I built a custom Slack inbox. It was easier than you’d think. | Yash Tekriwal (Clay)

I truly wake up to maybe 100 to 150 new Slack notifications. Not even just like, oh, these are unread channels. Truly someone has tagged me. 60 to 80% are more in the FYI category. So my 100 to 150 that's giving me anxiety is actually more like 30 to 40 that I really need to be on top of.

You can use AI to do a task for you, like categorize things, summarize things, or you can use AI just to build a tool that would've been much harder to build before with very straightforward APIs and structured data.

Exactly. Think about like a Kanban-style board. You have in red on the left, action required, urgent, Yash needs to get back to it. In the middle, we've got a yellow need to read column. And then on the right in green, much more easy, I have a bunch of FYIs. I can just go ahead and click this archive all button. They'll disappear from the dash, and then those notifications will also disappear on my Slack.

Ugh, that's magic. And this is such a better way to just get through your queue.

My dream is for someone else to watch this video and say, I want to build that app on top of Slack. And then I can go pay that person $15 a month for this app to be maintained. And used, and then I can file bug reports with them instead of having to fix it myself, because I would happily pay that.

Welcome back to How I AI. I'm Claire Vo, product leader and AI obsessive, here on a mission to help you build better with these new tools. Today we have Yash Thakraral, head of education at Clay, and he is a hyper optimizer showing us how he uses Perplexity Computer to work through the hundreds of Slack messages he He gets every day. We're also going to debate, is SaaS really dead? Let's get to it. This episode is brought to you by Guru, the AI layer of truth for your company's knowledge. Here's the problem. Your AI is only as good as the information you feed it. Most companies are getting confident but wrong answers from AI because their underlying knowledge is outdated, incomplete, or just plain incorrect. Bad information doesn't just slow you down, it costs you money and puts you at risk. Guru solves this by adding a verification layer between your company's knowledge and AI tools. Instead of just hoping your AI gets it right, Guru automatically scores content for accuracy, flags outdated information, and ensures your team gets trustworthy answers every time. It works with the tools you already use so you don't have to change how you work. Thousands of companies trust Guru to keep their AI accurate and compliant. Ready to stop playing Russian roulette with your company's knowledge? Visit getguru.com to learn more. Welcome to How I AI. Yash, I'm so excited. We've been trying to make this happen for months and we've been trying to make it happen over Slack for months. And what I love about that is what we're going to start this episode off with is how you get yourself unburied from the deluge of Slack messages and emails and work you have to do on a daily basis.

Yeah. So I wish I could say I had built this back when we were organizing it, but maybe that gives you a little bit more leeway for understanding me losing Slack threads all the time. For context, you can see my Slack screen right here, right? Right now, I cleared this truly 2 hours ago and I already have like another 40+ messages of which you can see like 8 or more are DMs. Uh, and it's just gonna keep going up right now. I truly wake up to maybe 100 to 150 new Slack notifications, not even just like, oh, these are unread channels. Truly someone has tagged me. It's a DM, they've tagged me, it's a group DM or something else, which all feel very important. But not all notifications are created equal in Slack. For example, I care much more about getting back to you on scheduling our podcast recording than I do about my colleague's really fun comment on their dog that they posted a photo of in the fun dog channel. But I get an equal notification for both. And so what I sort of started doing with Perplexity Computer when it came out about a month ago is thinking if I could truly just design any software or like paradigm myself, what would I do, how, and why? And so Perplexity Computer is actually not exactly how we initially solved the Slack problem. We'll come back to that. In just a second. But I think the framework is also what matters most, is I needed to be able to envision what does a better world look like instead of just asking Claude or Perplexity or Open Claw, make my Slack easier. And so that better world that I thought of is if not all notifications are created equal, then what if I could better categorize my notifications by DMs versus group DMs versus threads? Versus group @mentions, because I treat all of those differently. I try to clear my DMs ASAP because I tell everyone if I'm not responding within 24 hours, DM me. So that urgency needs to be there. But then on top of that, of course, people are DMing me about random things like who wants to go dancing this weekend, who wants to go to dinner for hot pot this week, and all sorts of other fun things. So even within each of those 4 categories, right, the DMs, the groups, the threads, the @mentions, I also want to subcategorize by what requires real action from me, what do I need to read but maybe doesn't need a response from me, and what are more of the like FYI, for your information notifications. And as you might guess, a little precursor that I'll give you is like 60 to 80% of my notifications every day are more in the FYI category. So my 100 to 150 that's giving me anxiety is actually more like 30 to 40 that I really need to be on top of, and that makes things a lot easier, but I need to build a system to get there.

