you2idea@video:~$ watch 1em64iUFt3U [1:15:12]
// transcript — 841 segments
0:00 You are a product manager shipping product   without knowing how to write code,  barely knowing how to review code.
0:06 I have zero technical background, did  music in high school ... when Sonnet  
0:10 3.5 came out. I remember watching a YouTube  video building apps using Bolt or Lovable.  
0:15 It basically felt like someone came up to  me and said, "You have superpowers now."
0:19 These days, you're using Cursor with Claude Code. If you're non-technical like me, code  is terrifying, but AI just makes it so  
0:26 much possible. In the next coming years, I  think everyone's going to become a builder.  
0:29 Titles are going to collapse and  responsibilities are going to collapse.
0:33 The main challenge people have is  reviewing the code that AI has written.
0:36 It's very difficult for me to catch  mistakes. What I'll do is basically  
0:41 /review. This tells Claude to start reviewing  its own code, but what's even cooler is I have  
0:47 Codex as well as Cursor open. I will  have each of them review the code.
0:52 This comes back to this quote. I think  everyone's always hearing. It's not that  
0:54 you will be replaced by AI. You'll be replaced  by someone who's better at using AI than you.
0:59 It's the best time to be a junior, contrary  to what a lot of people are saying,  
1:02 how there's no more junior roles  out there. Yeah, that's true,  
1:05 but also when else in history could you get out  of school and just build a startup on your own?
1:11 Today, my guest is Zevi Arnovitz. Zevi is a  PM at Meta. Prior to that, he was a PM at Wix,  
1:17 and this is a truly remarkable conversation that  every non-technical product person needs to hear.  
1:23 Zevi is super young and has no technical  background, but as a smart, young,  
1:28 ambitious person, has learned how to use  Cursor and Claude Code to build significant  
1:32 and real products completely on his own, and he's  created his own very clever and effective workflow  
1:38 that everyone listening can copy. To make that copying even easier,  
1:42 at the top of the show notes of this episode, you  can download all of the prompts and /commands and  
1:46 start doing all of this yourself. Zevi shows you  how to work with cursor to quickly add your ideas  
1:51 to Linear to explore your idea with AI, how to  develop your plan, how to then build the thing,  
1:58 and then have different LLMs review  your code and update your documentation,  
2:02 and then use all of this as a learning opportunity  to develop your own sense of how things work. 
2:07 I haven't stopped thinking about  this conversation since we had it,  
2:10 and everyone needs to pay attention to what  AI is unlocking for non-technical people.  
2:15 A huge thank you to Tal Raviv for encouraging  me to meet Zevi. If you enjoy this podcast,  
2:19 don't forget to subscribe and follow it in  your favorite podcasting app or YouTube. 
2:23 It helps tremendously. And if you become  an annual subscriber of my newsletter,  
2:27 you get 19 premium products for free for  an entire year including Lovable, Replit,  
2:34 Bolt, Gamma, n8n, Linear, Devin, PostHog,  Superhuman, Descript, Wispr Flow, Perplexity,  
2:38 Warp, Granola, Magic Patterns, Raycast,  ChatPRD, Mobbin, and Stripe Atlas. Head  
2:42 on over to lennysnewsletter.com and click  product pass. With that, I bring you Zevi  
2:47 Arnovitz after a short word from our sponsors. This episode is brought to you by 10Web,  
2:53 the company that pioneered AI website building  before ChatGPT. In the last three years, over two  
2:59 million websites have been generated with 10Web's  vibe coding platform. 10Web's vibe coding platform  
3:04 is a powerful way to build websites. Think of it  as lovable for WordPress, front end and backend.  
3:10 Users can build any website at any complexity,  e-commerce, portfolios, information websites,  
3:17 blogs, and it comes with the WordPress admin  panel and thousands of ready to use plugins. 
3:21 10Web also offers website generation as an API  as a service for SaaS companies, marketplaces,  
3:27 hosting providers, MSPs and agencies. SaaS  companies can embed it via API so that users  
3:33 can launch AI generated sites directly inside  of their platform, connected to their own data.  
3:38 Agencies and MSPs can get a white label  dashboard to manage clients and resell under  
3:44 their brand. Hosting providers can self-host  the API builder on their own infrastructure. 
3:49 Check it out at 10web.io/lenny and use code  Lenny for exclusive free credits and 30% off  
3:56 API or white labeled solutions. That's the number  10web.io/lenny, vibe coding platform as an API.  
4:06 Today's episode is brought to you by DX, the  developer intelligence platform designed by  
4:11 leading researchers. To thrive in the AI era,  organizations need to adapt quickly, but many  
4:17 organization leaders struggle to answer pressing  questions like, which tools are working? How are  
4:22 they being used? What's actually driving value? DX provides the data and insights that leaders  
4:27 need to navigate this shift. With DX, companies  like Dropbox, Booking.com, Adyen, and Intercom  
4:33 get a deep understanding of how AI is providing  value to their developers and what impact AI is  
4:39 having on engineering productivity. To learn  more, visit DX's website at getdx.com/lenny,  
4:46 that's getdx.com/lenny. Zevi, thank you so  much for being here and welcome to the podcast.
