked. Speaker 2: But he forgive us. Joe the go to market lead Conversation email campaign system we built was we wanna, Joe I'm just putting here in in my cursor is that we have the sales pipeline, and we wanna automate the process and family healthcare agency prospect. And I'm researching them and really generating personalized, AI emails. We use Bravo right now, but we might update in our tools in the future and track in the engagement, whether they, like, open the email and Kelly the emails. This, so the architecture we I Aviel, we have to been through a few, you know, like, Conversation. Speaker 2: So the first 1 was we do use a CRM product too, and we get a lease script from, the HCA, which is Kelly with the license. Linking, we use the manual because I'm so scared of getting banned by linking, using some tools. So if you got great tools, let me know. Then we populate the unique leads into our agent. We have a old workflow, so we use the Fire Pro API, using, Comet and launch the campaign, then finally, generating the email. Speaker 2: So we wanna send the scheduled emails in, like, a 7 day retreat campaign. So we send them day 1, then we're tracking the Entertainment, and the same same, you know, 2 days later. So you think about it's a long running agent, and we managing them right now using Chrome jobs, to managing the, to managing our, agent. So I did, kinda polish this a little bit, in the second version, which we have, I use the SQL right now to manage the agent's date, for the scheduler. So down here on each layer, so we have this enrichment. Speaker 2: We do use an enrichment by using the FireCloud to try to understand our target customer, their prospect website. Then the email generation have a 2 parallel system. The old old workflow is, as I mentioned, it's a multi step for me and fire prescripts, and your workflow is totally fully automated, using the prescript website content and using, PurpleXity. So my prompt, you need the email, generate HTML, the update, the contact to Breville. The email delivery tracking, I have a tracking system, for my email, and I read the SQL schedule. Speaker 2: I I ran I just use a JSON file because it's easier for me to track in them. I have a crunched up every hour to post Breville for the delivery, open and click and bounce events. Then, my AI agent can go, you know, decide for next step. The 1 the cool thing is I have a Slack web host just like all you guys have integrated in my Slack for you to send email, then storage on the disk. On the database type, right now, I just using our cloud support right now. Speaker 2: We host our AI agent on Google Cloud. The email service, it's reality or workflow, bypass the database in there using the j on the JSON files. So few of the key trade offs I made was that the the speed to market architecture, which should glitch JSON disk and cron jobs. But I recently found a really cool tool called Temporal. I don't know if you heard Speaker 8: about it. Speaker 2: It's called long running agents. In the future, I may use that to manage my agent's state. We're using AI email generation over templates, using purple at Sysorno, to generate every email fresh rather than using Tableau and just a variable substitution. So, another key decision right now is using Slack file or date database because, it's just no database for me to mention. But in the future, my, when it's getting complicated, we need probably getting, you know, observers and database to managing our thing. Speaker 2: And other is single thread prompt or event driven. So we have a Chrome Joe polling every 30 minutes instead of webhooks event driven scheduling. Apps, a lot of that is that simple. The webhook server to keep alive. Joe event bus. Speaker 2: I hate event bus right now. But there is Joe no favor from missed weapons. So, overall, this agent has has been helping us reach out to customers and to scale that in the future. I think 1 Kelly thing I'm my concern is how do I managing all the long run agent or efficiently robustly in the future. So that's it. Speaker 2: We have been, it takes us to 300 emails to getting 1 paid customer. Joe it is a lot of emails. Months, like 3 months. Yeah. Speaker 6: How many mailboxes do you Speaker 8: have in Speaker 2: 1 month? We have right now, I have, like, maybe 2. Yeah. You need, like, 50 at least. Yeah. Speaker 1: Yeah. 25. I love Speaker 7: the availability. Because she's going to the spam folder at this point. Speaker 1: Your your mailbox is cold. Are you Speaker 3: 25 max? 25 max per Speaker 2: yeah. Okay. Speaker 1: Yeah. Per day and per per mailbox. Yeah. Speaker 2: Okay. That's a that's a good point. So it's like quantified. Speaker 6: Look at Smartly. They have a good API. Or yes, Smartly Joe a good 1. Speaker 0: Joe Speaker 7: wanna go really big deal with that email bison. It's good millions of emails. Speaker 6: Which, you know, spam. Speaker 8: There's a couple of great tools out there for 1 that email. They'll just like send tens of thousands of emails legitimate to their network. That works. Your email address up. Speaker 9: Yeah. Speaker 2: But that's Yes. That's Yes. Speaker 5: That's really good. What are you Speaker 8: what is that? What is it about, like, the customization? Speaker 2: So the first paid customer is is actually we we did analysis. He is very business driven. He's in California, near the Bay Area. So they go to all the events. They actually are kinda have that consciousness of AI speed up their workflow. Speaker 2: Another personalization is, about the pain point that they really care about, you know, driving up their revenue for their home care agencies. So understand that it was it really matters. Yep. Speaker 3: Cool. Yeah. Brian? If you would not have built this, how many third party tools do you Speaker 7: think you would have used Speaker 3: to daisy chain this together? Speaker 2: I mean, at least maybe 5, 6, 7, 10. We have to get linking to Linking, just think about Reddit and also the, emails. But the the thing we I think the true amazing thing about beauty is is this is like a test run for us to understanding, like, building long horizon agent. After building this, we we automate, like, everything out of our Slack channel from, like, leads, from managing our tickets, from our, you know, emails. So that comes down to the question is, do you build or you wanna just use Joe tools? Speaker 2: Like, I think every phone I have just think about in the future is that the the the trade off. Speaker 7: And not not just how many tools Speaker 6: have you replaced, how many people Speaker 7: and hires have you replaced doing doing all that work? Speaker 1: Can can I say something on the the just to add to using the LinkedIn stuff is just so hard? Yeah. I did discover that perplexity has scraped all of LinkedIn, and they tried to make it not clear that they've scraped Speaker 9: all LinkedIn. But if you Speaker 1: you really get aggressive with your prompts in in Sonar Pro, it's Speaker 8: extremely clear that they have scraped all Speaker 1: of LinkedIn. Yeah. Speaker 2: Okay. That's Speaker 1: it. Joe it costs money, but it's pretty fucking unreasonable. It's just Speaker 7: So that's for like prospecting and and pulling that in. Speaker 1: Yeah. Because it's I think there's various sketchy services. I'm just saying it's yet another service where you're like, here's LinkedIn, here's the person I swear I know and Speaker 4: then, like, play with the Speaker 1: prompt a little bit. And you're like, it just gives you a little history and you're like, awesome. Speaker 7: Like