Roll Claude out from one laptop to the whole GTM org | Sam Gong & Zach Vidibor

I wanna do it. I have everything set up. I just haven't had the time. But the killer use case I heard is that GroqBot is really good at watching content. So you can send it YouTube videos, podcasts, things like that to extract information and that it's supposedly exceptional at that.

Hmm.

That would be useful. Yeah. All right, I'm gonna start to get this kicked off here. A few more people will still roll in. Good morning or evening. This is the 4th of our back-to-back back-to-school sessions, and today we're gonna get into what I think is like the hot topic, the context layer. And, you know, not surprisingly, this actually came up in the other webinars that we did, and I had to hold people off and say, oh, we're We're gonna talk about that even more on Thursday. So I would say to people, like, if your company is not really investing energy here and figuring this out, like, stop, reprioritize what you're doing. This is like the biggest unlock right now that you could be doing, maybe working on to get AI to be the most helpful for your business, for your goals. So We are gonna dive into all that stuff. I would say I should probably introduce myself, and then I'm gonna get to introducing our featured guests here. And let's see here. What I wanted to do is, is find the share button here. So I'm Jared Waxman. Some of you know me. And hey, hey, Claude. Hey.

Hey.

Hey, Todd, you guys are— you guys have been regulars to this. I love it. So I started Go-to-Market Engineer School last year, and we do these sort of free sessions ahead of the program. It's an 8-week program, and I'll give you a little rundown on that in case any of you are interested. Then we're going to get into this, the topic for for today. So the program looks like this. I mean, we've done 4 cohorts so far. We've got nearly 300 students that have gone through it from lots of great companies. And what we try to do is bring in, bring together— can you still see this?

Yeah.

Bring together really like top folks in the industry that are going to be teaching classes live and that you're going to be able to interact with in the live sessions and in small group sessions, and you're going to get— students get access to tools during the program, from Clay to Octave, Swan, Assemble. We're going to be using Claude or ChatGPT throughout, whatever it is you want. And then the last 4 weeks of it, you're going to be doing a capstone project, something for your company or business that you sort of take, that you adopt for the period. And then present that capstone project at the end. And we even have recruiters in the room and things like that. And then it leaves you something for your portfolio. So if this is interesting to you, registrations are closing out next week, and I hope to see you there. Okay, so I'm gonna get back to this now, the poll. I am very interested to see where you guys are in terms of the context layer. Did everybody have a chance to fill out the poll before I close it? We have 30 people in the room. We've got 16 people filled out the poll. So I'm gonna leave it open for, for another couple, another few minutes, and then we'll close it out. So everyone is building a company brain right now, or should be. Almost nobody has one. really for the whole company. I ran a workshop a few weeks ago for portfolio companies of Unusual, and everybody, they know how important it is. Most of the companies, if they really had gotten into it, what we're finding is that people are able to do something at a department level, but trying to get company brain sort of across context layer across the company is, you know, it's still kind of elusive. And so we're gonna try to figure out how to do that. today with our guests. And so let me introduce our guests. Sam Gong ran the marketing org around Claude and the context layer at Workspan. And he now runs Fresh Context advising, I think, much bigger companies on the same move, what he learned at Workspan and applying that to other companies. Zach, as you guys probably know, he built Octave, and that's for the problem that's underneath what Sam was working on, keeping that context layer current once it exists. I'm running Go-To-Market Engineer School on a folder of Markdown files in a GitHub repo, and I think that that is working, for a company of my size. But we'll see, we'll get into it. Sam, I want to start with you. Give us the backstory. You spent a couple of years playing with this, building this stuff on your own. I understand you had a vector store, an agent kit, cloud projects stacked up on top of one another, some N8N in there. Walk us through what The problem was you were tasked with it at Workspan. And then, you know, how did you start and how did it— why did you eventually pivot?