So if I could repeat your problem back to you, you are very important at work and also super fun and popular, and this is just causing you tons of @mentions. And you're gonna show us how you use Perplexity Computer, although I think you started with something else to kind of solve this problem and prototype your way out of it.

Exactly. I think the other, like, general framework I'll give here is that I do think something that I coach a lot of people on my team about as well is when to use AI purely like an MCP even, versus when to use AI to build something deterministic like code or an API call. And so for example, here in Slack, there's enough API endpoints available that we could get into that I know I should be able to build the information organization system just using code. And so actually I did that via my little— we can do a detour here, Jarvis, my Open Claw. Let's see if I can even find my— here we go. Thread. It's a very long thread that we can come into all the way up on Slack. I'm not going to go through and do all of this, but you can sort of see in here at a quick glance, it's building the digest for me, right? So it's looking at, okay, what is the timestamp? What do I want to mark? For example, Slack has a whole info chart that we could find online and then put in this if we wanted to on how they've built their notification system. It's very intentional whether or not it's unread or gives you a number or how many numbers it gives you is all dependent on the stream in which it comes in. And so for me to actually pull not every single new message in my Slack, because that would be overwhelming, but to only pull the ones that I care about with only the context that I care about requires a little bit of just systems thinking. I need to look at what was the last time I looked at the message? Did I look at only the most recent message or have I looked at another message in the thread before? Because then I don't need to see the whole thing. I just need to see only the most recent messages. So all of this data is tracked in Slack via these timestamps. And you can see me going back and forth with Jarvis here for a long, long time on what is unread, what is not, how do I look at these channels? And so this back and forth, which truly goes on for like thousands of messages, so we won't do all of it, took me like a full day to really prototype and build and understand to get to a point where, to be fair, now those 100+ notifications Come into this Jarvis Digest channel in Slack. It does group it into direct @mentions. You've got those 3 sub-buckets. Then we've got DMs, then we've got group mentions, and then we've got threads. So those are the 4 overall buckets. And then within each of those buckets, I could now, and this is what I did for a week, just command-click into each of these, open up a new thread, Decide when I need to respond, come in, respond, and then go back.

Just to kind of narrate for people that are maybe not watching the YouTube what we're seeing here, you basically said, look, I get this all-purpose inbox from Slack with notifications and unread. I get hundreds of them, of which maybe 2 dozen are actually interesting to me. I have a pretty clear sense of how I want those organized and prioritized in my own workflow. I'm gonna spin up OpenClaw as a coding agent in Discord. I think I spotted, spotted Discord.

Correct.

And why Discord real quick?

So my one reasoning for Discord, I started on Telegram with OpenClaw, like I think many people do, but I really like the threading nature of Discord for organizing all of my chats. For example, in Discord, my favorite command similar to Slack is Command K, and then I can quick find, quick search any thread that I want. I can pick up the context where I need to, and I can open and close those threads to keep it really clean in terms of what I want to stay active and what I want to stay inactive.

Great. So you used OpenCLAW and Discord to basically say, hey, I know we can reverse engineer how Slack determines what unread messages are and how they categorize them, which would be incredibly painful as a human because I'd have to go like knee-deep in the docs, Read a bunch of stuff, memorize all these codes. Super annoying. You go figure that out. We're going to do a back and forth, and then we're going to build this automated Slack digest feed that pushes every day a very targeted list of unreads grouped by, again, things that are coming directly to me, things that are coming to a group, things that are in channels I care about for work, things that are in channels I don't care about for work purposes. And then even within that, they are prioritized and then they're deep linked. So you can actually go into that deep link, take action, bop out and work through your, through your inbox almost like a prioritized email inbox or a task list.

Exactly. The only thing I would add to your great narration is that it's important to also note that pretty much all of what you just described, while built with AI, The only place in which AI is repeatedly used in the system that was constructed is in the categorization of messages that need action from me, need to be read, and are FYIs. Everything else is custom code built using an AI assistant, but it's truly still deterministic in the categorization that gets put out into the digest.

And that just goes to your earlier point of like, you can use AI in these 2 ways. You can use AI to do a task for you, like categorize things or summarize things, and or you can use AI just to build a tool that would have been much harder to build before with very straightforward APIs and structured data.

Exactly.

Okay. And then, you know, this is, this is great. It's still, you know, to my, to my naive eye, it's still pretty overwhelming, right? It's like a really long list of things. And so what was the next step you took here in terms of making this even more usable for you in your day to day? Yeah.