Yep. That that wall gardens down. Has anyone, automated sending messages to LinkedIn or like inserting data? If you're bringing your brain Speaker 1: I don't know how. I've been, Speaker 8: in a Speaker 1: cold world with LinkedIn for like the last 6 years. Speaker 9: Yep. So Speaker 8: that's something that you have to use software you download, makes your laptop 1000 degrees. It's actually Speaker 1: Joe This Speaker 6: is my growth machine. Speaker 7: My growth machine, there's there's a ton of them. 1 thing that I found is if you go to linkedin.com, you get an entire compiled source of their entire JavaScript app, and you can just feed that into an agent and say reverse engineer this and make it pretty. And That's maniacal. You can reverse engineer the whole thing. That's great. Speaker 7: And then the only thing you have to do on top of that, there's like their internal API is, it's bifurcated and there's some confusing things that you have to work through. But, you can have an entire inbox, going to your agent Speaker 9: with No. Speaker 1: As of the last 4 weeks, there's 3 different JavaScript versions. And so and they are always constantly rotating different shards. Yep. So you have to like if you do that, then you have to say this particular user Speaker 7: As in this particular version. Speaker 1: Yep. And then as soon as they change it Yep. You actually have to, like, look at the DOM and then change it yourself. Speaker 7: Yep. As you fuck. Yeah. Speaker 5: Be very careful with LinkedIn. I will get your LinkedIn. Speaker 3: Yeah. No more than 20 25 messages a day. Speaker 7: Yep. But that means that it's a high value target at the same time. Speaker 5: We had tremendous success on LinkedIn. Speaker 9: Yep. So 24 messages today. Speaker 7: Cool. Thank you so much. Adam's up next. Speaker 6: Consistent content Speaker 0: on a regular basis? Speaker 8: Oh. People Speaker 6: have been do you wish you did? Yeah. Okay. Yeah. I've been there. Speaker 6: It's like, okay. I got an hour. I got time block in my calendar. It's just some content. I was like, oh, fuck this. Speaker 6: I'm gonna go check my message or something like that. So, this is marketing machine, and I use the new playground, skill from to make something pretty. But before I hop into this, let me just demo it real quick. So marketing machine sits as a, intermediary trade automated content from your meetings that are happening on internal or external meetings. Right now, I use ReadAI, and they've got a white web hub integration. Speaker 6: We are capturing marketing hooks from the meeting. Speaker 3: Let me just do Speaker 6: a quick demo. So we're scanning a transcript based on our brand voice, brand guideline, running it through kind of, like, different content pillars. Let's just pick, let's see, like that 1. We like 1, 4, and 8. Right? Speaker 6: These work. There we go. Okay. We got a little lighting going on. So it's taking those hooks. Speaker 6: It's extracting them. I'm gonna let that run-in the background. Flip back to the demo. Kinda cool. So reading the webhook, meeting comes in. Speaker 6: It's running it through our brand voice. It's extracting those hooks. It's doing some other stuff. I have no idea. Oh, it's doing the the Slack messages. Speaker 6: And then based on we're not there yet. Okay. So we got some prompt or we got some posts. So let's take a look at this 1. What do we think here? Speaker 6: It's like, you know, make it punchier. Right? Joe I'm gonna rewrite that 1 or, like, remove client name or remove, like, that specific, dollar amount. I'm gonna revise that 1. Let's approve this 1. Speaker 6: Boom. Here we go. This is well, actually so the the thing that I like about this, it's it's stuff that I've talked about in meetings or or anyone on our team has talked about. It's like real actionable things. You have a meeting with a client, they talk about this problem. Speaker 6: Those things just die. Like, they they go maybe they make it to your product team. They hardly ever make it back to the marketing team. Now it's like you add that you add my email or any other email to that meeting invite. They're sitting in there on that meeting transcript, and this is all running. Speaker 6: So let's approve that 1. And now we got an image generation. So we've got a brand guidelines, it's got our pantones, it's got our colors, it's got our, kind of style that we use. And then once we get that back see. It should take a second. Speaker 6: So after we approve the image, it goes in our queue. Queue is literally Google Sheet right now. Here's our boom. Alright. Let's see. Speaker 6: Yeah. Looking solid. Boom. Approve. That's in the queue. Speaker 6: And you can set your, hosting cadence right. I was testing between 1 or 2 a day, different times a day. And then it all goes through the LinkedIn API. I created a custom app. It's a lot easier when you can do it. Speaker 6: You know, you can do a solo app. It's a lot harder to adapt for multitenant or multi user. But that is Speaker 8: Yeah. That's yeah. Very cool. Yeah. Speaker 7: So for those of us who aren't doing it, what's the lift that you see of creating consistent content? Speaker 3: Let me show Speaker 6: you. So that's post. Yeah. And it's like Woah. Right? Speaker 6: Just consistent like, they reward consistency. This is just all these are all top of funnel, mid mid funnel eyeballs that I wouldn't be getting otherwise. Because I I I know my stats before. It was 0. Right? Speaker 6: It was like, I don't want to do this. So I needed to make it as easy as possible. There's friction. Speaker 9: Is the posting automatic or it still goes Speaker 8: to you? Speaker 6: Yeah. All post yeah. Posting is all once I approve that last that image, it's it's all automatic after that. Speaker 1: Yeah. Have you tried my client doing this via API? Speaker 6: Yeah. I've heard I've heard you get better reach if you're posting organically versus through the API. I haven't Entertainment that, but it's like I wanted it's 80 20. I wanted to get something going, and I and I can always modify the other. So Speaker 1: I noticed you have a lot of unaccepted connections up there. Speaker 6: Yeah. But I have Speaker 1: the same problem. I need Speaker 2: an agent with a I Speaker 6: like that. Yeah. It's like never right. Like, they're they're really weird with the permissions you can get through the apps. It's like, I post, but then I can't get stats on the post back. Speaker 6: So I have to, like, scrape. I haven't built, like, the self correcting loop because it's not just an API to end point. Speaker 4: Actually, do you Speaker 1: would you keep doing this even if it didn't? Because I I have a hunch that these will become completely walled gardens. Like, is it worth it even if you can't keep doing that? Like, they don't want this to all the agents. Like, they they want a human to still be. Speaker 6: Yeah. I mean, it's the same thing with SEO. It's like, if it's good content, it's good content. Like, these are actual insights that are coming from our internal meetings. I think there's good value there. Speaker 6: I is there some threshold where they where they hit it? Yeah. Does does anything care? Speaker 3: But but I but I guess Speaker 1: I'm saying, like, is this the beginning of the strategy? Or do you feel Speaker 4: like you would continue doing this? Speaker 3: Even if you didn't see Speaker 1: more results, would you be doing Speaker 6: Yeah. I mean, other channels, right, it's like adding other channels. I wanted to do it where it was actually pulling our PRs and auto