How Figma’s PMs actually use AI every day | Ezra Mechaber (PM, Figma)

Ezra Mechaber is a growth product manager at Figma whose goal is to be “lazy in the right places.” When he joined Figma, he found almost no one outside engineering using AI to get real work done, so he and a couple of fellow PMs turned his personal Claude Code setup into PM OS, a shared repo of 31 skills that every PM at Figma now uses. In this episode, Ezra walks me through his favorite skills, how he bug bashes with Codex, AutoXP—an automated growth PM that found a huge win on its very first experiment—and a real-time writing assistant he built with Jev. What you’ll learn: Why PM OS is a repo, not a productThe skill Ezra runs six times a day to tame SlackHow one repo keeps 60 to 70 PMs in syncWhy his bug-bash agent catches things he never wouldHow AutoXP runs experiments with one human checkpointWhy Jev took over every meeting he was in for two weeksWhat to do when you want to yell at your AI— Brought to you by: Optimizely https://www.optimizely.com/howIAI—Your AI agent orchestration platform for marketing and digital teams Jira Product Discovery https://atlassian.com/howiai—Prioritize with insights, build with confidence — In this episode, we cover: (00:00) Episode preview (00:51) Introduction to Ezra Mechaber (02:04) The origin story of Figma’s PM OS (06:30) PM OS demo: the “chief of staff” setup (10:28) How the repo works: skills, hooks, and private sub-repos (12:00) How the skills got built, roasted, and whittled down (14:15) Build artifact, FigJam MCP, and coaching skills (17:33) Transitioning from the terminal to a local web front end (19:30) Bug bashing with Codex (22:08) AutoXP: an automated growth PM (26:26) Claire’s Experiment Engine backstory (26:58) Building a real-time writing assistant with Jev (33:15) Lightning round and final thoughts — Tools referenced: • Figma: https://www.figma.com/ • FigJam: https://www.figma.com/figjam/ • Figma MCP server: https://developers.figma.com/docs/figma-mcp-server/ • Claude Code: https://www.claude.com/product/claude-code • Codex: https://openai.com/codex/ • Cursor Automations: https://cursor.com/docs/cloud-agent/automations • Jev: https://typesafe.ai/ • Marker: https://marker.page/ • Instinct: https://instinct.com/ • Dots: https://openai.com/index/introducing-dots/ — Other reference: • Jev for beginners: how to use it and what to build: https://www.lennysnewsletter.com/p/jev-for-beginners-how-to-use-it-and — Where to find Ezra Mechaber: X: https://x.com/ezramechaber LinkedIn: https://www.linkedin.com/in/ezra-mechaber Website: https://ezramechaber.com/ — Where to find Claire Vo: ChatPRD: https://www.chatprd.ai/ Website: https://clairevo.com/ LinkedIn: https://www.linkedin.com/in/clairevo/ X: https://x.com/clairevo — Production and marketing by https://penname.co/. For inquiries about sponsoring the podcast, email [email protected].

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  I, as a growth PM, I feel like the biggest sin you can commit is wasting traffic. What if we could just automate myself out of a job and like create a little growth PM who could just like run and run and run? I call it AutoXP. And right now, one of our engineers is running it and has like five different experiments going that I have not taken a look at. And it found a huge win on its very first experiment. AutoXP has its own channel where it can communicate with itself. Hey, AutoXP, go find an experiment to run on our settings page. See what you can do. Go crazy. And it will trigger a data scientist to go find drop off and like look at the funnel and look at mix shift and all of those things.

• 0:37Claire Vo?

  The whole hypothesis was like, how do you exploit all this traffic and maximize the number of experiments? And then like the time to be alive, it is here. You can just like five code it there. This is amazing.

• 0:51Claire Vo?

  Welcome back to How I AI. I'm Clara Vell, product leader and AI obsessive here on a mission to help you build better with these new tools. Today I have Ezra Mechaber, product manager at Figma, and he's going to show us Figma's internal PMOS, the 31 skills that every product manager at Figma uses to make their life a little bit easier. We're also going to talk about, of course, my boy Jev, and how you can build a real time AI editor with that little, cheap, fast model. Let's get to it. Optimizely Agent Platform for marketers is built on a bigger idea, a whole directory of agents now joined by a newly revealed squad of virtual teammates, each with a defined

• 1:33Claire Vo?

  role and a personality of their own. This isn't about faster drafts, it's about handling the coordination, too. The research, the approvals, the handoffs, the routing, the back and forth that quietly eats your week and never makes it into anyone's highlight reel. Get the story behind it all. Hear the confessions from marketers who have handed off the busy work, meet the squad, and understand what virtual teammates can do for you at optimizely.com/HowIAI. Ezra, welcome to How I AI.

• 2:08Speaker 00

  Thanks so much.

• 2:09Claire Vo?