So for context, right, for anyone that's not watching the YouTube as well, this message is helpful, but I have to scroll at least 4 to 5 screens down just to be able to even see all of the notifications that are already summarized for me. So you can also imagine just how much worse my actual Slack is. I used this for like a week and it was more effective, but it was also just draining. And so then my wish became, what if I just had an actual software that I could build on top of this that looked clean, Felt like Superhuman for email, had navigability, and let me sort of categorize this long digest of just text and emojis into a real interface. So that is where we come into Perplexity Computer. And so I will pull up Perplexity Computer here, and you can actually see right here, this is the digest that I ended up building. What we'll first do is even show you. The thread in which I built this digest. So this one also had a fair bit of back and forth, but 80% of what you just saw, which I'll describe for people not watching in a second, was built in like the first 4 messages. And I do think this is actually worth digging into some of the guts, right? I'm asking it to analyze the message structure in the digest channel because via connectors and the browser, it has access. To everything that I'm looking at. And then what I think is so good about Perplexity Computer that is unique to Perplexity and not Claude or OpenAI because they're frontier model companies is that they are shameless about using all of the different AI models to build each part of the task in subsequent order. So you can see here for fetching the, the digest, it was using Sonnet 4.6. Then you can even see down here, I believe it starts running different tasks in parallel. And you can see that it's using Gemini, I believe, for planning at some point. You can see that it's using Gemini for coding Python. You can see it's reading different skills that it's building. It then uses Opus for the actual build because that's more intense and it wants a little bit of a better layer of reasoning. And I could just keep going down and down, but all of this work happens from just my first message. So there's an additional layer of troubleshooting and intelligence and sort of testing for Am I building the right thing? That removes the need for me as a human in the loop to constantly go back and say, uh, you tried, good job, but also it doesn't work. Like it literally doesn't even do what I asked you to do. Try again. That frustrating, I have to reprompt you loop over and over again is much, much better with Computer because of how they've built this ensemble orchestration.

So let's take a minute to talk about Perplexity Computer because we haven't actually seen anybody demonstrate this tool on, on the podcast. And so why, you know, above dropping this in Claude Code or Codex, which as you called out, has its limitations in terms of being single provider in terms of model, you know, you showed you used OpenCLAW as well to do a similar task. What has drawn you to Perplexity Computer and what do you think is unique about the setup?

Yeah, it's a good question, right? So I think there's, probably 3 things that really draw me to Perplexity Computer, ranging from like simple to a little bit more involved. The silliest but simplest one is running things in parallel, right? In Cloud Code and Codex, I'm still having essentially a chat-based back and forth conversation. And so I'm doing one thing at a time, but often I maybe want to kick off a task. I could do this in separate terminal windows if I really wanted to for Cloud Code, but it's annoying. I can actually just kick off, you can even see here in the screen, right? I kicked off like 4 different tasks. in the span of 10 minutes of each other this morning and then had them all running at the same time. So that's one thing that I think is very simple, but concurrent runs and long-running tasks are really nice in Perplexity Computer. Number 2 is that Claude, Cowork, and Codex sort of have these connectors so that they can access your different apps, but they're primarily built for code generation. They're built for app making. I would actually say Perplexity Computer, while used often to build these UIs and tasks, and like those are the fancy things, You might see if you go to use cases here, like a lot of people are building apps because everyone thinks in apps. It's actually far greater than that, sort of in the same way that people discovered Claude Code can do so much more than just coding. It's a really good local agent. But the reason I like the Perplexity computer element is actually because it's in the cloud. Because it's in the cloud, it is much more natively connected to all these different tools. I don't have to go give it the— in the way that I give OpenCLAW, right? Like access to different skills. to be able to go do things in software. It can just go do it and ask me for logins on my own. And that native connection to all those tools helps me be able to do things much more fluidly at the speed of thought because everything is being recorded everywhere. So like a really easy example that I actually deleted because it was an accident, but I can just talk through real quick, is I also use Notion AI to record all of my meeting notes. Perplexity Computer is connected to Notion AI. And so what I am always doing in all of my meetings is saying, hey, here's a bunch of follow-up things that we should probably do. Let's make sure we get them all done. I try and put all of those most important ones in my task management, but because that transcription exists, it's in Notion and Computer has access to Notion. I can then also go into Computer and say, look through all of my meeting transcripts from the end of the day today, gather the list of action items, Categorize the ones that you think are important, throw them into my Asana if they're longer-term action items. And actually, for anything that's just a message or a notification or an email, draft me the response.