4:56 Thanks for having me, Lenny.  I'm a huge fan of the show and  
4:59 tons of people that I've admired most and  learned the most from. I've been on here,  
5:03 so it's a crazy moment for me.  I'm really excited for this.
5:06 I really appreciate that. I want  to start by reading actually a note  
5:09 I got about you from Tal Raviv,  who is a previous podcast guest,  
5:14 many times newsletter collaborator. One of  the most AI forward product managers that  
5:18 I know I've learned a ton from him. So here's  what he said about you when he introduced us. 
5:23 Zevi is the most hands-on vibe coding PM I know,  and I've personally learned so much from him. His  
5:28 engineers at Meta ask him to teach them how  to do what he does. Every time we get coffee,  
5:33 I repeatedly get this feeling of  everyone needs to be hearing this.
5:37 That's so nice. And so that's the goal. That's the goal of this  conversation is to help more people hear what you  
5:41 figured out. We're going to get very hands-on.  We're going to do a lot of show versus tell,  
5:46 showing people what you've figured out about how  to be a PM, a non-technical PM building stuff. I  
5:52 want to give people a little bit of background  on you because I think this is going to inspire  
5:55 a lot of listeners to feel like they can also  do what we're about to show you. This is going  
6:00 to look very advanced, but just give people a  little bit of sense of just your background.
6:05 I'm very non-technical. I have zero technical  background. Did music in high school. A lot of  
6:10 Israelis do technology units in the Army. I was  not in a tech unit. And basically a year ago,  
6:18 I was traveling with my wife for three  months in Asia and we were in Japan and  
6:23 that was around when Sonnet 3.5 came out.  And I remember watching a YouTube video.  
6:29 I think it was either Greg Isenberg or Riley  Brown and they were basically building apps using,  
6:37 it was either Bolt or Lovable, just using AI. And it was like a crazy moment for me because  
6:42 I was watching this and it basically felt like  someone came up to me and said, "Hey Zevi, there's  
6:48 this cool new technology you should check out. You  should really give it a try. Oh, and by the way,  
6:52 you have superpowers now." And the second I got  home from Japan, I didn't even unpack my bags,  
6:58 ran to my computer, opened Bolt, opened an  account, and for the past year I've been building. 
7:03 And the last thing I'll say on that is we talked  about this a bit before we started recording,  
7:07 but I was prepping with Claude for the episode  and I was trying to clarify what my goal is for  
7:15 this episode. And Claude said, "If people  walk away thinking how amazing you are,  
7:20 you failed. And if people walk away and  open their computer and start building,  
7:23 you've succeeded." So I really hope that  we can inspire some people to do the same.
7:27 I love that so much. I feel like that should  be the goal for my podcast. If you're like,  
7:31 "I love that guest." It's less of a win.  If it's just like, "Oh, I'm so inspired  
7:35 to do the thing that they figured out, that  is the real win." I love, Claude is the best.
7:41 I agree. Okay. So let's dive in and give people, let's  start with kind of a high level overview of  
7:47 how you operate and you use AI in your job. What  are the core tools and just what's kind of like  
7:53 the frame of reference for the workflow  that you figured out and how you operate?
7:56 This all started where I was a project's  power user. I love projects, GPT projects.
8:03 ChatGPT projects? Yeah, exactly. GPT projects and Claude projects,  which are basically a shared folder of chats which  
8:10 share both custom instructions and shared  knowledge base. And I think it was around  
8:15 when GPT started using memory where I thought  it was interesting, but it really annoyed me  
8:23 because I do a bunch of different things. I'm  a terrible runner, I'm a PM, I was a student,  
8:31 psychology student, so I had all these  different facets of life. And what happened was  
8:37 the memory feature was mixing stuff up. So like I talked to GPT about running  
8:41 and it would say, "Oh yeah, after this 5K, you're  going to crush all your next product reviews." And  
8:47 it's like, okay, I understand that you have that  in your memory, but it's just not relevant. And  
8:52 projects basically allows you to compartmentalize  and have things within the right context. So  
9:01 tracking back to the story I told when we came  back from Japan, I started building this app. 
9:05 The first thing I noticed was that these products  were built in a way where, and when I say these  
9:11 products, I mean Bolt and Lovable, were built in  a way where they were super eager to write code.  