Yeah, so great to be here, Jared. Thanks for bringing the band back together. I want to give a plug for Go-To-Market Engineer School. We actually sent someone from Workspan through and that was how we found Octave initially. And we'd been looking for something that would help us execute this pivot. The company was going through kind of a product repackaging. We were having a bunch of training for the team at sales kickoff. And the marketing team needed to produce just way more segmented materials. And so for us, Workspan sells partner tech, and every partner uses the same software, but the way they use it is different based on whether you're doing a cloud partnership or a services partnership, even which direction that partnership is running. And I had like 2 people on my marketing team who were close enough to the product marketing to produce good materials. And then the rest of demand gen and the BDR org and everybody was just kind of stuck. And so what was a little different about my AI journey is I was actually not that deeply Claude-pilled. I'd been really trying to work on AI infrastructure for the company, and I'd say I was like a C-minus Claude practitioner. And the two go hand in hand. When we, when we brought Octave in and found Octave and it worked right out of the box for our use case, The way we set it up, I was going through— I don't remember what I was trying to do, but Zach and Julian were just like, why aren't you just ingesting your sales materials through a Claude code harness? Why aren't you just running this way? And I was like, I can do that. And so like, they're partly to blame for me now being fully, fully Claude-pilled and running all of Fresh Context out of 5 repositories. So we'd been looking at company context and trying to solve for the team-wide thing. My first instinct, exactly like you said, was to just solve it for marketing. And I was talking to the CIO and the head of sales. and making the case that we needed to bring in the software that I'd tried and knew worked for marketing. And the head of sales stopped the demo and he was like, wait, hold on. So this, this Claude instance now that you're showing me that you're using to produce these materials, like, can I just chat with it for a little bit? And so he started asking it questions like his sales team would ask it questions. And he said, okay, yeah, we're buying this and do what you need to do for marketing, but I want all of my salespeople to train out of this. So we, we crossed that line from departmental solution to team-wide solution as soon as the head of sales realized that Octave was going to be the PMM consultant for every salesperson as soon as you connected it to everybody's cloud instance. And he said, don't just use this for marketing, like everybody has to sing from the same song sheet. And so that was kind of our very quick conversion to, hey, here's this tool that's going to help me produce better specific marketing for the use cases that we're selling into. to this is actually going to help the whole go-to-market team look at these use cases. And if you see a great marketing asset, it doesn't work for the BDR who talks to you after you see that marketing asset to not know the use case and not be able to hold that conversation. And then for the sales team to not have the pitch and the deck and everything. So the shared context spread very, very quickly because it was so sharp and effective. And then the other kind of aha moment for me after, like you said, we'd been trying to use all the wrong tech for this. We, we were hacking N8N. We were, we were storing stuff in N8N tables that like stretched the limits and pointed to stuff. It was like so kludgy. And, uh, we always ended up with either a Claude project that didn't talk to other applications. And so you had this silo of really good context that people could go to and like consult the oracle. But then if you wanted to use it, like to, to generate something or to plug into an orchestration or something like you just couldn't, I don't know why Anthropic made that decision to silo Claude projects, but. they're only for people. And so immediately pulled everything out of cloud projects and then just had one place that was a shared resource. And once it's working, everybody wants that shared resource to stay sharp, right? So from that point forward, head of sales, head of product, and me leading marketing, we would get together whenever there was a campaign or a feature release and we'd say, okay, step 1, let's make sure that we've got it loaded into Octave and that the results out of Octave are good. And then we'll go build our training materials off Octave and we'll build our campaign launch and all this other stuff. So it went from being like, this point solution to the starting point very, very quickly once people realized how good it was to have everyone working off one set of files.

It sounds like you got buy-in right away, which is great. And is that because you guys had tried some things and they failed, or just the organization had seen the light? I'd imagine there's somewhere that friction occurred and you had to use your powers of persuasion to demonstrate small wins. What was a small win that got people to really see that this context layer could work across?

So I think friction, right? Friction is, is definitely a function of complexity. And, and I think also the, the just kind of the personalities and the mindsets in the room. And Workspan's CEO is a PhD data scientist, right? He's like a very, very sharp guy. He's a systems thinker. And he immediately understood the value of not repeating yourself, not having stale context in multiple systems. We talked about context debt. Right? And how much context debt we carried and all the systems and all the end-to-end flows that had hardcoded prompts that were going to need to be updated as part of our repackaging. And I was like, hey, I think we found another approach that's going to work better than this. And he got it right. He had been leading the push for, for go-to-market to figure out AI, right? He had given us space and said, like, I remember talking to him, we'd had a good quarter for revenue and a quarter where the pipeline was just a little bit shaky. And I was like, like, hey, I just— I want— I know we're spending a lot of time trying to figure this AI thing out. Like, am I good? Are we good? And he's like, you're doing the right thing. Like, the, the experimentation that you're doing, I fully believe it. So I, I was lucky. I had a very supportive executive team that was pushing us to go find solutions and immediately saw the, the benefit of a shared context approach. Um, at, at the clients we advise is totally different. At any other company I've worked for, I think it would be a much harder 2-step where first you, you have to take a sales leader or CRO or IT and people who are not fully bought in on AI or have different philosophies about adopting AI. And then I think you would also need to persuade people that we need to trust one team to do this. And, and, you know, I've come from big, big companies, public companies with thousands of marketers and thousands and thousands of salespeople. And it is, it is really hard for people to agree on what the right way to position a product is. Just when you get that big, everyone's got a different viewpoint of the world and trusting one team to say, no, we're going to drive this thing from, from kind of the top down, I think is much, much easier at a startup than a big company. So in Workspan's case, there wasn't a ton of friction, but I see it right at the other clients that we advise that it is hard to put in a shared layer. It's much easier to put in a point solution. And I think that's partly why you have so much fragmentation and people just solving for their, their little use case with AI and not thinking in AI systems.

Now, I don't, I don't want to wait till the end of this webinar today to like get in, get questions and discussion going from, from all the participants, because I think everybody here is probably having these discussions at their companies. I just closed the poll. I wanted— can you guys all see the results?

Yeah.

No? Okay, let me— I'm gonna share the results, and then I want people to, like, come off mute if you have questions or comments, put them in the chat if you— you know, what do you want to learn? What do you want to figure out in this? What are you struggling with? Let me figure out how to Let me find the share button again and show you these results. And I want to get your take on this. And if you're not speaking, you could put yourself on mute. But if you do want to come off mute anytime to ask a question, go for it. So you could see this now, right?

Yeah.