Can you show us what connectors you, you are using regularly just so people can have a sense of what you're talking about? Oh, you're using like all of them.

So many. I mean, not even close to all of them, right? But to, for everyone listening, I've got the bare basics: Google Drive, Gmail with Calendar, Notion, Asana, Slack, Forms, Tasks, Typeform, Zoom, Spotify, which I haven't really used, but I thought it would be fun, you know, Airtable, Google Slides. And then there are tons of other ones I'm not using, like Linear, Supabase. I've looked at trying to use maybe like Snowflake more, Datadog more, but I'm focusing really just on what am I actually getting value out of today instead of chasing. I have what I would call shiny object syndrome. So one day I'll maybe get back to all the shiny objects, but for now, this is what provides the lion's share of the value back to me.

Great. And then just to kind of like loop the thread back to what we were originally talking about. So you've shown us Perplexity Computer, you really like it. Multi-model, multi-threaded concurrency, nice to use, lots of connectors. And then you flashed to the beginning, but let's just show what you built for that, that Slack management tool, because I think it's really cool.

Yeah, so I'll, I'll do my best to visually describe this for those not watching. You can imagine a dashboard UI. It looks sort of like any task management tool you've had before. Think about like a Kanban-style board.

Mm-hmm.

But instead of multiple cards for all the sections of tasks that you've got, you've just got 3 main ones. You have in red on the left, action required, urgent, Yash needs to get back to it. In the middle, we've got a yellow need to read column. I should make sure that I'm understanding what's going on here, but I probably don't need to respond. And then on the right in green, much more easy. I have a bunch of FYIs. Hey, here's what dinner is. Here's where the address is. Here's what someone is doing for the launch that we just planned. And so then there's a bunch of other smaller dials that I've customized on here. So I can group this like we were talking about earlier by order of operations. Do I want to go through my DMs first? Do I want to go through my group mentions first? I have a sidebar on the left that I can categorize those by, and then The best thing about this dashboard for me is that I use this all the time. The FYI notifications on the right, I can just go ahead and click this archive all button. They'll disappear from the dash, and then those notifications will also disappear on my Slack.

Oh, that's magic. Because I know you can in Slack, I think you can do like mark all as read, but you can't be like, mark my DMs as read or mark FYI as read or like multi-select. And so you're sitting there with either like 100 unread messages or zero. There's nowhere in between. And this is such a better way to just get through, get through your queue. And one thing I have to say was like, there's this incessant debate of like, is SaaS dead? Is this like SaaS apocalypse happening? Like, is who's going to build the new Slack that's better? Is Slack the new Slack? And what I love about this moment is like Slack is still great. Like it's—

It's so good.

It's great for sending messages. I don't know if it's great for reading messages, but it's great for sending messages. And you can, with very low effort, say my company is using Slack, we're happy with it generally, 9 out of 10. And to get it to 10 out of 10, I'm just going to build the thing that works with my brain. And it doesn't even have to be about a deficiency of the existing software. It just has to be, you know, closing the gap between your ideal workflow as an individual and what you get out of SaaS. And it's— I just think it's such an interesting model as we think about like, what are these, what are these productivity tools going to become? What are these collaboration tools going to become? Is there like, you know, Slack Core and then Slack Custom?

Right.

And you just build on top of it.

I, yeah, I so agree. I would actually say maybe my very hot take here, I think we will see an explosion in software being created and used because of all these tools. I don't think the average person or even the proficient person is going to start custom by building and coding all these tools. I don't think the intelligence of AI, at least the way we're seeing it ramp currently, is getting to a point where my mom could go in and say, make Slack easy, and then it builds this. I think instead what will happen is you sort of, like you were saying, optimize these tools. And I'll be so honest, if someone else— my dream is for someone else to like watch this video, look at this and say, I wanna build that app on top of Slack, and then I can go pay that person $15 a month for this app to be maintained and used, and then I can file bug reports with them instead of having to fix it myself, because I would happily pay that.

I, I agree. And it's— I, what I think is so funny is I'm like a B2B girl. I've done product for a long time. There has always been this long tail queue of like customer requests being like, I would really just love it if that button in the bottom left of this page did this very niche thing for my specific vertical. Like, that would be great. When's it on the roadmap? And the answer is, a reasonable SaaS PM is like, never. And now you can kind of like replace this concept with Okay, point a forward-deployed engineer at that and say like, we're never going to build that for you. That's a you problem, babe. But let me show you how you can get that on your own. And then again, yes, I agree with you. I think kind of like 2 things are going to happen. One, just so many more creative opportunities for software to be built, just as you said. Like somebody maybe will watch this and be like, yes, I'm going to build like Slack Digest as a product. Um, the second thing is like, there's probably like TAM of this big for that product. And because the cost of building it's so low, who cares? Like somebody could go turn it into a $10K a month like project or a $20K a month project. Doesn't have to be venture scale, doesn't have to like hit a billion dollars, doesn't have to have a million users and can still like be useful.