  One of the things I love about this podcast in particular, because I know we have a lot of crossover listeners to Lenny's podcast, is PMs are really leaning into AI, but they're all doing it in completely different ways, right? I started like a jillion AI years ago with like Chat PRD and like, how do we make documentation and go really fast? We've seen so many different product flows. I think everybody's sort of crafting their like artisanal product platform now because they can. And so when you said you had a PMOS, I was just so excited to have you come on and demo it. So tell me, what was the origin story of what we're going to see today? And why did you build it?

• 2:50Speaker 00

  I like to tell my engineers that I, my goal is to be lazy in the right places. And so for me, a lot of it was, I walked into Figma a couple of years ago, and I thought everyone was going to be using AI. It was going to be crazy AI-pilled. I was just gonna, I was already feeling behind. Like, I just, I felt behind because we're all on Twitter and we all see like, you know, Tebow just posting about like the future of the universe. And I walked in feeling really behind, and I realized basically no one was using these tools outside of engineering to like get real stuff done. And so I started just trying to automate some of my own productivity. Like, I was the person who was trying to do

• 3:32Speaker 00

  too much in Zapier like many years ago. And the more time I spent with it, one, Claude Code had just come out, and I started to realize like there were things I could do in the terminal that were like really interesting. All of a sudden, I could like build CLI. We, you know, we were in that CLI phase. And like, no other PMs who I was working with were doing it, and they were all sort of curious. And we had this really interesting challenge from Yuki, who was our CPO at the time, which was basically like, what would it take to get all of our PMs kind of AI-pilled? Like, what would that look like? And working with a couple of other PMs, shout out to Jinan and Lawrence, started thinking about

• 4:15Speaker 00

  like, what would it look like to turn this thing that I was doing, just like hunched over my laptop, like a little creep, into this kind of like reusable thing that lots of other PMs could, could delve on top of?

• 4:26Claire Vo?

  Great. And so what, what were like the problems that you thought were most important for PMs to think they could solve with AI?

• 4:35Speaker 00

  First of all, I feel like the, the like really trite thing about being a PM is that you're doing too many jobs. And the question is like, which jobs do you either do really poorly or not do at all? And so for me, it was like, okay, how do I do, for the jobs I'm not doing, and the jobs we are not doing as a team, how do we do them at like, the minimum allowable level of like missing things? Like, my IR support team is incredible, and the thing that I forget to do on every launch is to loop them in at the right time. And so one of my questions was like, how could I do this to make sure that I don't forget Katie, one of our support ops people? Or, meeting notes? Not my friend. I'm really bad at them. But like, we have transcription tools. How do I just get like

• 5:19Speaker 00

  good enough so that I'm not blocking other people? So a lot of them started from my jobs of what I wanted to do. Find someone in the org who works on such and such thing. We're like 2,000 people now. So just understanding who the code owner is is really useful. And understand who we are in relationship to each other. And then also just like, how do I get fun productivity out of all of our other PMs? So one of the skills that's in there was this like really detailed spec skill that started to walk through, how does our codebase work if you wanted to do this really intensely? Like, if you wanted to do this the right way, what would that look like?

• 5:57Claire Vo?

  Amazing. And so I love this idea that you had a couple tasks on the PM plate that you were like, we need minimum viable doing of these tasks. Like, it just needs to-

• 6:10Speaker 00

  Just like making not bad at it.

• 6:11Claire Vo?

  Yeah. And I talk a lot about, you have to like both raise the floor, like, we have to do just a little bit better on our baseline and take off the ceiling. And I think AI can help you do both. And often you have like one set of people taking off the ceiling and saying, hey, everybody else, like, take off the ceiling instead of like, this is how I can help you raise the floor. Okay. So show us what PMOS is, its skills, what it does. How are you relying on this day to day?

• 6:39Speaker 00

  Cool. Okay. So this is the rough structure. It started just as like a really simple repo inside of Claude Code as a TUI. So it's got this, you're, you're the PMs chief of staff, you know their work. I hate the term chief staff, but like, that's, it's really useful for like teaching.

• 6:59Claire Vo?

  It's good shorthand.

• 7:00Speaker 00

  It's really good shorthand, which makes me die a little bit every time I read this. When the session kicks off, it reads some information about like who the, who the PM is, who they work with. It's got a memory file. And then it's got a couple of different skills that we can walk through. So one of them was just like, we had no good data on like which set of people work on X thing. Like, we have several different org chart tools, and like, none of them are stitched together in useful ways. And so one of them is just like, literally who works on this team, who owns X, and it calls it when it writes docs. So if you need review on a doc, you, the PM, do not need to remember

• 7:42Speaker 00

  to say, my, my engineering manager is Amen. It just adds all of the relevant people as, like, Charlotte is my designer and Amen is my engineering manager. And it adds them as the relevant people and remembers who your skips are. So that's skill number one, which was just like a nice way of reducing some of the like repetitive work inside of new docs. I use it a lot as like a meta Slack. So our, our Slack is insane. I feel like anyone who's in a large enough org in tech, just like Slack becomes really unruly. I archive roughly 10 channels a day. It's bananas. And so one of the things that we were really

• 8:21Speaker 00

  focused on was how do you help PMs know where to focus and what to do and like go run down all the things that are important? And so I run this skill, what's up, like six times a day. It pulls all of my relevant Slacks and projects together. It pulls all of the things that I like, who I owe stuff to, what decisions are outstanding. I have a version of this that runs through all of my live experiments on the platform and like gives me a, not that I peek at my, you know, I had to stop listening.