Okay, good. So it looks like, yeah, it looks like department wins so far. But hey, 6 people are saying they have a context layer that is like company-wide, and I want to hear from you guys and gals what you're doing, how you're doing it, what problems you're running into. And really, you know, almost nobody has not started this yet, which is great. And then, you know, I just put a handful of potential technologies in there, like what's part of your stack. And yeah, Google Drive. Google Drive is winning. GitHub, Slack. That seems, you know, I don't want to say universal, but that's kind of the shared experience, if you will. More kind of point solutions or, you know, open source projects, your GBrain or Trust Graph, a little bit lower down. A couple of people using Okta.

Good.

And then this is what I really wanted to get to is like, who's in charge of the company's context layer? And Nobody. Marketing team. People on this call. Good. The quantitative team, Qualtrics team. Not sure what that— so CTO, CSO. Okay.

What—

So folks on the call, what do you— someone who's like actually got this thing working company-wide. Anybody want to come off this mute and kind of tell us how it's going? Anybody brave? Somebody?

Yes.

Or if you have not, if you're struggling to get it going, or you just have it going in your department, Tell me what you— come off mute and tell me what you're using and what problems you're running into, if at all. And if nobody speaks, we'll go back to grilling Sam and Zach then with questions. But I definitely want to make this something where everybody gets a chance to ask the questions or chat.

Let me just poke on company-wide and maybe this stirs the pot and prompts some discussion. But when I said, look, we rolled out company-wide, like company-wide was go-to-market. It was the customer-facing team and product and engineering had different systems of context for managing the codebase. There was like a back office AI home and we actually, we split the Claude instances. There was a little bit church and state where there was a corporate license for Claude for product and engineering. And that made sure that, you know, code never leaked out into sales and marketing. And then for sales and marketing and customer success, where I was responsible for the AI transformation, That was where we rolled out. And in the end, it was 2, 2 kind of 2.5 contact systems, including Octave. But it wasn't company, company-wide. And we had Glean, we had some other tools that were trying to be company-wide. And we, we didn't find that the company-wide approach was what AI needed most of the time or what people needed most of the time. You wanted something that was use case specific, front office, back office.

And what do you— did you find then that people were trying to figure out, well, when do I put something into this system? When do I put into that system? Where does the thing live? Like, we now just have— we now have another set of surface areas where docs could be kind of lost.

That's definitely a thing. Can I show a picture?

Yeah, for sure.

So we— where's my share button? We kind of settled on this 3-layer context system. And this is, you know, what I've seen work. I've been talking to lots of people building company brains and what's their approach? And are they looking at it as concentric rings? Are they looking at it as layers? Is it individual department team? I believe context is a product. And I think the storage and retrieval of context actually isn't hard, right? That you can do all of this and get GitHub, you can just have files. If you want to go write the files and read the files, AI is great at that. What makes context valuable is that it stays current and that you have agreement on it. And sometimes that means that it needs to be updated very, very quickly and everybody can read and write to some types of company context that are valuable. And sometimes it means that actually you want like one subject matter expert to hold the pen. And if you're doing a product release, product marketing, tells you how to take that product to market, right? The collaboration that exists in the org creates a business process that context needs to tie into. And so we split it into 3 systems. We've got brand, which is a very squishy hodgepodge of everyone can write stuff on the marketing team that contributes to, you know, if we're running thought leadership on LinkedIn, somebody said something interesting over here. I want to take that thought leadership post and interview the CEO. And what does he think about that? And then I'll put a note into our brand context. It's just, it's every day there's stuff happening that changes the way that your company fits into the world a little bit. The canonical context layer, this is where we tried a bunch of stuff. We knew we wanted it to be shared. Octave was what we brought in and Octave's, you know, what Fresh Context uses for this as well today. That's a much more controlled, okay, we've got a product strategy, we've got a go-to-market strategy, we're starting a new motion. We're going to reach out to this buyer, this persona for the first time. How do we coordinate across all the different people and go to market so that that strategy reaches the end buyer, right? And that's not everyone should be able to chime in and have an opinion. You want a feedback loop. You want to be able to react to the reception that you get when you have sales calls or BDR calls or whatever. But it's not a free-for-all. And so there's governance and there's almost a test and release cycle for the context that we have in the middle layer. And then this last layer, this is How do you— it's Claude force, right? How do you have Claude go into your systems of record? And if you ask it a question in plain English, this is what used to take a team of analysts saying, you know, what is our, our forecast for Q3? How many deals do we have that are closing in Q3? You don't need to go and configure reports if you build your semantic layer on top of your deterministic systems. AI should just be able to go and get that. And so when people say, oh, well, we're building a context layer, I said, which context layer are you building? And it's funny, I ran Claude deep research on just how many B2B martech websites, sales tech websites mentioned context on their homepage. And 3 years ago, the number was zero. And today, it's something like 60% of anybody in go-to-market has to claim that, yeah, we're building a context layer. But if you click in on it, like they're doing very different things, it's a different product. And so I don't know that the right answer is looking for one solution to rule them all. I think you build the context to serve the business process and stay up to date. And then your question, Jared, is when you've got this stuff that goes in different places, it has to— this is why it all kind of goes together. You have to have the harness written with instructions for look here for this, look here for this. If you're thinking about writing, go here to write this, go here to write this. This is the process to do these things. So that was the way that we solved it. Otherwise, you're right, stuff writes on top of each other. People disagree about what should go where, and it's impossible to just keep it organized.