doing that, like, kind of new releases. So it's like anything any more friction I can move in the system. But, yeah, I think other channels will be the next Speaker 5: 1. Yeah. Speaker 7: Are you on X here? Speaker 6: Yeah. Yeah. Yeah. Yeah. I've and I I've this is, like, the third iteration of this I've done. Speaker 6: The first 1, I was I had an X bot too, and that was, like, all in in make.com. But this 1 just works simple. XBA has a pretty nice engagement Mhmm. API. So you can, like, come Conversation, obnoxious. Speaker 6: Okay. Sweet. All set. Thanks, guys. Speaker 9: Joe, yeah, this presentation is a bit different. I brought my open Cloud Mac mini. Right. So I hope we leave the building alive. It's sheer force in here. Speaker 9: Yeah. So we just needed to water things in. Hi. I'm Uzair. I moved out here not long ago, loving everything about this place. Speaker 9: Done so many things in my life, a lot about special needs, but right now, the creator of AI Game Master is the song portal. It's, it's AI driven storytelling game, in the realms of art and role playing and and d and d. And we're completely bootstrapped, so that means no Conversation, and already generating tens of thousands of dollars a day. It's not my first project, so I'm very proud of that. Thank you. Speaker 1: Thank you very much. But if you've Speaker 9: dabbled with AI, you know that comes with a cost. So AI costs a lot about tens of thousands and thousands. So it's we're still in the trenches, But it is a success and it is going down, Joe, a lot to do. But being, frugal as we are, when OpenClub came out, I said, like, hey, maybe we can get this cheap employee, this apprentice working for us. Because I've been nagging the team for a long time, about about, like, hey, everyone's talking about you, Reddit, about, like, role playing. Speaker 9: Maybe we could answer everybody. It's like, yeah, we have our own jobs and stuff to do. So I was hoping OpenCloud would help us. I love Reddit. I've been, I started, like, finding Reddit when generative AI came out. Speaker 9: So I did a whole bunch of things, that got some attention, like, like, photos and stuff, and some less controvo more controversial, less controversial things. But I love it. It's the bootstrap heaven. It's where you can post for free, and it sticks. And there's, like, SEOs, and Google reads that, and real people read that. Speaker 9: So it's where you can really talk, but most places have a pay go up. And there are, like, every single moment, 1000000000 conversations going on. Obviously, many of them about your domain. Joe, theoretically, I would love to jump in and have a conversation. Okay. Speaker 9: Side disclaimer, I I hate spammers. Alright? So we're I think everybody touched on that. Like, we wanna utilize our force, but very be very minded about it and and respectful and just, and delicate about the the approach. So, I hope this can bring some value to a lot of you. Speaker 9: This is the big takeaway from the top, spoiler. But f 5 bot is a tool. It's a free tool that listens to any keyword you want on Reddit. It's free up until point you, like, make sure it's worth the money and you can add, like, more terms or whatever, more skills. But I was pretty surprised to find it, like, only a couple months ago, although it's been around for, like, tens of like, 10 years, I think. Speaker 9: And it's really good. Like, I find the lead, like, say, we're in the AI D and D. So, like, okay, search for, I'll talk about this later, but AI D and D and then we jump in a conversation and say, hey, try our tool. And then, like, AI Dungeon is a competitor. It's, like, they're the first company that came out with such a game. Speaker 9: So if you could tell it, well, don't surface things that are in the subreddit of AI dungeon because I'm not allowed to post there. It's a waste of everyone's time. And, again, be respectful. So it's incredible. And I send, like, the team, hey. Speaker 9: There's a conversation going up. Try you tap on that, and immediately, the tool throws me a message. The team responded because, like, it's very instant, these, alerts. So if you're using listening to Reddit, definitely, worth a shot. So it's called fibot.com. Speaker 9: And, again, it's free to start out with, Speaker 1: in which it costs us. Speaker 9: Yeah. That so once you have that set up with time, your inbox starts looking like this. Like, everybody on Reddit talks all the time. So as as long as you put up those search terms, they will surface up and this becomes overwhelming. And there's a lot of work to do in, like, finding the good ones. Speaker 9: Again, do not spam. I think you wanna come in the most organically you Speaker 4: can, and Speaker 9: that means finding the right conversations. Not anybody that said AI means Speaker 1: I should jump in and say, hey, hey, Gambier master. If Speaker 9: they were looking for, like, an AI assisted adventure, I could come in and say, have you tried our tour? So, right. So we got, a bunch of these, like, these leads, and I was sending them to the team. And then it's like, okay, it's overwhelming. Enter open cloud or cloud bot or mold talk or whatever. Speaker 9: The big promise of, like, the agent that will do everything for you. It says there's, like, there's a quick start. In practice, it wasn't a quick start. It wasn't easy. Like and I'm a good developer. Speaker 9: It wasn't the easiest setup to to put up. Then once you start reading beyond the hive, you notice a lot of people running through the same thing. But eventually, it was up, and you get this. If you ever wondered what it is, this is like a dashboard. So it's running on the on the Mac Mini or you can go on VPS, like a virtual server. Speaker 9: And then there's a constantly running thing, and you have a dashboard, with several agents and, like, prod jobs. You can do things, like, in a repeating manner. They could just chat to it. It's like, hi. I'm at a high tinker to Leon. Speaker 9: It's like at the setup. Yeah. It asked for API Kelly, Joe, like, you're paying for all this. Yeah. So that was cool. Speaker 9: So this is open claw and, like, okay. Now I want you to work for me. So I'm getting all those emails. Oh, sorry. So the first thing I did was connect it to to Telegram, and then you can chat with them wherever you are, which is convenient. Speaker 9: And to make sure everything works, I was like, send me a daily joke. So, sure. I'll send you whatever. I timed it to be right when we're having lunch, and I'll suffice the family. I was like, why don't skeletons fight each other? Speaker 9: They don't have the guts. Speaker 1: That was awesome. My kid knew Speaker 9: it already. But then the next day came, and here's today's joke. Why don't Joe each other? And then, no, no, no, no, no, no, no, no, Speaker 8: no, no, Speaker 9: Joe, no, no, no, no, no, no, no, no, no, no, no, Joe, no, no, no, no, no, no, no, no, no, no, no, no, no, no, Joe, no, no, no, no, no, no, no, no like, it's shiny upfront, but it there's a lot of devil in the details. You need work to make it efficient. Like, this is an easy example. I could have told him, save all the jokes you've already sent me. Make sure you're not sending them again, for example, if I wanted to make this really efficient. Speaker 9: And then the next thing was, like, okay. So I connected to my email. I'm getting all these f 5 bots of conversations going on Reddit. This was making sure it could read my inbox, and it it did. It felt good, Once I finished connecting everything. Speaker 9: Yep. Speaker 7: Just 1 minute. Speaker 