• 8:53Claire Vo?

  Wait, hold on.

• 8:54Speaker 00

  Not that I peek at my experiments.

• 8:55Claire Vo?

  Say it. Say directional. Say directional. I dare you.

• 8:58Speaker 00

  Right. Right. And so I, I find this thing to be incredibly useful. And let's see if we can find anything. So it will look at like certain channels that we've identified as useful. It will read the whole thread. This is probably a little overspecified. Like, I feel like we're all moving towards like the models are just good enough now. But there's a lot of stuff in here about how to find exactly the right stuff for the way that we are all set up in Figma, which is just enormously helpful for how I run my day to day as someone who's just in way too many meetings and trying to do too many projects.

• 9:36Claire Vo?

  What I like about what you're showing, because so many of these like PMOSs that we've seen, skip all this like foundational piece, which is, in large organization, it is actually really hard to know, like, who should have an opinion on this product? Is somebody else working on it? And so I like that you invested in sort of like the baseline context gathering as skills. And then I'm presuming that helps like level up some of the more like complex or strategic workflows that come out of PMOS.

• 10:06Speaker 00

  Yeah, it's been a long ti- I like tried to build a benchmark sort of unsuccessfully. And I'm curious how you think about it for all the stuff you're doing now, but like- I would be really curious how this performs relative to like totally vanilla Codex these days. But I do think like in general, especially in that early, like, I don't know, like last October through like May, as like the models were getting better, but were not like superpowered yet. Felt like it was really, really helpful for how it navigated all the weird parts of like Figma's IA.

• 10:38Claire Vo?

  Yeah. And one thing, again, we actually didn't specify this for folks, but Product Team OS is a repo, right? So like, take a step back and tell us, like, what is it? Is it a repo of skills?

• 10:53Speaker 00

  Yeah.

• 10:53Claire Vo?

  Is there more to it? Like, what is actually the thing?

• 10:56Speaker 00

  Yeah, so it's a really, really basic repo. And so it's got an agents.md, which just tells it how to behave. It's got a bunch of related skills that tell it when to fire and what to do. It's got a couple of command hooks. So things like on boot, go find out if there's a new version of PMOS, because we wanted a good way to like keep our 60 to 70 PM kind of roughly in sync. We've also got a couple of built-in folder structures. So the base template when you onboard is this like, it's a dot folder, so it's invisible and you can't mess with the context. But then it duplicates and it sort of builds your version out.

• 11:41Speaker 00

  And then it's technically two Git repos. It's the parent repo that everyone can download. And then their workspace is a sub-repo, so that if you want to write stuff back, you can write it back privately.

• 11:52Claire Vo?

  Amazing.

• 11:52Speaker 00

  Again, because like none of this was in a UI at the time. And so we were all just trying to figure out ways to like make GitHub work for PMs who had never touched GitHub before.

• 12:00Claire Vo?

  And was there a small group that sort of like brainstormed what these initial skills should be? Was it like, hey, everybody just put it in your favorites? How did that work?

• 12:10Speaker 00

  Yeah, so Jen and Lawrence, I think, talked about this a lot at Config this year because they did a little presentation on this. But basically, we had a really small Slack channel of just the three of us hacking away and firing off as many PMs as we could for a little bit. We had that sort of like Cambrian explosion of like 1,000 skills and then started to like whittle stuff down. And then we did an offsite with every PM at Figma this spring, where we had everyone install it, use it, hack on it, come up with like what was good, what was bad, just like roast it. And so some of the skills that are in here came out of that. Some of the skills either got whittled down or whittled away. And I would say as the models have gotten better, I've been trying to basically take stuff out that's no longer useful.

• 12:53Claire Vo?

  That's, that's great. And so I have a question. Do you have, as a team, any telemetry on this, like how it's being used? Or is it instrumented?

• 13:03Speaker 00

  No, I thought about doing this and I decided, one, I like didn't want anyone to feel creeped. And like I wanted it to be fun. And I also just didn't have the time for it because all of this was, you know, like sort of 20% time. The biggest thing for us was just we, like, PMs at Plac are chronically online. And so we got a lot of feedback. And then we had a public channel. And so we started getting like, all of the engineering managers started using it and asking us if we could like, come help them talk about how to do knowledge work for engineering. Our data scientists have one that is like essentially a fork of this. And so we were getting a lot of feedback about what it was good at and what it wasn't good at.

• 13:46Speaker 00

  And I think for us, the, the big next thing, and we can talk about the evolution into what it is today, if that's helpful. But like, mostly for us, it was, we managed to get a bunch of PMs comfortable in the terminal. But like, surely that's like not the right end state. Like, especially if you're doing a lot of writing, it like sucks to read monospace font all the time. And so for us, like that was where we wanted to spend our time is like, okay, what, what would be next? How do we get people out of the terminal now that we've got them comfortable to concept?