Adam's got a question. Go for it. Come off mute and ask.

Yeah, no, this is great, Sam. I think someone added in the chat talking about how different departments are building their own context. So sales builds their own context, customer success, like the sales cycle end to end includes all these different teams and departments, but the way a lot of companies are structured, For instance, I work for DocuSign, big company. They think of it as still siloed processes. So as you're thinking through this, you know, go-to-market context layer, how are you accounting for sales, customer success, all these different teams that plug into this?

Great, great question, right? And so I think to the point about not everyone agrees on the context, like, I think all positioning is a theory, right? All how do we sell this product in this market to this persona? And you can have the smartest people sit inside the walls and review transcripts and talk with customers and say, here's our theory. And it's the Mike Tyson quote, right? The plan only lasts until you get punched in the face. And so the sales team is out there getting punched in the face, trying to win deals every day. And so you have to have that feedback loop where this canonical context system isn't a dictionary that one person writes and then everyone else has to follow. This has to be a living thing that every sales call is adding to the intelligence of, and it's not, hey, this one customer didn't like the pitch, go change everything. That business process and that how does the org internalize feedback? And does that mean that maybe the pitch is wrong for this industry or that pitch is wrong for this buyer? Not this pitch is wrong 100% of the time. There's a lot of judgment that goes into those updates. And so I think having this snap into that, let's bring sales and product marketing, let's bring customer and product marketing closer together. All this stuff is being transcribed on every call these days, right? Everyone is getting that feedback in real time. Making sense of all that noise is a big part of why I would never try to run this in GitHub today, right? Like if, oh, we had a call, let's go search GitHub, let's compare the notes on that transcript with GitHub. Like this is a thing that just needs to be always on. This is a job for a really good platform.

Awesome.

Thank you.

I think, and maybe just to build on that too.

Sam, can you go through the— yeah, yeah, go ahead.

Go for it, Jack.

No problem.

I was just going to say, to build on that, I think it's really important that, you know, these systems, I think both when Sam and I, you know, no surprise, pretty philosophically aligned, like these are lower in the stack and those like those end jobs of like how the context is interpreted, delivered through an agent to a person that's gonna final, you know, go the last mile. Like, that's where you control a lot of that config. And like, it will look a lot different for a CSM preparing for a QBR at one of your top 10 customers versus a, you know, SDR that's doing pre-call research for a, you know, net new account in your SMB segment. You know, like, those are different jobs. And like, you'll have agents and tooling that knows how to connect to these systems. And there's The skills and agents and workflows and the harness is what is going to help interpret this lower level where we do need shared agreement into that end output and product.

You custom rolled the harness to point different types of questions or projects at different systems. Am I understanding correctly, Sam?

Yeah.

And then when then— so Octave was sort of one system and some— there is permission control there. Not everybody could just change our, you know, who our customer is or who, you know, what our product, you know, how we're positioning our products for different audiences, right?

They were—

but some of the other systems on the prior page or layers of the context stack there, where did they live? Like the deterministic customer context, I imagine that call transcripts and CRM records, those underlying sort of data systems are going to be tapped into from multiple teams, you know, not, you know, from customer success, sales, you know, even, even, you know, product to figure out, well, you know, where are we having, you know, the most complaints, problems, you know, etc. So how do you— where did you build that? Is that even the right question? Where did you build that to such that other teams can also access it? Or did you— or is the harness that you built for everybody, or is the harness just for people in go-to-market? How'd that work?

Yeah, that's, uh, 2, 2 good questions there. So deterministic, like you already have data that lives in databases, you have your CRM, you have your marketing automation, you have, uh, records, structured records that are out there, and you need to build a layer and you need to build a router, you need to put it somewhere so that if someone goes in Claude and they ask a question, Claude has the plumbing to find the right data source and then Claude has the right instruction. So when somebody asks this, what they really mean is like translated into this Salesforce report lookup, translated into this HubSpot bounded query, right? Like it's deterministic because there's a right answer. And if you evaluate the prompt and your forecast pipeline for Q3 comes back and it's different than what the CRO is looking on on the dashboard, something's broken, right?

It's—

you can go and debug it until the numbers line up and the, the answers line up. And when is the last person this account came to one of our webinars, right? That's a deterministic answer that lives in your CRM or HubSpot or Marketo or something. So that we didn't, we didn't build the data store. There's like a skill, or if you've got Snowflake or Databricks, right, you can build a semantic layer in your data warehouse so that the agent knows where to go and find those things. And it's consistent. For the harness and the routing, We definitely had kind of 2, 2 classes of people. We had RevOps and marketing and not, not even everyone in marketing initially went into Claude Code and like lived in this context-rich, everything's there. I've like got the whole company at my fingertips world. The campaigns team and the RevOps team kind of went first. And then once we proved this out and it was working, we scaled it up and then people were like, okay, it's not that scary. I can, I can move into cowork. I'm comfortable working out of a desktop. I want the extra connectors. I want the extra files and so on. Some people worked out of a desktop harness, and then most people worked out of just Claude chat. And through the corporate instance, we installed the Octave plugin, we installed the router, we gave it instructions so that you didn't have to be Claude-pilled in a builder and like in the terminal to use the different systems. You could just be a salesperson and have no idea what was going on under the hood and benefit from somebody coming in and putting a there is one right way to do this into your Claude harness. So I don't think the future is every salesperson moving into Claude code, right? I think you have builders and you have consumers, and the consumer should spend their time hitting the number out in the field, engaging with prospects. They should not have to worry about, is this context coming from the canonical system or the deterministic system? Like, they should just get what they need to engage with their accounts.