Yeah.

And I feel like there's so much useful software that we could have Because it's been so expensive to build and not worth people's time. And I am just, I'm with you. Like, let, let's go do more.

Agreed.

Yeah. And then, you know, and then Slack can do the thing where they just go like this and like scoop 'em up.

Exactly. Slack can then acquire all those people. I think you'll see a Cambrian explosion of like businesses that wouldn't have existed without venture funding or that people wouldn't have wanted to build that can make, make their own money or get acquired by larger companies. Yeah. Who knows?

Yeah. That's— I mean, that's basically what I'm doing right now. So it's like, I'm living it. I love this. This is genius. People take this idea, think about the most— I often tell people to build their anti-to-do list and then spend an hour a day burning down that list, which is like, I never, ever, ever want to go through my Slack unread again in an unprioritized way. That's going to be on my anti-to-do list. So I'm going to spend an hour a day Trying to figure out how to like dig myself out of that problem. Another one like I have is I never, ever, ever wanna delete spam out of my email by hand. Like I've, I so often have to go through and like click the checkboxes and delete. So I never want, so like how do I solve that problem? I never, ever, ever want to like manually, you know, hand to Asana, enter my action items after a meeting. And so this idea of an anti-to-do list, which I can share, I have a, like a list of like 100 ideas that could go on an anti-to-do list. And spending time with AI to automate those is such a worthy use of time.

Totally agreed. And I think computers may be another step in the journey of people just being able to do more with less.

Yeah, it's just, it's all getting better. This episode is brought to you by ThoughtSpot. Product leaders know the struggle. Your users want data insights, but they don't want to leave your app to find them. ThoughtSpot Embedded solves this by putting analytics directly into your product. Your users can search in plain English and explore data instantly, right where they work. No separate tools and zero context switching. What sets ThoughtSpot apart is that it's not just another bolt-on dashboard. It's a search-driven, AI-powered experience that feels native to your app. Developers can embed it with just a few lines of code and then fully customize the look and feel. The result? More engaged users, faster decisions, and a product that delivers more value every time someone logs in. If analytics is becoming core to your product strategy, visit go.thoughtspot.com/howiai for more information and try the free trial at go.thoughtspot.com/howiai/trial. Okay, so you are a pro user. Any other use cases for Computer that you think are useful? Any other like beautiful UIs you've built that you wouldn't have been able to do before?

Yeah, so I'll actually show one other one that I've built, and then I'll show one that like someone on my team has built as well, 'cause I think I'm trying to get everyone to be thinking about and using these tools. And they're both sort of different use cases. So I think mine unsurprisingly falls in the same thematic form of Yash has too much information and Yash wants to consolidate the information and be really helpful as quickly as possible. So what I can show you here is similar to the Slack digest. I first just wanted to make sure, okay, am I actually getting the right things into this digest before I spend more time trying to make it pretty? And so I have a bunch of Slacks and emails. I realized I stopped reading the news because I've been so deep in Slack and email, so I want some news in there. And I could actually, and probably will, add a couple of other communication streams to this long-term, but I just wanted to get those 3 in there. And so I got a couple of days of the text working in here, and then I was like, okay, the text, similar to Slack, also boring. Why not code up a UI? I will say, computer seems to default to Kanban-style pretty UIs. We've got another 3 Kanban cards here. It goes in order of AI news, email, and Slack in terms of what needs my actions. Obviously a little bit of a Venn diagram overlap with my Slack Digest bot, but if I want to consolidate to just like truly mission critical, this is also really good. And in terms of talking people through what the iterative process here looks like, I can tell you, right, step 1 was getting the text in here and then saying, oh, that's actually not an important message. Can you tell me why you categorized it that way? And then explaining to computer how to rebuild the background skill it's using. To make all of this categorization. Step 2 was then getting the UI in here, and then step 3 is actual usability. So when I had the UI in the first place, it didn't think to default link me out to all of the messages so that I could go back and look at them. So I asked to, and now on the Slack notification, I can link back out to it. I have to go back and add it for email and then add it for news, but that way it becomes a real command center of all the things that I want to go see.