• 14:15Claire Vo?

  Well, before we go to what's next, can you walk us through like just a couple more of your favorite skills in here? Maybe ones that teams haven't thought of or you think are particularly unique or kind of helpful?

• 14:26Speaker 00

  Yeah, sure. So I do think for us, like with a busy Slack, the, this kind of like meta what's up that is tuned to just like literally what do you have to do every day? And how do you judge whether you're doing a good job is really useful. Build Artifact is one of my favorites at the moment, which is if I have something like, actually this,

• 14:47Speaker 00

  this demo like dashboard that I made for this podcast is just like a really simple HTML artifact. And then we have the ability to publish it. We've got a little hosting tool that we use internally. And so it takes it through, it's got, in this case, it has a template HTML. And I've got a couple, I don't know where they're sitting in. I've like totally lost track of my folder structure at this point. But there's like a couple of different templates of little CSS that I like. And I've sort of grabbed from different places. And so build stuff like this when I want to show stuff off. Figjam, the MCP for Figma is insane. I don't know if you get to use it on a regular basis, but I feel like I use the MCP all the time.

• 15:29Speaker 00

  And one of our PMs built this amazing Figjam builder that produces these like wild workshop templates very quickly. And so just figuring out like how to load it up with, like, what do you think looks good? Or how do you think this should work? And so there's a lot of taste in here on how it's designed. There's, you can also tell there's like, we've got a lot of Claude slop in here from earlier.

• 15:55Claire Vo?

  I was gonna say, that is such a Claude slop sentence right there.

• 16:00Speaker 00

  But it's a mix. Like, there's a lot of stuff in here that like, if you look at this output compared to some of the Figjams that we would put together natively internally, like it's, it's just a really good set over time. And then maybe my favorite other one is coaching, which I use. As someone who used to manage people and is now ICing again, like I spend a lot of time.

• 16:24Claire Vo?

  Living the life.

• 16:25Speaker 00

  Oh, it's so nice. It's so good. I used to spend a lot of time thinking about how to coach people. And, you know, I feel like managers are very busy, especially now. And so thinking about like, I always wanted to make sure I was doing the first pass of my own coaching before taking it elsewhere. And it's just really good at saying like, hey, here's our career ladder. Here's where you are. Here's all the feedback that I have about you. I'm gonna listen to your problem, or I'm gonna listen to this thing that you're working through. And I'm gonna tell you basically like, get over yourself and don't be like, don't worry about it. And just like, keep working on this thing. Or like, learn how to communicate with this person.

• 17:07Speaker 00

  Because I remember that the last time you had an experience with this person, like you misunderstood them. And when you had a conversation about it, you understood that you had just misunderstood. So I really like it for coaching. I think it's really, really strong on that one.

• 17:19Claire Vo?

  And you have about 30, 31. And like just looking through the list, they seem like both a combination of very tactical, interpersonal, and then like generative in some ways. Like that seems to kind of be the, the, the catalog of them. Okay, so how do people actually use this? Is this a kind of like app that you built? Like, how are people using this?

• 17:43Speaker 00

  Yeah, so I think it's been a little bit of a journey. So we started in Claude with the like literal TUI in terminal. We had like a lot of debates about like the best terminal app because people were coming over from UIs. Then we built a little front end for it, which runs locally. And we tried to make it really fun. It's got like all of the like lead code type ASCII art in it. And the goal here was just bring it into the web browser so that people could type, like, we're doing a lot of writing. It's hard to read monospace. It's got some nice markdown formatting. We built it, it works on mobile. I actually like, right after, this was, this was pre-Astra,

• 18:25Speaker 00

  but like I spent a bunch of tokens building like a local native version just to like see what it looked like. Now I'm basically in Codex 100% of the time. One, I think browser use has gotten really good.

• 18:37Claire Vo?

  So good.

• 18:39Speaker 00

  Really good. We have, we don't have permanent desks in the New York City office. And I don't like thinking about booking a desk every day. And so every once in a while, I have it running on a schedule to just like book my desk one day out. But it's also, it's a total beast at bug bashing. And so like in Codex, I have a little bug bash. I have like a threaded conversation that just lives here all the time. And when something new comes up for our team, where we're about to launch an experiment or redesign or something, we have these Figjam files that explain, here's, here's how to bug bash it. Like, here's the test account to take. Here's the path to follow. Here are the actions to do. And like, here are the things to look for.

• 19:23Speaker 00

  And I have started basically snagging one account for me and one account for Astra and just having it run through. So in this case, we are trying to make some of our request modals inside of Figma just like much more straightforward and much more streamlined. It's all UX stuff. It's all UX writing. It's fairly straightforward, but the like logic behind it is really complex. And so Astra is just much, much more diligent about following directions than I am. And so I will run it through and I will say, like, keep doing all of these specifically. And just like, it will run through all of these scenarios and it will use the browser. And it's a dream.