So the MCPs are just sort of entry level that you built on top of that to disintermediate, because at larger companies you have information that to the, the LM could look equally valid across different systems. And so you built a skill or layer on top that everybody can access no matter which surface area they're using in Claude, whether it's the desktop or Claude Code or whatever. So that when a question comes in, you're directing them to the right, or it's directing them to the right source for that question. Did that take a lot of— like, is that kind of like a constant? Somebody's kind of on that and tuning that and looking like— do you have to see what questions are coming in, how people are using it, so that you know where people are actually ending up in the wrong, you know? In the wrong alley so that you can then update your skill? How do you, how do you build that file?

So it definitely takes a little bit of, it's not like months of work, right? But you, you ask a question and the agent comes back with something. You're like, well, that's weird. Why did you respond with that? And especially if you've got, I think one of the problems people in the chat were talking about, Claude parses Google Drive, Glean parses everything, right? Like you can have too much context. And, and especially if you're looking for What is the latest positioning for a person in this industry who has this problem and use case? There is one answer. This is why we call it the canonical layer. And you don't want to go parse 7 years of company history on the internet and wikis and glean and yada, yada, yada. Like it's going to be wrong most of the time and it's going to be slow. So putting, putting the router in the harness that says, hey, if someone asks anything about how to engage an account, how to build a campaign, if somebody asks a go-to-market question, go here, get the latest and greatest answer, use the context system that was built specifically to stay fresh and give that answer. You can quickly see, did it go there and did it get the latest positioning or did it pull up a random sales deck that Google Drive had that Claude found from 6 years ago?

Right?

Like you can, you can debug this. The team who, who builds the harness has to run a bunch of test prompts and what, what's the sales team going to ask? Sales team, give me, give me 10 questions that you're going to want to ask Claude. And you run it and you just go look in the toolchain and what did it call and what did it get? And that's weird. You went the wrong way and then you tweak it and you improve the harness and it's just prompting, right? You just prompt Claude and you say, hey, when you found this thing, you should have gone this way instead. And you, you get it to the point that that harness works. And then every once in a while you publish an update to your harness and it just, you know, refreshes for everybody using Claude.

Yeah, gotcha. Yeah. Too much context. The LLM is gonna just, I don't wanna say make stuff up, but it's gonna sort of choose its own answers and everybody's gonna be kind of running in different directions. So that's why you have to have this sort of oversight layer.

Totally, totally. And, and I think especially for go-to-market, like good, good aligned sales and marketing is the holy grail, right? Like if you can have a point of view that is put out in the market and the thought leadership ties to the campaign, ties to the talk track on the webinar, ties to the BDR follow-up and the sales, like, The buyer has a great experience when it's like, hey, here's what we can help you with. And every single person and every single piece of content that buyer interacts with helps reinforce like, we can help you with that. And we've got our act together and we're here to help with this exact thing. That is hard, right? That's the holy grail. And the more you put in there, the harder that gets. And so we're not deep on the data structures and things underneath this, but one of the things that Octave does really well. One of the reasons I, why I love Octave, and I've, I've tested this with a bunch of different systems, is it knows where to go and it knows when to stop looking. And when you provide too much context and there's just an ocean of stuff that you can find, the agent has a really hard time knowing, did I check everything? It can't be greedy. It can't check every document on every search.

Right.

And so Octave gives a map to the agent that says, what is the motion that we're working in? What is the product that's being sold? What is the industry, the buyer? Right. There's an ontology And you can geek out on ontologies. You can waste a lot of time debating ontologies. A good one is opinionated and it tells the agent, go here, go here, go here, and then see what you've got and send it back. And so for go-to-market use cases, that's usually what you want is the one right answer. So everyone sings from the same song sheet and not let's look through 7 years of history and return everything that might be interesting.

I want to get to some of the questions in the comments here.

Sam, Zach, feel free to chime in here. A couple of people asked, like, so, okay, so this is more of like an advanced stage, essentially, like, how do you get started? What's the V1 of this? You know, if you're at a startup, where do you, where do you, where do you begin?

Yeah, I'm happy to take that one. Or Zach, do you want to take that one?

I would say, you know, where to start, I would, I would probably say don't try and build a company brain to start. Like, pick a problem area that matters. A company brain might be, you know, what's your PTO and HR policies, and what is— what are your sales team's, like, rules of engagement? And, like, there's a million things that are context. And so I'd say, first thing, don't do that. Like, that's probably not going to solve a business problem. That's a panacea that, great, we'll get to someday. But pick an area that really matters. You know, is it is it the kind of canonical context that we're talking a lot about today of like, hey, our products and how we think about our market and we want to build agents that are really smart and sensitive to that? Or is it, hey, we're a PLG company and like we got a ton of product telemetry data in Snowflake and like that's another type of context and maybe that's the problem to start with. So I would say my first thing I would Say, don't make the project build a company brain because it's going to go who knows what direction. You know, it's too big of an elephant to swallow.