9:16 So their system prompt was you're a coding  agent. So when you'd write something, they'd  
9:21 straight away start coding. So at the beginning  of a project, this was super fun and exciting  
9:25 because they just go and start building your app. But later on when things got more complex, this  
9:30 created much more problems because planning  is really important when you're implementing  
9:35 something technical and let's say you're  implementing payments or something that's  
9:40 going to be a change to your database. If  the coding agent is just like, "All right,  
9:45 I got it." And just starts writing code,  this always results in terrible things,  
9:49 some really gnarly bugs that I had. And to mitigate this, what I did was I created  
9:54 sort of a CTO. So again, I'm not technical. I  have been in product for a while, but I know  
10:02 zero stuff about code. So basically what I did was  I created a CTO with the custom prompt of it being  
10:11 the complete technical owner of the project. So I  told it, "I own the problem. I own how we want the  
10:17 users to feel. You're the complete owner of how  this is going to be built. I want you to challenge  
10:22 me. I don't want you to be a people pleaser." All these things that kind of mitigate the  
10:26 regular ChatGPT-isms. I always think about this  where for some reason, the easiest way for me to  
10:32 think about AI is to imagine it as people. And  I think ChatGPT would probably be the worst CTO  
10:39 because it's such a people pleaser and it's so  sycophantic where ... Just a short story I had  
10:46 a few weeks ago, I was trying to learn about  Bun JavaScript, which was acquired by Anthropic  
10:52 and I was trying to understand what they do. So I was talking to GPT and this wasn't within my  
10:57 co-founder CTO project and I asked it if it's  similar to a different framework that I have  
11:04 in my app called Zustand, which nothing  to do at all with what Bun JavaScript  
11:10 does. And basically GPT goes, "Oh yeah, it's  exactly the same." And then it started talking  
11:16 about what it meant and I was like, "Wait,  no, these are not the same at all." And it  
11:20 said the most terrifying and hilarious thing. He goes, "Oh, I'm sorry. I thought you were  
11:24 just making this up and I was riffing  with you." And I was like, "Oh no, no,  
11:28 no, this is terrible." So basically if regular  ChatGPT was a CTO, that would be the CTO who goes  
11:34 along with your dumbest ideas. So creating  the project allowed me to mitigate that.
11:40 So this is, just to be super clear, you have  a ChatGPT project that you've given a prompt  
11:45 to be your CTO of your product  and being a non-technical person,  
11:51 this is kind of like the thing you talk to  when, and we'll get to what you're actually  
11:55 using to build when you have questions about  architecture and decisions that are technical.
12:00 Yeah. So now I'll show my full workflow  and I don't involve GPT anymore,  
12:05 but I definitely would recommend, even though the  technology has gone ... So when I started this,  
12:11 there was no plan mode or ask mode. It was just  build on these products on Lovable and Bolt,  
12:17 and they've progressed a ton. A lot of what  I had as workflows have become ingrained  
12:22 in these products, which is really interesting. I would still recommend start with a project for,  
12:27 first of all, the reason that I said, and also  it kind of puts you in a place where you're in  
12:31 a chatbot and not writing code. So you take the  time to converse and to learn, which I think is  
12:37 critical. And the second thing is if you're  non-technical like me, code is terrifying.  
12:44 It's the scariest thing in the world to look at,  and I look at it as kind of like exposure therapy. 
12:49 I think if you see this where I'm working like  in Claude or in Cursor, you might be excited to  
12:56 start using those, but I would really recommend  starting slow with a GPT project, beautiful UI,  
13:03 super simple, then maybe graduate to like a Bolt  or a Lovable, and then go to Cursor in light mode,  
13:09 slowly, slowly, gradually ease  in until you open a terminal,  
13:13 go full dark mode, go full dev. So I would  really recommend doing this gradually.
13:18 That is awesome advice. And so just to be clear,  
13:22 these days you're using Cursor with Claude Code  powering it. And what I love about that is that  
13:28 you've never written code. The way you put  it, you're afraid of even looking at code.
13:32 Yeah, 100%. You can do exposure therapy, and I love  that Cursor is useful to you. And what  
13:39 you're telling us is that graduating  from a ChatGPT project that is kind of  
13:43 your technical co-founder kind of taught you  enough to feel more comfortable going straight  
13:48 to cursor. You said that you actually went  to Bolt and Lovable kind of in the interim  
13:53 and then you went just straight to Cursor. What's  the reason to just go straight to Cursor? Is it  
13:57 just Cursor can do everything and once you get the  hang of it, it's actually the most powerful tool?
14:02 Yeah, I think I graduated from each tool when I  kind of outgrew it. So Bolt was awesome until I  
14:08 was trying to connect payments to my app  and I kind of started losing it and then  
14:12 I graduated to Cursor and I've actually fallen  in love with Claude. So I'm using Claude Code,  
14:16 but that also runs within Cursor, and  I think this is Tal who told me this. 
14:22 I'm not sure who he's quoting, but code is  just words at the end of the day. So it's  
14:27 just files on your computer. So basically you  can be working on the same project and carry  
14:31 it from app to app. And especially now,  I can work with multiple models and apps  
14:36 on my project. So start slow, but definitely  there's a lot of places you can graduate to.
14:42 Awesome. Okay. Should we dive into  screen share showing how you operate?
14:47 Awesome. I pulled up cursor. Can you see it? Mm-hmm. Perfect. So within my code base, what you can see  here on the left, these are all my code files.  
14:57 Here on the right is Cursor. So this is  basically like having AI, which has access  
15:02 to all the code. And here in the middle, I have  Claude Code running. And what you can see here,  
15:06 I'm going to close cursor for a second.  What you can see here are all my /commands. 