5: 1 minute. Okay. Speaker 9: Joe it can read the inbox. That's great. And then it's like, can you send me that on Telegram? It's like, sure. I'll run a a writer, watcher in the Chrome Joe, and we'll pull the thing and send you the thing. Speaker 9: Then I realized my little apprentice minion can code. That's cool. It's like, wait. You can Joe. So okay. Speaker 9: Can you, like, write me a dashboard of everything you're doing? Like, sure. Here, I wrote up an express better SQLite 3, which I haven't ever used, with CRUD blah blah blah blah blah. Just run this and it will work. And it didn't have a control panel. Speaker 9: I can give it tasks here and, like, they go somewhere. I actually haven't used this, but it feeds the leads from f 5 months to here. Mhmm. And this is where I come in organically manually, and I can see, like, the suggested reply. Nice. Speaker 9: I can just click follow-up. It opens up the lead. Yeah. I can, like, put in the, response I want. And again, I haven't connected it manually because responses aren't all that great as you've seen. Speaker 9: Still requires a lot of manual of a lot of attention to make sound human, and Reddit is very sensitive to that, so it's extra important. And yep. But it's a nice tool. Take away from this talk. If I bought less Joe OpenClaw, there's a lot of products coming out for, like, your own agents. Speaker 9: I'm not sure OpenClaw will be the best 1. And finally, I just ran into this in Redmond. I'm thinking of collecting that to OpenClaw and see what happens. Speaker 3: Yeah. Speaker 5: Why open call and not just have, like, log code? Speaker 9: This yeah. I'm biting the hype. Right. Joe so I didn't connect it still because it's not a 100% yet and I do all the fine tuning by myself manually. But at the beginning it was like, hey, I read the entire Reddit thing, and Speaker 1: it was it Speaker 9: sounded all very AI. So I made I I asked it to sound like me. I gave it a bunch of posts that I posted, and made it sound like me. And then it still sounds like AI, maybe giving me. Like, oh, Yamamoto. Speaker 9: I love what you're doing. Oh. It's like, yeah. That's not exactly me. So I don't have a good, answer. Speaker 9: Like, it's still manual, Speaker 3: to Speaker 9: get right. And I'm afraid of being banned by Reddit. My teammates have already pinned, Joe, like, we're trying to be mindful about it. Speaker 1: Yeah. Are you seeing your prompts with, like, different random variations? Because it seems like you're I don't know how many posts you're doing per Speaker 7: day, but if you're giving it Speaker 1: the same prompt over and over and over again, even though the content's the same, someone gets the same message. I'm just curious, are you experiencing with, like, Speaker 9: minimizing what that prompt is? So a a bit. Yeah. They gave it some some messages myself inside Joe it can have, like, organic things to pull from. But I I noticed like it built this huge machine for answering. Speaker 9: Initially, it had templates which I did ask for then all replies sounded the same because it was using the templates. And like I honestly haven't used this as much. I'm using f 5 bot, but it feels like it requires a lot more work to get really useful. Like, regarding, like, the exact replies and the the templates and the seeding. Joe, yeah, I was looking for, like, a quick win. Speaker 9: I didn't find it, but f 5 bot is amazing. And, and the future is crazy. Speaker 1: Yeah. So, Speaker 8: I don't know why everybody can do Telegram for Vociplo. Yeah. Speaker 1: If you switch to Slack, then allows you to do blocks and you can do apps. And I find that to be probably super useful for your workflow, like, what you're doing with the Apollo of Speaker 9: the Sun, things like that. Speaker 1: We have 2 at Speaker 8: our company, and they're on marketing Speaker 1: and it's also software. Alright. And Speaker 9: confused. I think it was higher and they're like set up scripts. It just did Yeah. Speaker 8: It's harder. For sure it's harder. Speaker 1: But it's worth it. Speaker 9: Oh, yeah. Yeah. Sounds good. Thanks. Speaker 7: So it seems like it's pretty good at building things. So why not direct it in that direction? Speaker 6: Like, why not put it in that direction? Speaker 9: I think it Joe maybe something. If if anyone tried to build things, with AI, tell me if you've experienced something else. But you've reached a certain point you can Joe beyond. Because you start throwing a prompt at it. You have no idea what it's done under the hood by now. Speaker 9: And then you start throwing prompts at it. It's like, I don't know. I don't know. I'm still I'm I'm a boomer. Still not convinced AI, is there yet. Speaker 9: But at the same time, I do believe we're all out of a job in 3 years. So I'm kind of living both sides right now. Speaker 7: Quick shout out to another AI tinkerers event, April 13. Is that right, man? Speaker 1: Yeah. That's our next, dev tools track event is April 13. We also have the AI tinkerer, the software factory, intensive coming out on the April 21, which is a 2 day training for software engineers to Speaker 9: become soft we wanna become software factory engineers. Speaker 1: So we're gonna publish that today. But the the next, DevTools track is already out. You can jump on that and people have been. We've already got, like, hundreds of registrations. Speaker 9: So good Speaker 1: good Jimmy on that. I review every day, but, like, the early you register, as as Dan said, like, it it counts. You know, helps you get in. Speaker 7: That's an that's an amazing track. I go, as much as I can every single month just get a preview of all these problems of, like, how do we build better with AI. Yep. And, like, it's Scary. Like, 3 years, Speaker 8: I don't know if Speaker 9: it will last Speaker 7: that long. Speaker 8: So Yeah. Speaker 7: Oh, 1 last question. Speaker 8: Yeah. Quick tip. We we had the same issue and we started using, like, humanized AI where I've passed open AI. I'd have open cloud take this stuff from, from cloud or whatever and run it through their API. And it because you end up spending so much time just DPA ing the content. Speaker 8: You run it through some of the specialty tools that Speaker 9: are really good at that. Humanized AI. Speaker 8: Humanized AI, I'm sure there's a few Speaker 9: I've heard of the Trope R and D, you said. I'll check that as well. Thanks. Yep. Awesome. Speaker 9: Thank you very much. Speaker 7: There you go. You're the second customer. Speaker 1: Here for Speaker 5: that. Joe Speaker 6: Thank you. Speaker 7: Here, real quick while I'm setting this up. Can you Speaker 5: give a explanation? I'm not sure. Speaker 7: Okay. Okay. Speaker 8: Let's meet Eddie. Speaker 7: Oh, you've joined? I'm not showing other cool. Okay. Speaker 6: You wanna stop it and try again? Speaker 7: Can we play the song again? Speaker 9: It sounds like there. Speaker 6: Good. Speaker 4: Alright. Cool. So I'm putting a conference on May 15 for starting customers. And for this conference, I decided to build the entire website, the entire ticket sales, everything from scratch, the management, the back office, just so I could learn as well. And I had an objective of, like, not writing a single line of code. Speaker 4: So I didn't write a single line of code for this. And I even wrote a very long article that you can go to the conference website and read, like, how I did each detail, like, the tools that I used, everything. I wanna also I actually minimize the number of services that I use, the tools that I use. Joe