• 20:05Speaker 00

  And then when it's done, it writes out back to the Figjam board of all of the things that it has done. And I feel like we are still tuning it so that it is not like annoying in the comments that it drops. But in terms of like what it finds for me, like it, it found for, two things it found in the last month. One, it discovered that if you put 1,000 emojis in a text box, it would break, which like weird edge case, great, good to harden. Two, it, it found that the way we had you request a seat for a given team inside of Figma. And when you do it, it shows your team name. And we had A, like A quote, team name, seat.

• 20:49Speaker 00

  And if your team name starts with A, you need to use and.

• 20:53Claire Vo?

  And.

• 20:55Speaker 00

  It had just not shown up in our earlier bug bashes because it really depends on what the team name is. And so it was just, it wound up being so diligent that it caught really small nits like that, that I would never have caught on my own.

• 21:05Claire Vo?

  This episode is brought to you by Jira Product Discovery. With AI, a single PM can summarize research, draft a PRD, and build a working prototype. But multiplayer mode? That's broken. You can't get everyone aligned on what should actually ship with just a markdown file. Jira Product Discovery gives your team one place to capture insights, prioritize with evidence, and build living roadmaps everyone can rally around. And because it's powered by Atlassian's teamwork graph, it has the context you need to decide what to build. Those decisions flow straight into Jira, so your developers or coding agents can start

• 21:45Claire Vo?

  building with the context behind every decision. More than 25,000 companies already use Jira Product Discovery. Try it free at atlassian.com/ how I AI. That's atlassian.com/ how I AI. Jira Product Discovery, better product decisions in the AI era. I, I love that so much. And so I think this like bug bashing workflow is a really great one. I highly recommend people do this. And then keeping track of it in a fig jam or some other sort of like kind of asset is helpful for you to just track what the feedback is. You got another automation that you're using in the team around experiments, right?

• 22:28Speaker 00

  Yeah. So one of the things that I had been thinking about for a while as our models have gotten better is like, what are all of the different ways that we can improve cadence? And so one of like, I, as a growth PM, I feel like the biggest sin you can commit is wasting traffic, right? To like have unallocated traffic that is not running towards an experiment at any given time. It's like, why are you wasting that traffic? And so thinking about ways that we could run an experiment on a surface we would never have spent time doing normally was really, really interesting to me. And then the models were just getting good enough across a little bit of engineering, a little bit of data science, the user, like

• 23:10Speaker 00

  the user experience writing was not so bad. The ability to come up with like, within a really bounded sandbox was getting better. And so I started thinking about, like, what if we could just automate myself out of a job and like create a little growth PM who could just like run and run and run? And there's, there's been a couple of other folks who have done this. I think some of the folks at both labs have done it, which is where a bunch of this inspiration came from. But I realized that internally, we finally had all the tools. And so I call it Auto XP. And right now, one of our engineers is running it and has like five different experiments going that I have not taken a look at and are just, and it found a huge win on its very first experiment. And basically, the

• 23:53Speaker 00

  way that it works is this, you can see, this is like one of the fancy schmancy artifacts that we made for this, is it works as a cursor automation. So there's, it, there are three cursor automations that trigger based on stuff in Slack. Auto XP has its own channel where it can communicate with itself. Someone can either tee up an experiment in the Slack channel, or they can, or it can just like run on a loop. So one of the things you can do is just say, hey, Auto XP, go find an experiment to run on our settings page. See what you can do. Go crazy. And it will trigger a data scientist to go find, or data scientist

• 24:37Speaker 00

  agent to go find, to go find drop off and like, look at the funnel and look at mix shift and all those things. It will trigger a UX writing agent that will come up with some ideas for copy in a given place. We're like moving towards more front end design now. And then lastly, it'll trigger an engineer that'll pick up a PR and run it. And then again, with all of these steps, it writes back to the Slack channel in the same thread. So you can follow over the course of the experiment, basically like what's happening. And there are two things that I think it does here that are really amazing. One of which is we are able to run experiments in places we would never have bothered to dedicate

• 25:19Speaker 00

  time or traffic. It's way faster to get from, I have an idea to this experiment is launched. And then two, when it's done, when, and an experiment finishes, it triggers essentially a webhook back to the original experiment to say, hey, it's time to analyze these results. And it has a bunch of instructions from all of the time that I have spent as a growth PM to figure out, like, do we want to explore more? Do we want to try to exploit? Is this a win? Is this a loss? And the goal for all of this is the only time you're allowed to look is before you launch the experiment. So you review the data science findings, the hypotheses, the UX writing and the code all at once. So like one human touch. And then

• 26:02Speaker 00

  it watches and then you can review whether or not it's a win or a loss, but it has already started data writing on the next rounds. I would say like, where we are in theory is pretty, like, is great. Where we are in practice is almost there. And so one of our engineers, Marian, has been taking this really, really far. But we're now at like, five live concurrent experiments on things that we would not have been doing otherwise.