15:10 Basically what /commands are, they are reusable  prompts that I save within the code base that  
15:16 I can run by writing / and then the name of  the file. So here you can see Create Issue,  
15:22 which is the first command that I'm going to  use. And basically what this tells Claude,  
15:26 it says the user is meant development and  thought of a bug or a feature and improvement,  
15:30 capture it fast so they can keep working. And then it basically says, this is the  
15:34 format that I want you to capture the linear  issue in, and it explains a bunch of things  
15:40 what exactly Claude needs to do to get there.  So the way I invoke this is basically I'll do  
15:47 /create issue and this injects this prompt  into Claude. So it says, "I'm ready to help  
15:54 you to capture this issue, what's on your mind." So basically when I'll do this is if I'm working  
15:58 on a big project and I suddenly come across  a bug or have an idea that I don't want to  
16:03 work on right now, but I want to work on later,  I'll do this really quick and Claude's main goal  
16:08 is to quickly capture what I'm thinking about.  So quickly to run through my full workflow. So  
16:14 basically it starts with creating an issue.  So this is the create issue /command, which  
16:19 basically tells Claude that I'm mid-development  and it should quickly capture what I'm thinking  
16:24 about and create an issue within linear. Then later on, when I want to pick this up,  
16:29 I have the exploration phase. Exploration  phase is basically telling Claude,  
16:33 we're going to only explore what we want to  solve here. It could either pull from linear  
16:37 or I can just speak freely to it. And what it  will do is it will analyze and understand the  
16:43 issue and just ask clarifying questions.  The next phase after we've done finished  
16:47 exploration phase is we're going to create a plan. So you can see create plan. This basically has  
16:53 a template that I love for creating plans, and  the output of this at the end of the day will be  
16:59 a markdown file with our plan that we can end up  building along with code. After creating the plan,  
17:04 we have execute plan. After execution, we  have review and then we have peer review,  
17:10 which is really cool and we'll get into  later on, and at the end, we update the docs. 
17:14 So this is updating documentation and everything  so that agents can write better code later on. So  
17:19 I think what we'll do is we're going to build a  feature live for my app, which I think is really  
17:23 cool. But first what I'd like to do is show  you the app so you have some context. So this  
17:28 is StudyMate. It's a platform for students, which  allows them to upload study materials and create  
17:36 interactive tests based on their own materials. So here we can go to the top. Let's upload a PDF.  
17:46 We can decide what pages we want to be quizzed  on. We can decide the number of questions,  
17:52 the difficulty level. And basically what happens  behind the scenes is we send the information the  
17:59 user uploaded along with the system prompt and  any other augmentations the users decided to  
18:05 Gemini and we create a quiz. These are challenging  questions that are meant to assess comprehension.  
18:12 You even have some hints and once we do a few  of these, we can submit ... I got them right?
18:21 Terrible results. Yeah, lucky. And so just to be really clear about this, this  is like a side business that you have, an app that  
18:27 you built that's making money. That's like a thing  you vibe coded having no technical experience.
18:32 Yeah, this is my weekend  project. Yeah. This is what I do.
18:34 Amazing. On weekends. Yeah. So you get basically  deep explanations into why each  
18:41 question was wrong or each question was  right. And at the moment, StudyMate only  
18:45 has multiple choice questions. And I was doing  some competitor research over the last weekend  
18:51 and I saw competitors who had true or false  questions and also fill in the blank questions,  
18:56 which I loved. So I think that'd be really cool  if we could build that live. How's that sound?
19:00 I love it. Crossing your fingers, this all  work. I just want to highlight the stuff  
19:05 you shared right before this in cursor. So this  is a huge deal, what you describe here. This is  
19:10 essentially what you've figured out is a way as  a person that has no idea how to write any code,  
19:16 how to build a product in Cursor as a product  manager using this series of /commands that  
19:22 you've concocted that you're going to be sharing  with listeners. They can download all these and  
19:27 just use them directly. They don't have to figure  out all these prompts that you've figured out.
19:31 Yeah, 100%. Basically what happened  was I formulated the backbone of this  
19:36 with the CTO, and it was basically  within the system prompt of the CTO  
19:40 project that I had within GPT. So it  said, step one, we do this. Step two,  
19:44 we do this. And now I'll keep building. And if I  see something that happens over and over again,  
19:51 I'll just create a /command and then it  will be automated within the workflow.
19:54 Amazing. So just to summarize the /command.  So one is create an issue in Linear, which  
20:00 I love. Linear is awesome. Shout out Linear. That's also from the product pass.
20:04 From the product pass. Oh my God, what a value?  Okay. So step one is create the issue in linear.  
20:09 So it's a command. So this prompt /command you've  created, just create issue. Then it's explore,  
20:16 which is explore the idea, help me ideate on  what this could be. And this is Claude helping  
20:21 you think through the feature and product.  And then it's actually create the plan. 
20:27 And so it's like the AI helping you build the plan  to build the product. Then it's actually execute,  
20:33 which is just build the thing.  And then there's this review,  
20:37 peer review step, which is awesome that  you'll share. And then there's document,  
20:41 update documentation based on the  new feature that we're adding. Sweet.