I want, like, everything in the same repo, everything in, like, running as few services as possible. It's a conference. Speaker 4: Like, 6 months after the conference, Speaker 9: no 1 Speaker 4: cares about the website anymore. So I didn't wanna put a lot of effort into it. And and then I was like, oh, I need to do marketing for the conference. So I made this complicated Google spreadsheet with, like, many tabs and, like, the people that are gonna post when and, like, the type of topics that I'm gonna post. And, like, I'm gonna mark everything. Speaker 4: I'm gonna do, like, how how you do marketing. I was like, this is stupid. Right? This is gonna take way too long for me, and I hate the tracking. I hate the generation of content. Speaker 4: Everything was like, you know what? I'm gonna just use a bunch of cloud skills, to do that. The interesting thing about skills is if you try to create them 6 months ago using cloud, that sucked. Like, you didn't even know what skills were. Right? Speaker 4: Even though it was Claude skills. Now it's so good at creating these skills. Not perfect, but good enough that you can do a bunch of them. So I created a bunch of skills to help me, create campaigns. And what I did is, literally, the entire conference is on this repo. Speaker 4: Like, there is no files there, like, there's no information, secret PowerPoint, like a PRD somewhere else. Like, everything is in this free flow. Right? So it was perfect to put the marketing content here as Kelly. Because as I run these fields, you can actually read the code or read the copy that I used in creating marketing campaigns and everything that it's gonna do. Speaker 4: So what I did was I created this similar how you you have PRDs and you have flows and you have other things that you do for coding. I did the same thing for marketing. Right? So I said, like, Here's my bank of copy. Here's my email templates that we can use, the messaging framework, resources just to to give an idea. Speaker 4: Here. Resources. Like, all the channels that I have, like, all my LinkedIn accounts, my social media accounts, all the Slacks that I'm a member of, all the emulators that I manage, like, I put everything here and I'm like, hey. This is a a place that you can look to create a company. I also put a style guide, like, what kind of voice do I want? Speaker 4: What tone do I want? What it should do, what it should not do, all this stuff. This is Claudia is not great at this, by the way. The copy is never great, but we we keep doing it. And then I generate campaigns. Speaker 4: So, like, instead of maintaining that spreadsheet, I just come here and I type things like this, like create campaign, and the campaign would be like, hey, the deadline to apply for the startup fair is April 13. And I type this, I'm not gonna run the entire campaign here because like it's gonna ask me a bunch of questions. What angle do you wanna tackle? Like, who do you wanna support on this this campaign and everything? But the the hard thing about a conference is I can I can do all the marketing myself? Speaker 4: But the power is in the speakers doing the marketing, the sponsors doing my marketing, and then the the attendees doing the marketing. So what I did, I created this multi system where it looks all the channels that I have, everything that is at everyone who's attending the conference, all the attendees, all the speakers, all the advisors, everyone. And the companies that I have and does an optimization by saying, like, I'm not gonna ask Brian twice in the same week to promote a specific thing, but I'm gonna ask him once this week to, like, hey. Can you upload this email or do this kind of thing? And and the campaigns, just to give an idea here, this is a campaign that I have Gambier 2 here. Speaker 4: Like, each and every day, all the activities that I need to do. By the way, I do everything not everything. The emails are automated, but all the social media posts are manual because I think they reach better if you post it yourself versus how to make that. Right? So, like, it gives me the least of things that I need to do and I just, like, do them in order. Speaker 4: And not only that, it's sophisticated because, see, like, I have to post my personal account because I have 10,000 followers on Twitter and then, like, retweet from this account and like from this account. Right? And then, like, ask people to like from that account as well. And it generates all the content for me, all the copy, for, like, each social network, which channel to post on, email copy. The email is automated Joe I don't have to do anything. Speaker 4: Twitter, blue sky, threads. Wait. I'm gonna get to some cool stuff. A script for me to read, to record videos for TikTok and Instagram and YouTube shorts. All the stuff. Speaker 4: It actually gives me how I should dress and how I should speak. Useless, but, it also tell me sadness and sometimes, like, he thinks like it's a videographer. Right? Like, you're gonna move the camera this way and you're like, none of that. Like, I'm not gonna do any of that stuff. Speaker 4: So I do a lot of manual editing, but it also prints the the people that should be receiving emails from me. Right? Because, like, if you receive an email very recently, it's gonna look at the logs of all the companies, by the way. When When I run a campaign, it creates the log for each campaign, like, who got what email, and what campaign was asked them, and tries to create a system that it it works on that context. But since we are talking about it, let me see if I forgot anything about campaigns. Speaker 4: I tried the copies, run the AMAs. It's really cool to run the AMAs from Claude because I've I've built some tools there that, I called the SKU to run the campaign. It's like, generate 1 of the AMEs. It's like, hey, do you want to preview? And I say, yes. Speaker 4: Send send an email to me. And then, like, I make some edits, and then I say, run the campaign. And it's just like, in 60 seconds, this sends, like, a 100 of AMEs. I'm using, Postmark for for sending the emails. And also, I learned the hard way as Kelly. Speaker 4: Like, I burned my main domain by mistake because I did a stupid thing. And I was like, last minute, I decided to change it. And then I sent a 100 emails and, like, got 50 people opening and, like, 0 opens after that. Joe what I did, literally, 2 recoverances, these people are interested, I sent an email to each person that didn't open. And I said, not from the same email address, my personal email address. Speaker 4: And I said, hey. I think you got an email from me, but I think it didn't open. Speaker 8: Can Speaker 4: you see if it is on your spam folder? And if you find it on your spam folder, can you reply to that with, yes, I found it. And literally, I recovered. Like, I'm not blocked anymore. Right? Speaker 4: So, like, I can send this emails again and, like, I'm off the hook. By the way, I bought a new domain and, like, use that now. So, like, don't make that mistake again. 