I'm gonna give you one other fun tip, which is I saw somebody on X posted that they had their Claude build a like newspaper style digest of all the things that the Claude, like that Claude code had done. So maybe you could put this in like a fun newspaper style where you like wake up every morning and you read, or like a magazine, something really fun. You wake up and you read. Yeah.

Extra, extra. Here's how Yasha's day is going.

Exactly.

I like that idea a lot.

An image generator.

Send me that thread.

Yeah.

Yeah, yeah, yeah.

Okay. So I, I really, I really like this. And again, you know, we've seen so many of these like daily briefing use cases and we've seen everything from like, I get, you know, markdown, very cold structured markdown in the terminal from Claude Code that I like store in Obsidian because I'm like an esoteric, you know, dark mode thinker to, you know, I just use like ChatGPT Pulse and every morning it sends me something cute and nice. To something like this, which is like I've built myself an app that again matches my mental model of how, how I wanna use this and is optimized for, for me. And I can actually use as not just, um, information, but as an interactive kind of command center for my work.

Exactly. And I think the thing that you just reminded me of, like 2 other thoughts I'll mention here. Number one is the other benefit of Computer for the non-technically initiated Is that even if I were to go code this up in Cloud Code or Codex, I still have to go like put the GitHub repo up. I still have to deploy it to Vercel. I have to still make sure that like it's all working in production in some way, shape, or form. Not that hard to do, to be quite fair, but still just an extra technical layer I think that has not yet been removed from usability of that type of software. For Perplexity Computer, it's already deployed, it's already live on the web, and you can see this button in the top right corner. I can just share it in the same way that like I wish Claude Code built me something that I could share like Claude Artifacts does if I go in Claude Chat. But again, my frustration is I don't wanna have to remember which one of Claude's 7 different tools to go into to figure out what my use case is gonna be. And quite often I fluidly wanna start in one and then maybe end up in the other.

And then my other question is, because this is pulling from emails and Slack, I'm presuming this app is reusing the connectors from Perplexity Computer, or are you setting those authentication tokens separately.

Correct. So all the connectors here are the default auth for what Computer's using in that app as well. And so that's really, really cool. Exactly. It's so, so smart because even if I'm coding with connectors that I have in Codex and Claude Code, it has to set up auth every single time. It's doing it all from scratch every single time. And the other thing that's really cool about Computer is it's intelligent enough that if the auth isn't working, it'll warn you, it'll try and go reauthenticate. And it'll even just try and do it in browser because almost everything exists on the web today. So they can still build you the proof of concept.

That is good stuff. This is good stuff. Okay. Um, just to recap for folks, I'm like staring in awe, you know, as, as an AI person, you get like a Clairvoyant 2-hand experience. This means—

I'll take it. Appreciate it.

So good.

Yeah.

Love it. So just to recap, you were just gonna build yourself custom apps for processing your work. Using Perplexity Computer, bunch of out-of-the-box connectors that can be both used to just like natively query your information and give you answers as well as be deployed as the backend for apps, personal productivity apps that are optimized for your workflow. And then is there one more you want to quickly show?

Yeah, you actually just reminded me as you were talking about that, that I showcase a lot of like information consolidation streams because that's what's most on my mind right now. But I've also sent Computer to a couple of other people on my team that I know are good tinkerers. And I've just been like, use it, ask me questions, let me know how it goes. One of my favorite uses of this that I would never even have anticipated on my own is that we have Clay University, right? So just to give you an example so that anyone watching can also see this, and for anyone not watching, I'll describe it to you. We have a lot of content on a website about how to learn Clay, and it's well architected. Our design team did a great job. Shout out to all of them on helping me with this. That's so beautiful.

Yeah.