20:44 Yeah. Cool. So let's go ahead and start  building. So I'm going to use Wispr Flow  
20:52 to dictate and basically this starts with /create  issue. So this basically sends that prompt. And I  
21:01 love this because I usually do this during when  I'm building something else. So basically it  
21:07 tells Claude that I'm mid building something and  I don't have a lot of time to waste time on this. 
21:12 So just ask some brief questions so that  you have enough to capture within linear.  
21:16 So I want to add fill in the blank questions to  StudyMate. I want this to be 30% of tests to be  
21:25 generated as fill in the blank questions.  I want there to be six potential answers  
21:30 for two blank spots, and of course there's only  going to be two correct answers. So one correct  
21:36 answer and two incorrect answers for each spot  and I want the interface to be drag and drop. 
21:43 So that's just basically a quick  think of how I want this to work.  
21:49 So it's going to ask me a few questions. Quizzes  are 100% multiple choice, question structure,  
21:56 single sentence passage to blanks and priority.  So one and two are correct, and this is  
22:05 not high priority. It's a nice to have feature. So now basically what Claude is going to do is  
22:13 it's going to use MCP, which is basically  a technology that was created by Anthropic,  
22:20 which gives AI the ability to use tools. So  this is connected to my linear. So what it's  
22:25 going to do now is it's going to use everything  we've said and create an issue within Linear.
22:30 And by the way, as this is loading, I just  love the way, the way you described this,  
22:34 especially doing voice mode, it's like exactly  how you would talk to an engineer describing a  
22:39 feature, "Here's what I want." And then they  ask you questions, here's the clarification.
22:43 Yeah. So at first when I was doing this with  the CTO, I would do it with ChatGPT voice mode,  
22:50 and that was crazy. That literally felt like  ideating with a person. It would push back,  
22:55 ask questions, and maybe one day the coding tools  will get there too, but that was exactly ... It  
23:02 really felt like sitting with my CTO. Great. So  created STU88. So if we open up Linear now, we  
23:09 should be able to see ... Let's see where STU88. There it is. Fill in the blank questions with drag  
23:17 and drop interface. So it has a TLDR, it has the  current state. It did a little bit of research on  
23:22 the code base, I think, expected outcomes, some  context. So yeah, so this is basically ready for  
23:28 me to pick up when I'm interested in building.  So now let's say a few days go by, I finished the  
23:34 current project I'm working on, I can pick it up. So when I pick it up, I do /exploration phase,  
23:39 which is what we said. And then instead of  pressing enter, I'll press tab and I'll show  
23:44 you this. So basically, exploration phase, what  it does is it will take an argument. This is  
23:51 basically a placeholder within the prompt,  which allows me to enter something that is  
23:57 extra context for the AI. So I can say here,  Linear STU88, which is referencing the ticket.  
24:07 And now what it's going to do is it's going  to go, it's going to fetch the Linear ticket.
24:12 And what's the idea? What's  the goal of the exploration  
24:14 phase? This ideate on the idea. Is that the- Exactly. Okay. So it's both for the CTO to deeply  understand the problem that we're  
24:22 trying to solve and also understand  the current state of the code base,  
24:26 what files need to be affected, and how is  the best way to implement this technically.  
24:31 And usually what happens is right now, Claude  is just basically reading a bunch of files,  
24:35 understanding the basic structure of the  code, and then it's going to come back with  
24:39 a bunch of clarifying questions that will  decide how we end up implementing this.
24:44 So it feels like it's talking  to your engineering manager.
24:48 Exactly, exactly. 100% this  is how I think about it. And you said that you're a CTO,  so you used to use ChatGPT prompt  
24:56 to have a CTO in there. Now the  CTO's living inside here in Cursor?
25:00 Yeah, because of the way the tools have developed  and they've become so good at both exploration and  
25:07 code execution. So now it's just a habit that  I call it a CTO, but it's basically all in one.  
25:13 The same agent will both do the exploration and  write the plan and end up executing the code.
25:19 Got it. So it's basically it's  Claude Code. Is there a prompt  
25:21 you gave it to act like that? To  act like the right kind of CTO?
25:26 Yeah, so within the Claude.md, which is  basically the system prompt that's loaded  
25:33 within Claude's context in every conversation, I  have some basic stuff like this is our workflow,  
25:39 this is how we work. Within exploration  phase, I want you to challenge my thinking,  
25:43 all kinds of stuff like that that can  be loaded within the Claude.md file.
25:47 Cool. One last question before we move on here,   just because I'm thinking about it as this  happens, the Linear issue that you generated,  
25:55 how often is it actually great and ready? How  often do you have to edit it? What's the quality  
26:00 of the Linear ticket that it generates? Because  a lot of people are probably wondering just like  
26:05 all these terrible linear issues are being  created by AI. Are they actually any good?
26:08 It's completely different because I'm a company  of one. So a lot of the context is within here  
26:15 and there's no need for me to talk to other teams  and understand. It's basically very accessible,  
26:20 and also I can easily see when  Claude understood something wrong. 
26:26 I don't want to say that I would create Linear  issues at work like this, but definitely if you're  
26:31 building your own side project, they're pretty  quality. And also, it just kicks off when I want  
26:38 to start working on it. I wouldn't say it's ready  to be built. It's ready to start being explored.