1 other thing that I did is this is the part that I love as well. Like, the conference website, you know, a conference website, you buy a ticket, like, speakers, yada yada. Speaker 4: Feels just like static content, which mostly it is. But also, like, the the back end of the the back office, of this conference is pretty damn amazing. I even built a CRM to reach out to small users. And it tracks open rate, click rate. I can reply. Speaker 4: It it does everything. But it's so integrated into the conference website that, like, you can pull content, you can pull names, you can avoid people that ask to be not on the list, people that cancel their ticket. Like, everything is it's fully integrated. And I can send broadcasted emails as Kelly, and I can select which ticket buyers get this email or that email that goes with the the campaign as well. So, really cool? Speaker 4: Really Speaker 5: cool. Have you thought about selling this as a product to conference people? Speaker 4: You know, I realized that I I look for products for conference all the time. Right? Like, Eventbrite, like, instant. Like, it sucks. It's terrible. Speaker 4: Terrible product. Meetup is terrible as well. They haven't evolved. Luma is great, but Luma is so, like, narrow. Right? Speaker 4: And I found 1 product that is very smart. They are ugly as Kelly, but, like, they do some very smart things with confidence. So 1 of the things that I do on on my conferences is every I think you get an email saying, like, hey. You can be part of the ambassador program. If you get 5 people to register for the conference, you get your ticket for food. Speaker 4: Right? And I did that because I saw another product doing that. I was like, oh, that's a smart idea. Let me do that as well. Right? Speaker 4: So I think there is a opportunity to do that. I'm not gonna do that right now. Someone wants to build it. I'm happy to explain. Speaker 2: Yeah. I have a question. How do you found, like, potential, like, speaker emails or, like, profile? Did you go to, like, LinkedIn or looking for people who are connected to you? Speaker 4: I am quite connected here in Seattle. So I know the speakers. I just reach out to them. And yeah. Joe mostly that. Speaker 7: Have you have you tried the Gmail MCP? Because, it actually has a pretty nice Speaker 4: So you launched there recently. Right? Are you talking about the ACLI or no? Speaker 7: No. The MCP, and you can create really nice and rich emails Yeah. In Gmail, and then it'll save it as a draft. And then it'll, like, you can actually make really beautiful emails in Gmail, then Speaker 1: you can, like, really craft Yeah. Speaker 4: The thing about my emails is they don't look polished. Right. On purpose. They look like they were sent from me. So I'm not trying to do anything fancy. Speaker 4: Sure. Like, I I don't even like the links are like like I copy and paste the links. Right? And at 1 point, I got, like, a few people upset with me because they didn't know if I was sending an email or if it was automated. That's the level that I'm trying to achieve. Speaker 4: Right? People, like, think that I'm sending by the way, all the videos come with, footer saying, like, this is cloud automation. Mhmm. Because I didn't wanna upset these speakers and these sponsors and everyone. I'm like, hey, Marcel is spamming you. Speaker 4: Mhmm. So yeah. Speaker 9: This was all built without a single line of code? I mean, Speaker 4: there's code. Yeah. I didn't touch the code. But yeah. You Speaker 9: didn't write a single Speaker 8: line of code. Speaker 4: So how long did it take to build? Speaker 9: How long did it take Speaker 4: to build? So I wrote there. It took me 3 weeks to build, but I only work about 2 hours a week, 2 hours a day. Joe it was like it was less than 40 hours that I put into to doing all this. Now I wouldn't say anyone can do this. Speaker 4: I've built many conference websites in the past. I did many conference, so I knew exactly what I wanted. Yeah. So that helped quite a bit. Speaker 9: And 1 of Speaker 4: the things that I did as well is, like, you know, I had some some documents PRD style. Some assets, some stuff that I've loaded here. Joe, here. Right. I can you know, design guidance, flows, like, what I want the flows to do. Speaker 4: So, actually, this is what I did. Right? Like I said, like, here, here's what I want. I think this should be able to do, the speaker should be able to do, the the staff should be able to do, you know, like, go at it. Right? Speaker 4: And the other thing that I did as well that I love, not related to, go to market, but I created my own, like, the email. Right? So everything is in the GitHub the GitHub. So I created the the template for YAML, like, everything is here. I have skills that say, like, hey. Speaker 4: I wanna build this feature. You know, like, these skills start asking me, like, what goes in here? Like, is the they're gonna have IDENTIFICATION for the front end or the the database or anything like that. Develop that and then, like, I just feel like, alright. We we are in agreement. Speaker 4: Go and build it. Right? It did everything. Thank you. Speaker 3: And Marcela wants it, so I will. He should go to the conference. Speaker 7: Yes. May 15. Speaker 4: May 15. Speaker 9: May 15. Yeah. What's the Speaker 5: name of Speaker 4: it? Startup Speaker 7: day. Startup day. Everyone's gonna be here. Speaker 4: All the Speaker 5: cool people. Speaker 3: Just gotta give myself an own goal. Speaker 5: I gotta request my end Speaker 3: and weigh in again. Sorry. Speaker 7: Yeah. No worries. So all of these talks are being recorded. Hopefully, this all works well. This is my first time trying it. Speaker 7: But, AI Tinkerer's platform, very similar to what you built for your event, it's actually gonna, chop up all that media for me. So I, as a as a organizer or program manager, don't need to do any of that. So I'm just sending the Zoom recording, straight to the platform, and hopefully, it'll be awesome. So every all the speakers are gonna get their own content to share, like LinkedIn posts and tags and stuff like that. So if you just watch some of the content that AI Tinkers itself is putting out, that's pretty amazing for a bar to go for. Speaker 7: So Speaker 8: and I Speaker 7: need to get Joe to come in and do a talk on that. Speaker 3: So I went to the last AI tinkerers Go to market engineering track. We sponsored it at AS. Speaker 5: It was the Speaker 3: first AI team I'd ever been to. And the first thing I noticed was that in between every speaking slot, there was this awkward 30 seconds of fumbling around and getting AV set up. And I was like, Speaker 5: this should be walk up songs. Speaker 3: I sent Dan an email. I was like, Dan, my 7 year old has walk up songs for his little league team. Why don't we? And so, I built a walk up song generator that is it's a little bit manual. There's some manual rough edges today. Speaker 3: But I think there's a nice road map I'll preview a little bit. And essentially, you go to the speaker page, and I have to manually copy and paste the bio and the abstract. I couldn't scrape speaker's website. I don't know if that was a limitation or not, but I know all of the stuff's API accessible. So in the future, you can imagine this being fully programmatic. Speaker 3: And so I pasted them in. This is Aviel as an example. And then I have a guilty pleasure song field here which is prefilled with a guilty pleasure or you could click the dice icon and it will cycle through a new suggested guilty pleasure song along the way. So you can pick whichever 1 you like. I think for Aviel, I did I did shoot the salt and pepper. Speaker 3: You can override it if Speaker 7: you want. Speaker 3: And then you click generate song, and it is very straightforward what's happening in the background. It has Claude take the abstract and the the speaker bio, and it will generate some style commentary based on the song that you selected. Speaker 8: You can't have any remnants of the actual song itself as identifiable. Otherwise, the music Speaker 3: models will choke all over it. Music