• 26:26Claire Vo?

  So this is a little bit of Clairvaux lore, but for those that are, are new to me, I have worked at every experimentation platform company in history and did a startup literally 13 years ago, something like that, called Experiment Engine. And the whole hypothesis was like, how do you get, how do you exploit all this traffic and maximize the number of experiments? And then like, what a time to be alive. It is here. You can just like vibe code, vibe code it there. This is amazing. Okay. I want to show one last thing, which is my boy Jev. And we have been doing a lot of Jev episodes.

• 27:06Speaker 00

  Or one out for Jev.

• 27:08Claire Vo?

  Poor, I, we, I don't know. I think Jev's got a, got a real shot at some longevity here.

• 27:12Speaker 00

  You have Heist Juice?

• 27:14Claire Vo?

  You, but you've built something really cool with Jev that I want you to show. Cause it's a use case I haven't seen.

• 27:20Speaker 00

  Yeah. Yeah, totally. So, right. So for people who don't know, though, how could you not know? I feel like if you're listening to how I AI right now, Jev is this model where you can ask it really weird, complex questions in text and you get like essentially a number back, right? So like, how good is this on a scale of one to 10? Or is this true or false? Or like, what is the meaning of life? Is it A, B, or C? And it's really fast and it's really cheap. And so I started thinking about all of the different places we could shove this that would really change the way that people work, right? So if you've used, Figma has an agent on the design canvas. I adore it. It's really, really good, but it is still subject to all of the latency that you get with all of these top tier models,

• 28:05Speaker 00

  which is like, it's slow in places where you want to have a coworker and it doesn't feel like a coworker. And so as we work on that, Jev feels like really, really interesting for some of those possibilities. And one of the first things I was thinking about was this idea of like, what if it could run questions about things you were writing as you were doing it? Like, what if it could know that you needed to, if you've ever written a blog post and had a section where you aren't sure what's going to go there, maybe you write TK or you leave some brackets. And sure, you could do that, like totally programmatically of just like, look for and identify. But also maybe you have a claim that you're not

• 28:48Speaker 00

  sure, like you have data for, or you need to go look up in your notes later. And so I basically, I built this prototype. I'm going to forget the name of it, but there's, I think it's called Marker, which is this beautiful looking writing app from this guy, John Steinbeck, which is an incredible name for a guy who runs a writing company. They have some of these ideas around it. And so as I was just playing with like, what could be possible with Jev? That was kind of one of the inspirations here. And I wanted to have Jev basically make a call to Jev anytime I stopped typing. And so the thing that it does is when you stop typing, it will look for places where you've left blanks or claims

• 29:30Speaker 00

  worth supporting. And it will use stuff in your library to go ahead and actually like substantiate those claims. So in this case, like I said, most people find it easier to keep writing than to stop and look something up. And it's making a call here as soon as I stopped writing it into the library where it has all of these interviews about like what these people actually did. Like if I go hunting for a number, the, it will, I forget the sentence I was trying to write. We recently put the API behind our own internal firewall. So let me see if I can actually run it live or if it's going to get mad.

• 30:13Speaker 00

  Okay.

• 30:13Claire Vo?

  Oh.

• 30:13Speaker 00

  And then.

• 30:14Claire Vo?

  Hot dog.

• 30:14Speaker 00

  Oh, no, hold on. But it doesn't actually want to, is it blocking it? What is blocking it?

• 30:21Claire Vo?

  I don't know.

• 30:21Speaker 00

  It wants to run.

• 30:22Claire Vo?

  It impressed me. It did something fast in my.

• 30:25Speaker 00

  Right.

• 30:26Claire Vo?

  Okay, so I just want to like reiterate this for people because it is mind blowing, which is as you write, as soon as you stop for a minute, it runs a quick check, which is very, very fast and says like, does something need to be filled in? Does something need to be supported? Like, what questions do I have about what you're writing? And then like hooks up to data sources and actually helps you improve your writing as you. I, Jeff.

• 30:50Speaker 00

  Right. Okay, so this works. So you've got, here you've got your title. So this is my title. And as soon as I stop, it's going to start checking things. So maybe I've got, I was thinking about how we serve writers. And I thought TK. Great. And it's like, you had a thought that you didn't finish. But then you'll see like, it also ran a check and it said, I have no idea what you were thinking because it's not in this library. But then I could say something like, we spoke with ex-writers and they said, and it both knows that I need to finish this. And it went and pulled from the library, like literal interview quotes from like writers we allegedly spoke to for this.

• 31:34Claire Vo?

  A thousand out of ten.

• 31:36Speaker 00

  Blew my mind.

• 31:38Claire Vo?

  I will like this. You rarely get one of these at How I AI. A full Claire hands up. Jeff.

• 31:45Speaker 00

  It's so cool.

• 31:48Claire Vo?

  It's so good.