26:43 Got it. So it's just the beginning of an idea.  Actually, let's come back after we go through this  
26:47 flow of how you would approach this if you were  at say Meta or another, maybe a smaller company,  
26:52 how this workflow might work at a larger  company that isn't just your own startup.
26:56 Yeah. Interesting. Let's come back to that. Cool. All right. So this is Claude coming back.  I have a comprehensive understanding of the code  
27:05 base. I thoroughly analyzed StudyMate live  codebase and understand the current system,  
27:11 feature quest and key areas that it's identified.  Usually I'd spend a lot of time going over this  
27:17 because this is super, super important, but  just for the sake of development right now,  
27:21 we're going to brush through this. Now Claude basically comes back after it's  
27:26 gone through the code base and understood the  way it currently works. It's basically telling  
27:30 me what the current understanding is. So it's  talking about how the app is set up at the moment,  
27:36 how the data is structured, what it understood  from the feature quest, and what it has identified  
27:43 as key areas, and then it asks me some questions. So it's asking about the scope, it's asking about  
27:48 the data model, the UX/UI of the feature, how  it should be validated, how it should be graded,  
27:54 what changes need to be happened to the AI  system prompt and all kinds of questions  
27:59 about the app. I've prepared answers to all these  questions beforehand because I don't think we all  
28:04 want to sit through this. So I'm just going to  paste that in and we'll see what Claude says.
28:08 Awesome. I love it. And I love just scanning those  questions I was asking. It's like such smart,  
28:13 sophisticated, important questions instead of  just, "Cool, here I go, I'm going to build it."
28:18 Yeah, and I think this is the big difference  between just vibe coding and going along with the  
28:23 vibes and really building serious apps. I spend  a lot, a lot of time going back and forth and  
28:29 understanding. Also, a very cool /command that I  haven't showed yet is learning opportunity, which  
28:36 basically when something is really difficult for  me to understand, I'll do /learning opportunity  
28:42 and then talk about what I want to learn. And this basically primes Claude and says,  
28:47 "I am a technical PM in the making. I  have mid-level engineering knowledge.  
28:51 I understand architecture and basically I want  you to explain what we're currently working on  
28:58 using the 80/20 rule." So this is a great way to  learn. I would definitely take this and every time  
29:02 you kind of see something that you don't fully  understand, I would definitely use this to learn. 
29:08 Great. So Claude basically comes back and says how  it understands the current data model and how it's  
29:16 going to implement. Yeah, so it's ready to create  the plan. So basically what I'm going to do now  
29:21 is I'm going to go and do /create plan and while  Claude is doing this, I'm going to show really  
29:29 quick what this looks like. So basically, these  plans are from a template that I found on Twitter. 
29:36 I forgot who it was, but it was just a  template that really resonated with me. And  
29:41 it's basically saying, based on our exchange,  create a markdown file that will be the plan,  
29:47 include clear, minimal, concise steps, track the  status. So this basically has like status trackers  
29:53 on each task that Claude updates as it's going  through and it will have a TLDR, some critical  
29:58 decisions that we've made and the plan itself. So Claude has finished writing the plan,  
30:03 so we'll be able to look and see exactly what the  plan is. So it has a TLDR, it has the critical  
30:10 decisions we've made and the tasks broken down.  And this is a perfect plan and it's also a really  
30:16 good way to write this because a lot of times,  I'll use different models to execute certain  
30:21 stuff. So Cursor has an amazing model called  Composer, which is superfast. So a lot of things  
30:27 that are not that complex, I'll use Composer. Gemini 3 that just came out is unbelievable  
30:34 at UI. So a lot of times, I'll split  the plan into backend and front end,  
30:38 and then I'll have Gemini just read the plan and  do the front end. So having this as a markdown  
30:44 file is really good. And also going forward,  it's really good to have within the app so that  
30:48 later on, if an agent is writing code in a certain  area, I can see what's already been done there. 
30:54 So what we're going to do now is we're going to  execute the plan. So now I think we're going to  
31:00 do this with Cursor just because Composer is so  freaking fast. So what we can do is basically  
31:06 just say execute and then we can tag the file. And  Composer is ridiculously fast. So that's it. It's  
31:15 off. It basically understands what the plan is and  it's going to go ahead and start writing the code.
31:20 Let me ask you a question while this is happening. Awesome. I have many questions, so this is  a good time to ask a few of them.  
31:26 You said that Lovable and Bolt and other  apps in that space are just not enough to  
31:32 build really serious apps and you have to move  to Cursor to do that. Tell us more about that.  
31:37 Just what's the limitation you ran into with  those products and why you switched to Cursor?