models Kelly choke all over it. And then it sends that bundle out to 11 Labs to have a song generated, which you can see, interestingly, the bios are all machine generated. And so it's just layers Speaker 0: of it. It's all machines. Speaker 3: It over indexes on whatever college you went to. Sure. And then you can add it you can add it to this playlist if you wanted to. I'll just go to the playlist that we created for today. And then on the playlist, you can change the order, the numbers, you can see the lyrics. Speaker 3: This was obvious abstract, song style, what have you. Future directions, this should just flow through. The only other detail I'd add is that I I had to give some prompting to 11 laps to make sure that the song was just 30 seconds, and then it went straight to the chorus. Right? A walk up song, you can't Kelly dally. Speaker 3: You gotta get right into the hook. And, most of these were actually, like, 1 shot generations. A few of them, I would regenerate. There is a regenerate button on the on the previous screen. And so I know this is go to market engineering. Speaker 3: I'll talk a little bit about the go to market project, but, like, I'm really just here because I wanted Speaker 9: to screw it up Speaker 3: with lock up songs. So thank you for indulging me, especially the speakers who may have had more or less flattering songs. Speaker 5: Can you let the speakers pick? Because I hate country. Speaker 9: Joe, I could. Speaker 3: 1 thing I I tried to do was to look up all of your Spotify profiles and then pick your actual guiltiest pleasure song. But it's a really hard matching exercise. I don't know if anyone's tried that before. Speaker 2: My request is can you do a gesture detection? Like You're trying to make a change. Just like Speaker 3: Figure out some some beats from entities beats from entities. Okay. So that's that's that. The the other thing, that at Anthos Capital, a venture fund I work at, we've been screwing around with me and 1 of our advisors is basically how do you build a go to market second brain for a company? How do you take all of the financial data, all of the presentations, Google web analytics data, interviews with the executive team, customers, stuff that as a venture fund, we might have in a data room or a diligence pack. Speaker 3: As a company, you might have just in the corpus of data that you keep to run your business. And how do we pull all that together, add context, and in our self interest, you know, make ourselves sound smarter. How do we go into a board meeting and ask an insightful question or suggest an experiment to run? I don't know. Speaker 8: And Speaker 3: in the hands of a company, how do you create what might be a virtual CRO? And so 1 of our advisors, use it for use case, I will show you in a minute. But I'm like, who made the comment about being a white belt? I'm like a no belt. I'm like a, Speaker 5: you know, Speaker 3: maybe a a an elastic drawstring. Speaker 9: So this is like Speaker 3: the lesson I've learned is like the dumber the questions that you ask Cloud Code, the smarter you will see. And so you kinda just have to like remove all of your ego from the exercise. These are the prompts that we fed in. The process is pretty simple. I'll show you the results in a minute. Speaker 3: Basically, you go to Cloud Code and you're like, hey, here's all this data about a company. Go crazy. Learn everything you can. Go spelunking through everywhere. Remember everything. Speaker 3: Ask me questions and I'll give you more information and just keep iterating. And then we went to codex And said, hey, this is what FloodCo thinks is a picture of the business. You're a skeptical analyst. What do you think? And use them adversarially. Speaker 3: We then Han,PhD them generate, some prompts to feed and some documents to feed into notebook l m. Speaker 9: We're like, Joe, it'd be cool to Speaker 3: have a podcast about this and a deck about this. And we also then went and had them generate a bunch of go to market experiments. Joe, hey, this is a business that has the following goal in mind, come up with 8 to 12 different scenarios that we may not be thinking about Joe that we can sound smart, the companies can run those experiments, and maybe we can model that with some of the impacts we'll see. 2 very different examples. The first 1 is a real example I'm sharing with permission, which is a business called Grillers Pride. Speaker 3: Where did they go? This is the result of it. So 1 of our advisors has a father who runs a kosher meats business online. And he had this entire, website generated for it. All sorts of background materials, interviews with his father about the business, with the goal of going from 2,000,000 to 8,000,000 in in gross revenue. Speaker 3: Right? He would never pay a consultant. He would never hire anyone to do this, and the results are sort of interesting. The next example is based on a real business. I anonymized the data, and I actually when Claude would ask me for more information, I would just send synthetic data that OpenAI would generate back at it. Speaker 3: So the numbers are a little bit off, but the result is quite interesting. I'm gonna scroll all the way down. This is a, property services business. I'm not gonna say exactly what it's in, but think big projects you might do around your house. Right? Speaker 3: Like landscaping, remodels, roofing, gutters, that kind of thing. And the good market experiments that it suggests are actually, like, fairly interesting. It's like create a storm response network. When there's a big storm in an area and insurance claims go up for repairs to houses, why don't we have a tiger team that is ready with a playbook to jump all over that and capture that demand? And by the way, here's some of the impacts that we might get from that. Speaker 3: You can imagine the next step where you might actually go out and start to execute some of these Joe to market experiments that are wrong. But again, 1 of the takeaways is this is like it's amazing how fruitful all this is. I know we've all experienced this. I actually completely forgot these prompts. And at the end, I went back and I was like, hey, this is the site that I generated. Speaker 3: What are the prompts that you think would actually best you Speaker 7: in this? Speaker 3: And then just loop through it a few times. But all this stuff is out there. It's dead simple to to go and work with. I just threw it up in a GitHub repo. It's all public. Speaker 3: You can take it and run with it Speaker 5: if you want. It'll take Speaker 3: a few hours and you just let it go crazy. Joe that's 1 experiment. But really I'm here for the walk up songs. Speaker 7: So if this is what your investors are doing, yeah, I think you gotta do it now Speaker 9: too. Yeah. So I built something Speaker 8: similar because I do advising and email testing and I was like, I wanna sound smart when I walk into them, you know. It had like stages where it's like, we we do some some stuff ahead of time based on what they provided and then we have like an hour long or 3 hour long conversation. And then we feed that into strategic analysis. Yep. Speaker 9: We help build all that. Speaker 8: This is more advanced. It's been a incredible process. 