I'll sharpen it one more time. And I think, I think I got lucky here, right? Not, not fully realizing how complicated and integrated this was going to be. But we started with, we need to launch all this marketing material. It's part of sales kickoff. It's part of our product launch as part of our campaign. And we need to keep the team coordinated. There was like a real pull. Right? We were looking for a way to do this better. And we didn't try to solve for every single thing. We said, well, let's load— and we worked with the Octave team. We took all of our sales kickoff materials, all of the recordings from all of the presentations, all of the workshops and all of the training things. And we're like, okay, we just got everyone in the room for a week and we hammered them with slides and we're hoping that they're going to leave and remember some of that. But just in case, let's put all that in Octave. And when they get back, they can just ask Claude and, oh yeah, that is what we went through. That is how I'm supposed to position this feature to a services partnership versus a cloud partnership. And so I think the more you can attach context to a motion and a thing that needs to get done, the sharper the context will be. People won't look at it from an academic perspective and be like, hmm, yeah, it could be a little better, and kind of just keep it in this like purgatory. They'll say, that's great, we need to ship this. We've got training next week. This is really helpful.

Right.

And so putting it in a position where there's already a momentum and a deadline and like it is part of an engage the market motion that gives you one, like you're not going to waste time just building stuff that doesn't get used. And 2, it's, it's part of an initiative that has to happen and the CRO gets and IT gets, right? It's attached to the business strategy and business outcomes.

So figure out the most important thing. Start there, start small.

Product launch, industry expansion, acquisition, right? Like there's, there's all these triggering events where now you need to change the way that you're engaging with the market. Pick one of those as a, as a kind of end around to fill in the context system and have it start with that specific thing to begin with. And then people will get it. They'll see how useful it is and you can always add to it, right? It's a, it's a two-way door.

Okay, then we got a couple more, I would say, technical questions. Adam asks, is there a preferred model for routing or pulling relevant context?

Like frontier model?

Adam, do you want to come off mute and say what you mean by that?

Yeah, I'm just curious, like when we're, you know, you're scanning through all this data transcripts, is there a preferred, if it's open source or frontier model that you guys use to do the routing and pull in context just in regards to token consumption?

So we, we have started playing with Kimi and we've been super impressed with Kimi and we, we're a Claude shop. We, we do everything out of Claude code and Claude has instructions for, I need to go parse a bunch of call transcripts, right? It's like a great brute force use case. We're not burning our Claude tokens to parse transcripts, right? That Kimmy can go and do that. And it's funny, even Claude, when it like gets the results back, it's like kind of, kind of salty about it. It's like, oh, Kimmy actually did a pretty good job parsing those transcripts. Like, um, whenever you have like a giant corpus that you're ingesting, yeah, it's, it's worth saving some tokens there, especially at your, your scale. Um, for the canonical stuff, like if you've done a good job with curation, You want a sharp model to help put it into the ontology and make those decisions. Like the write should be really sharp when you're, when you're, when you're going and gathering data. I think you can downgrade and use a cheaper model and save on tokens. When you're writing, use a smart model. That would be my recommendation.

And then Chris asks, how are folks solving for source provenance and entitlement?

Source provenance, like show, show me where this came from. And then what do you mean by entitlement, Chris?

So yeah, the provenance would be, you know, prove where this came from, prove that it's the most current up-to-date context. But then entitlement, if I've— if that customer's opted out or they haven't opted in, if there are internal policies that prohibit the use of that data for certain use cases, how do you actually ensure that contractual or compliance policy is enforced?

Yeah, I'll take the second one because that was a big deal for us at Workspan. And then, Zach, like, you could talk about provenance and all the stuff that you've built to show the chain there. But for customer references, this is exactly why we needed those partitions, right? Like, in the deterministic layer where we have every call transcript and we have all of the project delivery meetings and everything that the customer success team is working on, professional services team is working on. There's stuff that you don't want to show up in a white paper. And if, if context is just like this giant mass of everything's in there and you're relying on the prompt to say, don't use that, here's the list of things you can use, don't use it, you end up with this very messy system of like you're approving stuff and trying to track stuff and it just, it's too hard to keep it locked down. And the canonical context was our market-facing repo. And so, We would anonymize customer stories that needed to be anonymized and then put them in Octave. We would take our referenceable customer case studies and everything that was public and put them in Octave, but we would never let the like professional services week 6, let's go live and talk about how all your systems work together. That stuff never touched the endpoint that marketing would use to create market-facing materials and sales. Sales, you know, would generally work on the marketing rules. But if you got super deep in an account and you wanted to do a reference or something, sales would check in the other other context stores and say, who do we have that's doing something? Who should I go and ask about this to help do a customer reference? But it's a great point because it is a real problem if you don't have that partition where there is stuff that the company brain knows that you should not talk about in go-to-market.