31:42 I started using Cursor and Claude Code a few  months ago and I haven't looked back, but at that  
31:50 time, these teams have been moving like crazy.  So I don't want to say I wouldn't trust them. I  
31:56 don't know what the current state is. But for me,  it was basically the issue of I felt that Bolt was  
32:02 being very opinionated on how I should do things.  And I felt like my knowledge has gotten to a point  
32:08 where I can graduate and be more in control. By the way, I think that the main difference  
32:13 between all these tools is basically the harness.  So the models are all the same models. I'll run  
32:18 Claude within Cursor, I'll run it within Claude  Code, and it's also the models that Claude is also  
32:23 the model that is underlying Bolt and Lovable,  but basically, Bolt and Lovable will add a bunch  
32:29 of levels in the middle that will take all kind  of guesswork and hard decisions out for the user. 
32:37 So the user doesn't have to make these hard  decisions. So it's also very easy to build,  
32:41 but the flip side of that is that you have  less control. And basically Claude Code is  
32:44 just taking Claude and shoving it straight in  your code system and giving it full tools and  
32:49 to do whatever it wants, but also with that comes  a lot of decisions that you need to make. So I  
32:55 don't know if you can't build really amazing  production apps using Bolt or Lovable now,  
33:01 but I think basically if you want the most cutting  edge abilities of the models and you want to  
33:06 be able to make all the decisions on your own,  it's probably best to be on one of these tools.
33:13 What I'm feeling and hearing is  that planning work that you did,  
33:16 that's the stuff that Lovable, Bolt, and  would you put Replit in that bucket too?
33:20 Yeah, for sure. Lovable, Bolt, Replit, Base44. v0. Yeah, v0, all same bucket.
33:27 So essentially, they're doing that  planning for you. And as you said,  
33:32 they're very opinionated. They try  to make it easy. So it's just like,  
33:35 "Here's how to do it. Here's the way we think  is best for people." And what you're saying is  
33:40 once you're trying to get a little more serious  about it or want to go in a different direction,  
33:44 you don't have the power to change how  they plan. So Cursor lets you do that.
33:49 Yeah, I don't want this to come out like  I'm badmouthing them. Base44, let's say.
33:53 No, absolutely. Yeah. Base44 does an amazing job at basically  taking all the complex guesswork out of  
33:58 building product and just allows you  to just go with the vibes and build,  
34:03 but it will do sign in with Google  for you and it will do a database,  
34:07 but then you don't have decisions on what database  am I using. Do I need sign in with Google this way  
34:12 or another way? It would just do it out of the  box. So that's basically the trade-off there.
34:16 Awesome. Shout out Maor, the founder of Base44. Yeah. Maor Shlomo. Yeah. He's amazing.
34:24 Okay. I just love how this is like the way you're  like flinging, what's the word? Slinging models  
34:32 like Gemini 3 for frontend. I love that you've  never written any code and you're just like,  
34:37 "Cool, use Gemini for this and Claude  for this. And I'm just working on Cursor,  
34:41 talking to this CTO, helping you build  stuff and build significant product."
34:48 Yeah. We just live in the craziest of times where  basically the world changes once a week, it feels  
34:56 like. And there is just no boundaries. You can  use all of these just on your regular MacBook  
35:03 or regular laptop. And I have these moments,  I call them time machine moments, which is  
35:10 basically this week, for instance, I was prepping  for the podcast using Claude with a project. 
35:17 I was building, I was fully localizing StudyMate  from Hebrew to English, which I did in two days,  
35:22 which would probably take a dev team weeks. And  I was building a personal site, which went from  
35:28 no domain, no nothing, to live on a domain within  an hour and a half. And I was doing all three of  
35:33 these in parallel. And there was a point where  basically all three of the agents were running,  
35:39 so I didn't have anything to do. I just had to let them think,  
35:42 and these are the time machine moments where I  feel like I was in the future and I just stick  
35:46 my head out of the time machine and whoever's  next to me, at the moment it was my wife,  
35:51 I'll just say, "We live in the future." And she'll  be like, "Huh, what?" And I'll be like, "No, no,  
35:55 don't worry about it." But it's just basically so  crazy that all these things are just an API away.  
36:01 You can use anything. So I think it's an awesome  time to be curious and optimistic and hardworking.
36:08 These are my favorite kinds of podcast guests.  People that are living in the future, figuring  
36:13 out all these things and then are just kind of  come back, as you said, poke your head out of the  
36:17 rocket ship and just like, "Hey, here's this thing  that I figured out. Here's where we're going."
36:21 Yeah, best time to be alive. So  awesome. So it looks like it's finished.  
36:27 So now what we're going to do is we're going  to run the app locally and we'll be able to see  
36:32 what Composer ended up building and we're  going to see if anything else is needed  
36:38 on our end to maybe do some manual  review. Does that sound good?
36:42 Sounds great. And I love, that was  like, I don't know, a few minutes  
36:45 where if it was a human engineer, it'd  be like days, maybe a week for work.
36:50 For sure. Yeah, for sure. No, Composer  like the one thing is it's just so,  
36:55 so blazing fast, keeps you in flow.  So yeah, full features take minutes.
37:00 And it probably costs a  couple bucks in AI credits. I don't even look. I used to be so stingy about  paying for products and now I'm just basically,  
37:10 I look at it all as tuition,  as stuff that I'm paying for  
37:14 learning. So I don't know how much it  costs, but it's definitely worth it.
37:18 That explains why they're the  fastest growing products in history.
