The automated way to create it

0:00Ever since Karpathy talked about having a personal knowledge base, people who are into AI started creating personal wikis for themselves with all the information that their LLM needs to help them make good decisions. Well, what happens if you have a team? What if you have multiple people who you want to inform and you want to have access to the data that they need and that you want to get data from what they're doing so that you can make better decisions? Well, the CEO of grain.com is coming on. He's got millions in revenue. He's got a big team. He's got tens of thousands of users. He's taking everything organizing

0:32it auto magically into a wiki and he's going to show you how useful that is and how you can set that up for your team. Let's get started. [music] Presented by Zapier, the AI automation company. Everyone's talking about Karpathy saying that you need a personal wiki and that's what makes your LLM a lot smarter and your AI engagement a lot better. You all

0:49at grain have got something bigger. What's your version of it? Taking the main idea and basically making it a company a wiki with all the context of the company. So, essentially every conversation at grain we record all of our meetings. We capture all of our meetings, team meetings, except for the sensitive ones and give access to the whole team so everyone can have that collective

1:11context and intelligence. And then you have Claude somehow take it and structure it in a wiki so that everything is organized on your wiki. Why? Why not just say to Claude, "When I have a question, go to grain, look at my call transcripts, my team's call transcripts and look all over and bring me my answer." Why do you need the wiki? So, I do also that. It often times depends on the on the type of task being

1:36done. The wiki the wiki gives a lot bigger context. Not only are we pulling all of our calls, but I'm also pulling in Intercom conversations for context. I'm pulling in things I've clipped from the web, documents that are have important context, things we've written. And so it it's it's broader. And, you know, if you're every single time you're pulling contacts from an MCP, it blows through tokens pretty quickly. This has everything already synthesized. So, when it pulls the meetings, that it does save the raw transcripts, but it compiles them into a wiki with information about

2:18companies, people, projects. I see. So, you get all the transcripts on your computer, and in the at the same time you have summaries of it, so you can you can find what you need fast. Okay. Let's Let's look at what it can do. It compounds into recursive knowledge about about different entities, topics, people, companies. All right. So, you asked it this question earlier uh today. It's, "What's my strongest proof point for Grain's agent-first strategy?" What did it come back to you

2:46with? And is this Does this make sense? Yeah, it makes a ton of sense. So, we built an We want to We built a really early MCP back when MCPs were maybe not as well developed and supported by the ecosystem. We shipped it. We got some additional early interest, but there were still just like a lot of gaps. We're a startup. We have limited bandwidth, and so our attention went

3:06somewhere else. And then, we out of nowhere, you know, starting kind of early this year with some of the improvements, we just started seeing our MCP usage climb astronomically. So, as it says, MCP weekly users grew 15x with almost no investment. So, that was just kind of what we'd already built. We saw this We saw this signal and like, we have to This is a bright, you know, bright red center of the sun. We need to put everything behind it. So, that's basically been the company focus for the last several months is improving our MCP capabilities, improving the documentation, improving the examples, doing customer stories of how people are

3:41using it, just to, you know, inspire customers on how they can get value from it. And even with no investment, 15 times more use came out of it. The documentation was not very good, no marketing push, no dedicated team, and still it just absolutely popped. And as we scroll down, we'll see that there are even people who who are switching from Codex, from ChatGPT to Claude just so they could get this access. So, this is the kind of thing that you would say, "All right, I'm trying to make a decision here. I'm trying to understand if this is the right move. I need a place to ask, and I don't want some

4:12generic LLM." That's one example. Let's see another example of how you use it. And then we'll see how to build it. Yeah, so I asked this one this morning. Based on my own meeting transcripts, where's the gap between how I communicate strategy and how my team understands it? You know, a strategy is only good only as good as it is used to focus your team and help everyone on the

4:32team make coordinated decisions. Uh-huh. So, it came up with several gaps that I think as I'm reading I'm like, "Wow, this is really good." So, the first is that the strategic conversations happen in a small circle, and then I don't do as good a job as, you know, expanding them to the rest of the team. Uh it gives some evidence from different

4:48meetings. Wow. Let me read it. Evidence: In May 15th capture strategy session, Jake says outright, "I haven't seen any of that or viewed it." Mike had stayed up very late building the design and shared it only to Jeff. When asked why it didn't why it hadn't gone wider, quote, "I just ran out of bandwidth." So, that's one issue. Okay. So, fair analysis for you. Gap number two, written objectives are thinner than the conversations that shape them. Um here, let me read it. Q2 board follow-up, Mike reads the objective document and immediately says he doesn't know if they're aligned even though he was part of the conversation

5:25that produced the thinking. So, now when you see this, how what do you do based on this? So, the thing is, right, it's it's using the context it has, which is the conversations, which is what you want it to do. Mhm. And so, some of the times it's going to be like, "That is 100% true, and it's helping me see a blind spot." Other times it might say, "You know, this is partially true, but it's missing some context." And then I ask, "How do I get better context into the wiki?" So, it

5:45can make better judgments. Okay. Let's see how you put this company wiki together and you specifically went in one of these chats and said to to Opus he said, "Analyze this for me. How did How do we put this together and how can someone else do it?" Yeah, so I basically just said, "You know, I I I want to teach people how to do this. Look through what I built and explain what's represented here, how is it structured, how does it how does it decide what context to use cuz it's not going through every single context every time?" And this gets to the last