I guess, and to build on that and speak to the provenance component. I think that there are some things that are, you know, it's impossible almost to say, like the provenance of positioning might be, I don't know, there's years of institutional knowledge that like needs to be input in the system at some point. And so there can be a record of like, hey, it was this doc, it was this chat session, it was, you know, this URL, whatever. So there can be a record like that, but it might not be the like capturing everything that's behind it. But then for sure, and like part of our system, obviously, as you move forward with a canonical context layer, you want Git-style, a who makes changes, what did they change, when did they change it, right? You want to be able to version all this stuff. And then obviously a system like ours, and to Sam's point, the real value becomes keeping this current and sharp in the face of your market and your competition and all the things changing around you. So like Octave will produce, hey, you know, you said, you know, your canonical definition says when you're in this market selling to this type of buyer at this type of company, you know, the lead value prop is FTE reduction and that's where you want to save money. But then all our conversations like this buyer and this CFO, they don't really seem concerned about that. It's, you know, supply chain resilience. And so a system like Octave will come to you and say, like, hey, we have a mismatch here. Like, there's an incongruency. Here's what should be changed. Here's the record of, like, where we saw this on all the calls and all the people. And look at this, it was primarily voiced by the prospect side. This wasn't your team just changing the story, which maybe sometimes is valid. You want to say, like, okay, the team's changing the story and looks like it's working. But you want to see all of those factors of this data that it is qualitative in nature, and you need to have the provenance to have good judgment around it and then go thumbs up, yes, thumbs down, no, or, ooh, almost got it. Here's a little bit of extra human-in-the-loop input. And so having all of those workflows and the lifecycle management around this is a critical piece to making it a foundation you can really build on.

Yeah. So you, you can always see if you're in the agent, if you're in Claude or Claude Code, what, uh, what nodes did you crawl in the graph? Where did you go and look this stuff up? And then within Octave for that node, where did that node come from? Who wrote it, right? Was it the result of a, a product and, and sales transcript getting loaded in? Is this a doc that was put there, right? Like the team that, that owns the right owns putting it in Octave, and then the harness and Claude can always see like, where do we go and find this stuff?

And then we're talking here about internally, and then Jennifer Roberts asks, how do you share this with partners that benefit from the most up-to-date messaging and marketing materials as well? And I imagine Workspan, you know, dealt with that, you know, in lots of aspects being around, you know, co-sell pipeline. And then Claude Walton on the On the call here also asks in a sort of related question about, you know, aligning a team of 5 people, but then also external resources. So how do you, how do you take some of this, you know, curated knowledge here and then share it externally?

Yeah, I'll take the partner question. I think I have become a full partner convert, right? Partner marketing is marketing. Marketing is the ecosystem. And so loved getting deeper into those use cases while I was at Workspan and supporting people building these things. And Amit Sinha, our president, I was at a conference with him yesterday, and the way he put it is context stores are cheap. Like, storing this stuff is cheap. It's the tokens that you pay for. And every partnership has a reason for existing that's specific to that partnership. And, you know, when we were looking at Octave, I was like, hey, Zach, hey, why isn't there a partner section of Octave? And at the moment, there still isn't one. And so there are other tools I think that you would use. Workspan is one of them. Where you could look at how do we take our positioning, how do we define the better together offering, the specific thing that we do with that partner, and then pushing that outside the walls is like its own challenge context that travels and their sales team can see information about how your product and your company should work with them on a deal. It is a different problem space. And so there's a couple of vendors in that market, but Workspan is definitely one of them. And if you're running partnerships at a scale and complexity where you have to change the behavior of lots of salespeople who don't work inside your walls, like AI has a huge role to play there and not on my graph, not on the slides that I presented, right? It's, it's a, it's, you've got to get outside the walls.

Thank you.

And the, the partner object is coming soon. We're just, we're just trying to get it right. There's a, there's a big difference between is this a technology and integration partner, is this a resale partner, is this a services partner? So we're doing it the right way so it fits in the graph and I think people will get what they want. But we have seen, just to add some more color here and kind of like, I think how we're seeing other teams think about the architecture is we do have other customers who great use case for Claude Code, Lovable, your tool of choice. Build a skin, like build a portal for partners. That's like very easy. Go, go let Claude be great at that. And then you go, hey, you have access to these things in Octave. Here's how we talk about these markets and these partners and these products that I want. And like, it's a really easy way to like, great, we have a shared single source of truth for this. And then I want to like reinterpret it in a different kind of, you know, like we were talking about earlier, that agent, that output, that form factor. Those are like where I would say do your building there. Like that, that is the stuff to vibe code the heck out of. Sorry, if that was a question for me, I didn't quite grab it.

That got a little garbled. Can you repeat that, Taha? Okay.

That sounds like maybe another call.

Is that it?

Okay, little mute button solved that. I want to cover off on the other part of Claude Walton's question here. Claude, everybody's got to use your last name now so we're not confusing you with the AI. But in terms of a small team and getting going, we touched on this a little bit earlier and I know this is not exactly the same problem that you solved. that works for me, which is a bigger company. But Zach, you must see this. If you've got some MD files and Markdown files and Claude, what's the next thing you would add on to that if you're a small team? Is that your question, Claude?