I did make it in the past. No one liked it, right? But there's a lot of information here and it's not really persona-based or built. And as we as a company have scaled and are now scaling into further segmentations of different industries and audiences and who we sell to based on the features that we've built, it makes more sense to say, if you're in RevOps coming to the Clay website for university, where do you start? Because not all things similar to my Slack notification issue are created equal. for you based on the profession you're coming into on the website. That's a major design system overhaul to take on top of this website. And it takes a lot of thinking. We, none of us on the education team are designers. I went into Figma Make, no, no shade to Figma Make, to try and rebuild it. But the problem is that I had to like redescribe everything to Figma Make in order to even get it to look accurate to what we already had for designs. on the university. So it's not able to ingest the sort of like visual context layer that I have. What was really cool about what my teammate, shout out Chris Ming, did here is he went in and said, oh, Perplexity Computer can access it in the browser. It has all the different models, so it should also be able to visually recognize and then understand what I'm prompting back and forth. So he took an hour just chatting back and forth with Computer. It doesn't look as pretty for those not watching, but it's functionally Much closer to what we're trying to envision. Okay, now if I'm logged out, I can see all these different persona-based journeys. I've got SDR, BDR, RevOps, MarketingOps, GTM Engineer, and it even went ahead and built in this top right corner, what does it look like once you're logged in? So now Sarah Chen, random name, not a real person. How many courses have you completed? What are the next courses for you based on— you can see your persona on the top right. Being in RevOps. Let's look at what workshops you should go to. Let's look at what the cohorts are. And all of this helps our design team then better quickly understand what we're looking for. Because the other thing that I get frustrated with all the time is the like gap or the chasm in communication between design and any other stakeholders, because they know what they need to do, but they don't have all the context that we have. And so being able to build a visual bridge between those two is incredibly valuable.

I love it. Okay, so you're using it, you're getting your team to use it. It's not just for personal productivity. You can use it to prototype, pull in existing sites, really understand the context of them, and then build something that you can use to communicate to your cross-functional partners to get work just done better and faster.

Exactly.

And I'm going to give you a compliment. I get— I see a lot of B2B websites, Clay. Beautiful. Top, top, top.

I will definitely pass it on to the team.

Oh, definitely. It's gorgeous. If you have not seen it, go, go check it out. And if you have watched my, um, I think it was like my Opus 4.6 versus GPT-5.3 design head-to-head episode, if you watch closely, I redesigned the ChatPRD website and I say, hey, I love the Clay website, use that as inspiration. So excellent work there. Okay, so you are a hyper optimizer. Any other use cases you want to show for us before Before we get to our lightning round?

No, let's just go for it. Let's do the lightning round.

So we have seen a lot of personal productivity work here, but you said that like 70— I'm— this is how I remember it, so, you know, we're gonna— it's, it's God's truth— like 70% of what people are asking you to do is go get hot pot and like hang out on the weekend. You're just a very fun, fun person. So are there any fun— you know, now that you've dug yourself out of Slack and email and your digest and All your work's being prototyped by your team. What are your fun use cases of AI?

I think my, my preface to this is that I don't think I'm as fun as everyone else is with AI for whatever reason. Like, I treat AI truly as like a work tool. I have friends who are like in— they call ChatGPT Chat, you know, they're having like a personal conversation, they're sending text messages back and forth and then sending screenshots of Chat's response to the text messages to me being like, Look at how good it is. And for some reason, that's where I draw the line of like, I don't need a chat therapist. And also half of the time I'm pointing out to my friends that I would disagree with what chat is saying. I think it's just there to support you. And I'm here to tell you that that's not the right support. So that's one thing I'll put aside. The other thing that I'll say I probably use it for the most in terms of like personal fun in my life is brainstorming and research. So I love games. I love board games. I love activities. I love sports. Me and my friends host like a winter and a summer Olympics every year. We're now doing a spring and a fall, and it's typically a medley of a bunch of random activities. We'll do like apple bobbing, who can pull out all the napkins out of a box the fastest. Trivia is really fun. And we'll do like one of my favorite that I stole from a friend was a list of 10 things. You just have to guess. Was it a sword, a fish, or soup? And shockingly, no one scored above a 40% on that round. So we love doing these types of activities that if I had infinite time, I would just spend all of my brainpower thinking about how to make this more and more fun. But most of the time I have an inkling of the things that I want to do. So for example, this last Winter Olympics that we held with all of my friends, I knew we wanted to have like 2 or 3 ironic throwbacks to college in terms of drinking games. We wanted to have 2 or 3 more fun conversational games and maybe 2 or 3 actual games. So I had a long brainstorming session with Jarvis about what are all the activities that we've done before? What are the themes on them? How should we actually think about ramping new activities? And almost never is it actually exactly what I want, but it gets me in the thought process of, oh, that is pretty much close to what I want. Let me make the final modifications myself. And then it's also really good for Like the ops actually of organizing all that, we had 20 people come. So now then we wanted to be really intentional about matching people in teams. I put everyone in pairs, but I wanted you to interact with more than just your pair throughout the day. So then for each of the games that were 4v4, we were rotating which pairs were with other pairs to come form mega pairs to then compete in a game so that you got exposure to like everyone else throughout the day. And so that I thought was really, really fun. And I use that all the time. I actually bought 10 new board games because I found them via Claude yesterday and I was like, these look fun.