6:15question I'll talk about in a second. What advice would you give someone to build their own? And then like what's a strong prompt you could use tonight? So, that's the kind this to people in the in the description. Absolutely. So, it says, "What sources?" Like I said, it's meeting transcripts, customer feedback, uh web clippings, and other other uh this inbox, which is manual drops of files and and other things that I want to manually put in

6:39there for to give it context. Okay. And just just so we're clear on the app that we're looking at, this is VS Code. On the left, you see the files that are on your computer structured the way that uh Claude has structured it for you. On the right is a chat window, the same kind of thing that we see with any

6:54kind of uh chat. And that's it's using the docs as reference. Yeah, I'm just using the Claude code uh plugin in VS Code here. That's what I use. I love it. Yep. Then you can see how it's structured, kind of explains how it's structured, and then how how it decides what to use is it reads wiki.index. It tells It kind of tells gives instructions on how where

7:18to look for files, what to pull. Um and this is to your point of like what do I do sometimes when the answer maybe doesn't feel quite right is you can also think about okay, how do you pull better context, but also how do you improve the instructions so that they'll can kind of decide what to weigh and what to pull. You know, we had a conversation earlier about maybe it's maybe in this case my knowledge graph is biasing a bit too much to older context, and so I want to add a little bit more recently to it because we're constantly learning and changing. So,

7:45these are things you want to tweak. Respect the append-only structure. I like that. Yep. Then advice for someone building their own, these are I don't know if I endorse these. These are LLM this is LLM advice but start with one source. Raw files are sacred, append never overwrite. You you just mentioned that. All logical. I actually think if we were to start with one source and it's not because you're the the head of Grain the the transcription service but I would start with that. There's just so much value in our meetings and we don't go back in and mine it unless we're hunting for the one thing somebody

8:23said. Meanwhile, the way our company really runs is in there. The way our companies Yeah. the the things we're really working on are there. alignment, decisions, these all happen in meetings. The context is just so rich. So, I would start with that and the API for your company in many ways. Yep, exactly. Say did take everything in my meetings for the last month and we're all using meeting notes. Take it in there and if we've got Grain we've got MCP access, go take it and create a wiki for me using this structure or if you want you could adjust the structure. I totally believe in this one first and then build beyond

8:56it. Raw files are sacred, never edit them. I really like that actually. I hate when they edit because there's no version control. Yeah, 100% exactly. That's why the way this structure is you have raw is very different than the wiki. They're in separate folders. So, raw is is what? Raw just stores everything as it comes in and the wiki

9:17organizes it and makes sense of it. Correct, that's right. Exactly. It compiles raw compiles into the wiki. The wiki actually synthesizes, pulls out information into companies, people, topics. How do you create the automation that does that? So, uh this is what I'm still trying to work on. I think there's something we're all figuring out is like what's real here versus what is something that's more

9:39theoretical. Uh Mhm. These are huge contexts, right? And so trying to figure out how to use those efficiently. So right now I just have a skill that's uh uh raw compile or compile raw. Okay. So I just run that. I I have not done it. The I use Grain's API webhooks uh with a local ngrok uh server to basically pull every time there's a new meeting. So I've I've automated those coming in. Other things I'm I'm currently but I'm currently

10:08still uh compiling raw as a as a skill. Um might change that to be, you know, maybe a cron job hosted on Vercel or something like that, but for now I'm just kind of compiling raw when I when I trigger this with the slash command. I see. So I did an interview with the founder who does this for a personal wiki. He doesn't use a company wiki, though it's basically only company information, but it's only his information, not whatever everyone else is doing. And what he did was he used Coda Work as a way of triggering this automation since it's happening on his

10:39computer anyway. Oh nice. Yeah. So Coda Work does it and then everything is uh on his computer and then he can use it from there. Okay. Um we're going to give this to everyone. Is there anything else we should know about this? see there's a starter prompt here. That I think is pretty good. Okay. And of course we'll give it to them and since I mentioned that wiki that you can use if you're looking to do this in single player mode, I have a link for you right here where you can watch a video to see how to create this

11:07for single player mode. Yeah, I think one uh one caveat there is I'm still personally using this. I have We know we've actually had conversations about how do we create created true multiplayer version of this like it's effectively like a GitHub repo. That's our next iteration of that. So right now it's multiplayer in the sense that it's pulling context from you know, everyone's conversations, not just mine. Uh everyone's that I have

11:30access to. But I think that's the next step to make it truly multiplayer is like putting it in a get in a in a in a GitHub repo, everyone writes to it. I think there's questions about how do you manage context there? And and and so, that we're still kind of figuring that out, but I think that'll be the next

11:44iteration of this. And then, the other thing that I think is fun to show here is uh the way I've set this up is I have I have it um writing to Obsidian. So, this so Obsid- not writing to Obsidian, but Obsidian is synced with this uh this vault. directory And so, you can see like the knowledge graph build as you as the as the um Because this is Obsidian just looking at the same markdown files that we just looked at a moment ago in VS Code, but

12:13now in an organized structured way. Correct. Yeah, and you can see it sees all you know, all the context from the different entities. By the way, Zapier, my sponsor, they've got a new product. I can't talk about it yet that will do what you're talking about. Take it from single player mode to multi-player mode. I can't wait to

12:31talk about this. Amazing. That's going to be amazing. Okay. All right. Hey, my agent says that if you watch this far, you're going to want to subscribe. And Google thinks that if you watch this far, you'll want to watch that video.