1 thing I noticed was like, you almost need like 2 or 3 personas of like the skeptic and the optimistic person because he gets very sick of that Speaker 6: and it's like, oh, this is the best deal ever. Speaker 9: And it's like the best Speaker 8: thing ever and you need them all, like, the counterpoint, Speaker 5: you know, Speaker 8: side. I hope you've encountered that and, like, how you manage that stuff. How do you manage that data? Speaker 3: When you see those deals, Speaker 8: Joe not a good deal because it's a good deal. Speaker 0: Have that as kind Speaker 8: of, like, a model This is Speaker 3: a great point. So for my own, like, personal second brain that I use to do my job, I have basically in the personalization, I say you are a level headed, skeptical, analytical. I try to dial down the expectancy. And then 1 of the other byproducts of actually how we work at Anthos, which is perhaps different than other venture funds is we have a very Socratic approach to investing. We challenge each other, and we have a team of people that will do its own research and bring back its own conclusions. Speaker 3: And 1 of the things that I've done for my own second brain setup is I have a whole bunch of skills that correspond to the jobs that some of my teammates do Speaker 8: Mhmm. Speaker 3: That might run through an analysis of a company fed with a bunch of examples of decks that they've created that will play that persona for me, mostly just as a bundle of skills. But I I completely agree with the sycophancy that's embedded in a lot of these things. It's it's problematic for all of us, but especially, I mean, for the investor months. Speaker 8: Yeah. People are trying to figure out, like, the conversation you have in chat when you say that isn't right, I want to essentially, like, adjust it. That's data and that's information. If you pass on the deal, we do, like, a deck of produce, whatever. So how do you store that knowledge Yep. Speaker 8: So that later on, it's exactly in context. Speaker 3: Yep. And I haven't, like, measured this. It may be hard to measure. But the other thing about doing the adversarial step of hitting the models against each other and sort of, like, fact checking arguing over data, hopefully, the data is more accurate, and it does actually pick out inconsistencies in what the other 1 comes up with. But it should also drive towards a more neutral outcome as well. Speaker 3: And that stuff's so easy to do now. Speaker 1: Has anybody tried the GTM experiments before? Speaker 9: Yep. Speaker 1: And then do when you when you feed the thing back to codex, what do you exactly feed it? Do you feed it everything? Speaker 3: So I, I feed it the conclusions, which is basically just like a series of markdown files. And then I also feed it the source documents. And I say, this is what they came up with. I mean, you can see the prompt itself. This is what they came up with. Speaker 3: Joe out and just drive through all of the things you see on the screen here. Are you not sharing? Challenge the assumptions. Oh, I'm not sharing. Yeah. Speaker 7: Just go through everything that you see right there. Speaker 3: Yeah. Yeah. Yeah. I won't share it. But data integrity, challenge all the assumptions, find logic gaps, find inconsistencies, optimism bias, any data that's missing. Speaker 3: And it I mean, this is all, like, a Speaker 7: few weeks old. It seems to be a pretty good Joe. Speaker 3: On the have we tried the GTM experiments? No. The grillers pride business, the, multimillion dollar kosher meats business has started to. We haven't been now we're also not running these companies. Right? Speaker 3: We're investors. We give suggestions, and it's up to the founders to do what they will. I would hope and expect that over time, some of them get put into practice. Speaker 9: How do you do or do you do Speaker 1: any competitive analysis in your? Speaker 3: Oh, great great question. Yes. I don't I didn't I skipped over this. We also just Han,PhD, like, some basic deep research done by both Claude and Jack CPT on competitors, and we specify a few of them and just Marcelo overall and then fed that into the same Speaker 5: corpus of data. Speaker 8: For the CRO role, can you connect with watch guy and have it tell you, like, here's what the the watch that the CRO company should Speaker 3: Should buy. Here's what the president's club should give away this year. And here's how you get the best deal. Feature request noted. Speaker 8: That's cool. Speaker 0: I also did something similar maybe, you know, in that way. But, I made a PR became popular. It just it just lost all the all the it became very generic. So it lost a lot of details even though when you lose through all the steps, they are good. Questions are good. Speaker 0: But the overall result is just very generic document. What do you think is the reason why, because the for example, with the agent count is a great, great idea, great, great concept. What's the in your architecture, this seed that brings this feeds this, real, integrated ideas in place instead of making it more and more generic? Speaker 3: I I don't know. We haven't tested that enough to say we're just using the innate properties of the models, for the most part. In a in a full exercise, we should have some insights of our own. And Kelly, that's applied to kind of post processing. But you could imagine a world in which you feed this thing a few hypotheses that you already have, which is probably the right answer. Speaker 3: Right? A lot of this stuff is still going to be from between the years of human minds. Other than that, I don't know if you. Speaker 4: But what you described sounds like a context engineering problem. So you're getting garbage because the context window is not accumulating all the information that you need or is not ordering the right way. So if you try Opus 4.5 with 1,000,000 tokens, you might get completely good results. Like, I changed a 100% of all my skills that I was using when that model came up because all my skills were optimized for a 200 k context window. And now that I had a median, I was like, oh, I can do Speaker 7: things very differently. So, you know, Speaker 4: yeah, your net. Speaker 7: Or just 1 more 1 more question. Speaker 1: So I'm another person who's also built something similar. Speaker 2: I have Speaker 3: a similar question. But I closed my first customer report this morning. Joe there's the 2 different examples that I referenced. 1 is the e commerce side and the other is the, like, the, services business. E commerce side is plugged straight into QuickBooks and use that as a source of truth, which is nice in some ways, but you're still gonna inherit whatever data cleanliness is in the QuickBooks instance. Speaker 3: And then the other 1, perhaps it's more problematic. I just added a whole bunch of Excel files that came from, a company. And even worse than that, I transformed it with a bunch of synthetic data, and it's like it's super noisy. And so that one's not that was just for pure demo purposes. In a perfect world, I think we could rely on the data hygiene of of the ERP systems themselves. Speaker 3: We're typically later stage investors, so maybe that's that's true in some form. But I haven't encountered that problem and and looked into Speaker 5: it enough to really Speaker 0: help you to answer. Speaker 3: I I found, Speaker 8: if you use, like, Sonic, Speaker 5: do you stick it just to be doing, like, Speaker 3: yeah. Forces for forces. Speaker 7: Very cool. Alright. Thank you. Thank you. Speaker 1: Alright. That Speaker 7: concludes our demos for tonight. Thank you so much for coming out. Don't forget to do the survey. You can stick around for another 10 minutes, and then we gotta go. Speaker 9: There you go. Thank you very much. Hey. Hey. Welcome, Sonsborough. Speaker 8: So Speaker 9: great. Yeah. Good job. Good job.