Yeah. I work with a lot of startups, so just trying to get a mindset, what would be a V1 of jumping in, getting them organized for somewhat from chaos to align go-to-market, even with a very small team? You have graphic, you know, you might have designers or people doing paid media, but as a product marketer, I've always been in the core, you know, doing a document, teaching it to other people. But what would that look like just as a starting point on which you can build and scale the company with and not waste the time or have a debt for context right day 1, right?

So I think, I think the startup should not build capacity they don't need, right? You don't have time. And, and it depends on what are people using. If the whole org is product and engineering and there's like one person running go-to-market and everybody's working out of GitHub anyway, you could probably just manage it in markdown files. You don't have 7 years of history, you don't have the enterprise problem, and you don't need a live system that's updating and listening to gone calls. It certainly helps once you start going to market at scale and velocity, and you have 5 salespeople out there having conversations every day. But if you're kind of in founder-led sales, everybody fits in a room, I wouldn't overbuild it. I would wait until something feels like it's breaking and then context as a product, upgrade the product to meet the business need.

Yeah, I wrote in there that it felt, you know, starting point is a markdown file of product marketing deliverables in Claude, right? As a starting point for guiding.

If everyone gets that, there's some— we just did this with my team this week. There's some, some cool hacks where you can have a Git repository that live loads some other Git repositories. There's like pointers, so you don't even need to do a pull if somebody's pushed a set of plugins or skills or context. Other people can get it in real time. Or we use a hack for some of our non-Octave context that just lives in a wiki in Git. We have a daemon on everyone's machine that auto-pushes and auto-pulls all the time, so people don't even need to worry about it. And so, I'd say for a startup, just whatever gets the latest current thing shared, and then whatever business process you need so that there's feedback and it's not a one-way push. people ignore the answer and do something different because, oh, well, for me, for my use case, that's not there. If you're a small enough team that you can just talk about that and keep things fresh, I wouldn't overbuild it.

Okay.

Yeah. And I'd add to that, look, you know, the routines in Claude, amazing. You could schedule things every 10 minutes, every day, whatever you want. So you got your right connectors, you've got your routines going, And then even, you know, to manage the PR process, you know, GitHub Actions to auto-merge your PRs. If there's no conflicts, there's auto-merge that. So a lot of this stuff can be kind of smoothed out, I think, for when you're just getting going and you don't need all of the, you know, other stuff, you know, enterprise-level stuff that we've been sort of talking around. We are almost out of time here. Zach, Sam, any kind of last thoughts you want to— or resources you want to leave people with?

I think start building. I think, yeah, like everybody, like, I think to Sam's point that he just made, you know, start what feels right. Like, if people are working on local and that's not broken yet, Don't introduce context as a forcing function in and of itself. It sounds silly and pedantic, but focus on a problem you want to solve. Then I think context can become a true enabler to an agent you want to build or something like that where you have something really discrete. Yeah, don't overthink it. Start with what works. Then yeah, if people want to learn more about Octave, chat more, find me on LinkedIn. I'm always eager to talk about this kind of stuff. So appreciate you having me.

And my closing comment here is I'm at end of 2, kind of 2 and a half, watching these contact systems go to market and watching how they unlock AI and accelerate. I feel like there's this extreme pressure now where products are shipping faster, right? Go-to-market strategies are changing faster. AI has accelerated everything about the market and all of the internal context, human-to-human context transmission, getting people on an enablement call and taking them through slides and then hoping they remember it and having their manager follow up. Go-to-market is struggling to keep up. And so AI in go-to-market, I think, is crucial for any company that's using AI in product development, anyone that's selling into a market that's saturated with AI. Like, speed is critical. That's kind of my learnings from the last 2 years. And then All of this stuff, I think we will continue to evolve the best practices. And it's not, you know, what is Astra capable, what is Fable capable of, when Anthropic drops their next model, what are the agents suddenly going to be able to do? It's the interface between AI and the organization. That's the hard part to build around. And that's why you need these good systems and people need a mental model for how this stuff works and people need a way to keep up. And so who knows, maybe the next version of Anthropic can just read everything forever and come up with better positioning than every one of your product marketers, and none of us will have to work anymore. But I think it's far more likely that we're going to have a long time where the boundary between how do people coordinate and how does AI play an accelerating role in people coordinating, that's the game for everybody in go-to-market to work on.

And Sam, you just published a paper on this, right? A research paper on this.

We kind of launched our consulting practice with a manifesto where we said, here is exactly what go-to-market teams need to do to accelerate. And it's everything we've been talking about. It's building shared context, it's task clarity and teaching the agent what you need to do in the different channels and ways you go to market. And then it's team. It's how do you get people bought in and have a shared harness and It is a different way of working to have AI in, in as a connective tissue. So yeah, I can drop that in chat. Would be happy to drop that in chat if that's cool.

Yeah, yeah, do it, do it. All right, well, thank you everybody. I hope to see some of you in our program starting the end of the month and appreciate everybody jumping in with questions. It was a great discussion. I learned a lot. I hope you all did too and have a wonderful rest of your day.

Thanks, Zach. Thanks, Jared.

Take care.

Have a good day, everybody.

See y'all.

Bye.