So I think that's so fun. And you are not the only, um, board game gaming person that we've had on the podcast. We actually had an entire episode about how 2, 2 friends started a, um, a board game cafe.

Ooh.

In East Bay using ChatGPT and, and all sorts of stuff. So, uh, nerds be using AI to play. To play board games.

This is what happens when you give nerds AI.

Yeah, I mean, one of the things that I do like though is to be like, to throw big social events. I love the idea of putting more structure around like how people meet each other and what they do and how to make it fun. I also love a group activity. I recently ran a month-long March Sadness 64-song emo bracket where we decided what was the saddest emo song over all of— and I used like, I had like this vibe-coded app. It was like a nightmare to run before. It was so fun. fun to run. We had like over 100 people in it. And so I do think it is like a fun, um, kind of like jumping off point. It's never where you wanna end up.

Never.

But it gives you enough ideas that you'd be like, oh, I can pluck a little from that and a little from this. And then, and then get your friends together. So I, I love that. Okay. My, my last question, which I ask everybody, you seem like such a positive person. You seem like very proactive, very capable with AI. But, you know, you use OpenClaw occasionally, like, it's real dumb. Um, what, what do you do when AI is not giving you what you want? What is your prompting strategy? How do you right the ship?

I'll give you a very nuanced answer. I think there's 3 things, right? So thing number one is with OpenClaw in particular, it's even just recognizing that some skills maybe shouldn't be MCP skills. Case in point, My calendar. OpenClaw is really bad at dates and I don't know why.

Oh, so bad.

It cannot tell what today is. It cannot tell that I'm like talking in 2026 and I therefore am planning a trip for 2026. Why would I plan a trip for 2025? That makes no sense. But where like I can build in a cron job that basically then says, anytime I send a date message, just send the timestamp as well and recognize that this is the moment in time at which you are answering this question. And my hypothesis, it has something to do with like how the models are trained or when they were developed. So that helps. But like method number 1 is recognize that maybe sometimes the thing you are asking the AI to do is not the thing the AI should do, or that you need to give it better context. Method number 2, which gets much more silly, is be strict with it. Like I type in all caps, I will tell it like horrible things will happen. I'll be like, I'm going to lose my job. I'm going to have to fire my team if you don't do this correctly. My brother might not be able to make it back home. And all of these horrible— the more extreme you get with the examples, the more it's like, okay, on this shot, I promise I will get it. Even if you supply no reason, it has to be nowhere close to connected for you to actually get it to do better. But I will really tell it there are negative repercussions. And it does get stricter sometimes. And then I think the last thing I'll mention, which is really good for, I think in particular OpenCLAW and Claude, is I try to build skills for the things that it's not really repeatedly good at doing. So for example, I have a Google Calendar skill that I'm constantly refining and iterating. And so the more nuanced thing that I do is whenever it repeatedly gets something wrong, I ask it to explain to me why and how it arrived at that conclusion. And then I ask it to kind of go through the skill and tell me, what do you think is missing from the skill that would maybe make it better for the next time around if I just give you the correct answer? And iteratively, I have noticed it does get better and better over time. The warning is it's not a one-shot thing. It'll take 10, 12, 20 messages, but you will notice the improvement gradually.

I love it. And honestly, I've asked this question probably now 60-something times and no one has admitted that they are like, threaten the model with consequences.

I'm all, yeah, I threaten the model all the time. All the time. It's so unusually effective. The like data reason for it, I think, is the reward specification and the parameter and everything else. But like, it works, you know?

I mean, I always, I always reference parenting when we, when we come up with this question and I tell my kids, and I will tell the AI, I wouldn't yell at you if it wasn't the only thing that worked. We got it. Okay, this was so good. Where can we find you and how can we be helpful?

Great questions. You can find me on LinkedIn or Twitter. I think it's just my name, Yash Thakral, or Yash Tech on Twitter or X, whatever people call it now. Uh, and honestly, the, the cop-out answer I'll give you to how to be helpful is let me know how I can be helpful to you. I love teaching people things. I run education at Clay. I love this AI stuff clearly. And so the only thing on my mind is how to help more people have more fun with these things.

What a wonderful way to end. Thank you. And I'll talk to you soon.

Talk soon. See you, Claire.

Bye, y'all. Thanks so much for watching. If you enjoyed this show, please like and subscribe here on YouTube, or even better, leave us a comment with your thoughts. You can also find this podcast on Apple Podcasts, Spotify, or your favorite podcast Please consider leaving us a rating and review, which will help others find the show. You can see all our episodes and learn more about the show at howiaipod.com. See you next time.