• 31:49Speaker 00

  I for literally for like two weeks after Jeff dropped, I was just like running around going like, okay, but like, can we talk about Jeff? Like, it was every meeting I was in. My engineers, some of them were very excited about this. Some were maybe like a little eye rolly about it. But it's just, it's such a cool idea because I think you can imagine all of the different ways that if you could run checks on things at any given time, like what you would do with it. And I, the thing that sort of blew my mind was that they had built it for RAG or for essentially for like search and document retrieval as well. And so the ability not just to say like, this looks like a pattern. I am confident that this, you left something. But also I took this block of text and it relates to this section over here.

• 32:34Speaker 00

  And I did it in real time is just like, is kind of incredible.

• 32:38Claire Vo?

  It's unbelievable. This, this podcast is now How I Jeff. Just.

• 32:43Speaker 00

  It did.

• 32:44Claire Vo?

  We're rebranding.

• 32:44Speaker 00

  I did have this really bleak thought, which is it made me wonder if, like you could imagine you're interviewing for a PM job and you like yap your entire career into something. And then you have a voice to text that is just like searching against like what your examples are. And I did have this fear because I run a lot of Product Sense interviews that I was going to start running into people who were like jabbing their own little helpers.

• 33:08Speaker 00

  Which, I don't know, maybe I should hire them if they do. That's great.

• 33:10Claire Vo?

  Hire them. I was like, great, sign me up.

• 33:16Claire Vo?

  Ezra, this has been so good. Okay, we're going to get to lightning round. I just, to reiterate for folks, what do we see? PM OS with 31 skills, including what's up? What do I have going on today? Find me people in the org, write specs, coach me, the whole thing. You basically are now using that through Codex entirely. Also because Codex is magical for browser use. And then he built a real time Jeff editor, author, writing tool

• 33:46Claire Vo?

  that you're afraid is going to cause you trouble in future PM interviews or impress you in future PM interviews. So many good workflows in this. So the folks that have stuck with us this long, you have gotten some good stuff. Quick lightning round questions and then I will get you out of here. One, you told me a couple of weeks ago, you were using Instinct. Are we still on Instinct? What are we doing with our personal agent?

• 34:11Speaker 00

  Okay, so I still get texts from Instinct and I don't respond. It's kind of the way it works. So for people who don't know, Instinct is this, it's really sort of beautiful idea. Like, I'm really into it, which is this idea that what if you had a proactive agent? But instead of like needing a website to visit or an app to use, it just like texts you in iMessage. And I will say I still use it every once in a while. It was this kind of magical thing at the beginning where it was proactive in the right ways. And this like concept of memory that was like useful in this kind of mind blowing way. And I suspect there's a bunch of people who have been running OpenClaw for months who are like, what's the big deal? But it just, it productized this really nice experience

• 34:55Speaker 00

  in a way that was really great. And for the first few weeks, you could buy things on Amazon with it, which while not like a killer use case, was really nice. Because I would just walk around the house and I would say, hey Siri, tell Instinct that I need to buy paper towels. Or like my current use case is I bike my kids around Brooklyn and I will run into potholes all the time. And now you can like, you can 311.

• 35:20Claire Vo?

  31.

• 35:21Speaker 00

  Like affix my pothole. And I never remember to do it on time because like you have to look at the address of where it is and you have to report it. And the website is kind of trash. And so I have started like, hey Siri, text Instinct, there's a pothole at like 269 Atlantic. And it just remembers like what it's supposed to go do and submits it. So I do use it for that and it's amazing for that. But I mostly switched over to Dots now that I'm like fully Codex build.

• 35:49Claire Vo?

  I, I, so I

• 35:53Claire Vo?

  am trying to make Dots.

• 35:55Speaker 00

  You're like, you're like rocking Muse, is that right?

• 35:57Claire Vo?

  I, so I'm like max, Codex maxing. I love Codex. You could rip Codex out of my cold, dead fingers. But for work, I'm all Grok.

• 36:07Speaker 00

  Yep.

• 36:07Claire Vo?

  And then for person, you should check out Muse. It's a real, it's a real delight.

• 36:12Speaker 00

  I can't, I can't bring myself to do it.

• 36:15Claire Vo?

  I know, but they've been stealing our data for two decades, you know?

• 36:19Speaker 00

  I know.

• 36:20Claire Vo?

  That's my, as I say to people, the devil I know. This is really where I'm at with my personal agents.

• 36:26Speaker 00

  They already have the data. So fair.

• 36:27Claire Vo?

  They already, I said, my joke is my Instagram ads will get better. I'll, I'll end up spending more money.

• 36:32Speaker 00

  This is, I think this is like a real value prop.

• 36:35Claire Vo?

  Okay, and then my last question for you. When your dot is

• 36:40Speaker 00

  Yeah.

• 36:40Claire Vo?

  being obstinate, your AI is not giving you what you want. Are you a yeller?

• 36:45Speaker 00

  No, no, I'm not. I think we should all be nice to our AI. I treat my AIs like a friend that you love who's really like messing it up. You know, you're like, buddy, like get your life together. Like, you're too old for it. That is how I treat. So, hold on, I will show you. Because I yelled at my dot twice today, but I, I did it very friendly. Okay, so I ordered some shoes. They were too big. I need a half size down. And I tried to like find it, the cheapest version. And it was like, email search is temporary, rate limited. Can you go search your email? And I was like, I don't know, man, that's what I needed you for.

• 37:28Speaker 00

  And then it gave me some like really expensive stuff. And I thought maybe it was being lazy. And so I just said like, yike, assume you aren't going to check this. And I, I was trying to figure out why I don't yell. And I think partially like you can ask Math, my wife. I am like mostly not a yeller. But I also think, someone posted this on Twitter recently. And I, it really expresses how I feel, which is not that I have some, for people who don't know, there's like this idea of like Basilisk's, like this idea that if you're mean to AI, but then it turns out they're alive, like, wouldn't it have just been easier to be nice? Like, isn't that cheaper than like getting turned into a paperclip? And I don't, I don't think they're conscious.

• 38:13Speaker 00

  I don't think they're approaching conscious. But I do think it's really bad for the soul to just be mean to things all the time. Like in the same way that you like, like, there's just no reason. Like, it's just, I feel like I get worse when I like yell at things that can't help it, even though I don't think it's conscious.

• 38:28Claire Vo?

  You know, I like used to be a very hardcore not yeller. And I'm not like, I'm not mean. I wouldn't, I'm not saying anything nasty. But I am a very, I'm like an all caps, like why? Like, why, why are you like this?

• 38:43Speaker 00

  Why would you do this to me?

• 38:44Claire Vo?

  And here's, and here's why I'll say this. I'm spending thousands and thousands, if not tens of thousands of dollars on LLMs.

• 38:52Speaker 00

  Yes.

• 38:52Claire Vo?

  Like, a why are you like this after like $2,000 of wasted tokens?

• 38:57Speaker 00

  Yeah.

• 38:58Claire Vo?

  That's like, that's like a cheap.

• 39:00Speaker 00

  But that doesn't strike me as that mean.

• 39:01Claire Vo?

  Yeah.

• 39:02Speaker 00

  You know, like I have, you know, there's like, there's some real like

• 39:05Claire Vo?

  Like nasty kind of stuff. Yeah. No, I'm just expressive.

• 39:11Speaker 00

  Yeah. I think that, I will say my take is that normally what I want to yell at AI, it's because I have given it poor instruction. This is how you know I have kids. Is like, there's so much of that, which is like, when you get a bad result, the odds that you have like poisoned to the context and asked a bad question is like pretty high. And so I will occasionally find myself in loops. The Codex has this, Codex, if you're listening, OpenAI team, if you're listening, has this really bad bug right now where if you start a local session and you kick it to a remote session, even if you have YOLO mode enabled on your remote session, it inherits the sandbox with ask permission every time.

• 39:55Speaker 00

  And so I wind up in like clicking approval hell. And I get so mad and I go around in loops for a little while where I'm like, we already fixed it. Like, you have permission. I'm looking at the file. I pulled up the file. We're looking at the config dot Tom altogether. Like, what are we doing? And usually the answer is like, hmm, the context is poison that conversation. It's unrecoverable and yelling at it is not going to help. So start a new, just move on with your life. Which is not satisfying, but it is how I feel.

• 40:21Claire Vo?

  Let me give you a TikTok parenting tip. If you want to like really devastate your AI by being a kind parent, which is, I know that you're capable, but you're not.

• 40:34Speaker 00

  Oh, yeah.

• 40:34Claire Vo?

  You're not, you're not showing me how capable you are right now. And that's just signaling to me, I need to be a better parent.

• 40:43Speaker 00

  My, okay, I love that.

• 40:48Speaker 00

  That, one, stealing the parenting tip. Two, stealing it for AI. I feel like I really love to give my AIs like big expectations. And I feel like it's not in the math proof, but like I have been doing a version of this, which is like, buddy, like I know you can cook. Like, you would not believe the things that I've seen out there that you are capable of. Like, if you try a little bit harder, I am sure that we can vibe code this prototype that I can show my engineers so that we can go there. Like, I know this is in your capabilities.

• 41:17Claire Vo?

  I know you can vibe code Salesforce, just a little bit more. Okay, Ezra, this has been so fun. So many good use cases, so many good thoughts. Where can we find you and how can we be helpful?

• 41:30Speaker 00

  Yeah, mostly on Twitter. I post too much. Ezra McCabre.

• 41:39Speaker 00

  And then in terms of being helpful, we're hiring PMs. We're hiring a lot of PMs and we're hiring a lot of different kind. If you're an engineer or a designer, we are hiring. Come work at Figma. It is the most fun I've ever had in a job. We're just, we're building really cool things and the people are great.

• 41:54Claire Vo?

  Amazing. Cannot, cannot end it better than that. Ezra, thanks for joining How I AI.

• 41:59Speaker 00

  Thanks, Claire. Bye.

• 42:01Claire Vo?

  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 app. 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.