37:21 100%. Your marketing website sets the tone for  your brand and is the one touchpoint that  
37:27 every single one of your customers sees.  In today's age, if you're still having  
37:31 a hard time making small changes and simple  updates to it, you are doing something wrong.  
37:35 That is why so many companies from  early stage startups to Fortune 500s,  
37:40 including companies like DoorDash, Zapier,  Perplexity, and ElevenLabs turn to Framer,  
37:44 the website builder that turns your .com from a  formality into a tool for growth. Framer works  
37:50 like your team's favorite design tool and comes  with real-time collaboration, a robust CMS with  
37:56 everything you need for great SEO and advanced  analytics that includes integrated AB testing. 
38:01 Changes to your Framer site go live to the web  in seconds with a single click and without any  
38:06 help from engineering. Whether you want to  launch a new site, test a few landing pages  
38:10 or migrate your full.com, Framer has programs for  startups, scale-ups, and large enterprises to make  
38:16 going from idea to live site as easy and fast as  possible. Learn how to turn your website into a  
38:23 growth engine from a framer expert or get started  building for free today at framer.com/lenny,  
38:29 that's framer.com/lenny. Rules  and restrictions may apply.
38:33 So now we have this feature, which basically  we built and I can ask it to make some changes  
38:39 because it's running locally and once  it's ready, I'll be able to ship it to  
38:43 users. So now the next phase after I've  QA'd it and basically tested it manually,  
38:49 I'll have Claude review its own work. So  what I'll do is I'll reopen Claude Code.
38:56 I love this because this is one of the things  that comes up a lot in this podcast is writing  
38:59 code is now so easy. The main challenge people  have is reviewing the code that AI has written.
39:04 100%. And what you're doing here is you're  having Claude review its own code.
39:08 Yeah, so this is another thing where it's  very difficult for me to catch mistakes.  
39:13 So my review process has gone through a bunch  of iterations to really be as good as possible  
39:19 and to catch as many things as possible. So  I'll always manually QA at first to make sure  
39:25 if I can see any mistakes that Claude made.  And then what I'll do is basically /review. 
39:32 And this tells Claude to start reviewing its own  code. But what's even cooler and something that  
39:39 I'm really proud of is I will usually do multiple  reviews and I'll have Codex, which is ChatGPT's  
39:47 competitor to Claude Code, as well as cursor open,  and I will have each of them review the code. And  
39:55 then what I do is I have a /command called peer  review, which is really interesting. And basically  
40:02 what it does is it's going to take Claude,  which is usually the agent who I'm working with. 
40:07 And just to put this in a mental model, this  is basically my dev lead that I'm working with.  
40:16 The /command is basically saying, "You're the dead  lead on this project. Other team leads within the  
40:22 company have looked at your code and reviewed  it and found these issues." Don't take what  
40:27 they said at face value. The reason is you have  more context than them and you led this project. 
40:33 You need to either explain why the stuff they  found are not real issues and wrong or fix them  
40:40 yourself. And it's really cool because the way  I look at these things is I look at the models,  
40:45 I try to imagine them as people and I  can really tell you how each one of these  
40:51 would be as a real human because they have- Each model. Yeah. Each model has such distinct  characteristics. So let's say Claude, she would  
41:00 be the perfect CTO. She's very communicative.  She's very smart. She doesn't just go with the  
41:07 flow and do whatever you tell her. She's very  opinionated, but also super collaborative, which  
41:12 is I think why I'm always drawn to Claude because  I need to do so much learning and it's your dream,  
41:20 a very communicative, but very opinionated  dev lead, but then there's also Codex. 
41:27 So I use Codex 5.1 Max, whatever. I don't  know, they're not the best at naming models,  
41:33 but GPT's model. I always imagine it as like  the best coder within the company who comes  
41:40 to the office with a hoodie and sandals and  sits in a dark room. And you basically only  
41:46 bother him when you have the worst bugs and you  say, "Listen, we have this bug and it will just  
41:50 close the door for two hours and come out and say,  I fixed it." And you're like, "Wait, what? Are you  
41:55 going to tell us what happened or whatever?" And he's like, "Don't worry about it. I fixed  
41:57 it." It's like really not communicative, but  it solves all the worst problems. And let's  
42:03 say Gemini is like a crazy scientist who's  super artsy, super talented at designing,  
42:10 but if you sit next to it and watch it work,  it's terrifying. You would fire that person  
42:14 instantly. This might be just my experience,  but when I'm using Gemini within antigravity,  
42:19 which is Google's new competitor to  Cursor, when it's writing code, you can  
42:25 see the steps it's taking and it's terrifying. You'll say, "I want you to redesign the top of  
42:30 the dashboard." And you're looking at its thought  process and it will say, "Oh, first things first,  
42:35 I'll delete the dashboard." And then it'll be  like, "Nope, that was a mistake. I'll bring  
42:39 it back." And then it will say, "Oh, can I  edit the database?" And you're like, "No,  
42:44 do not edit the database. You're just doing a  redesign." And then it will end up designing  
42:47 something beautiful. So the way there is a  rollercoaster and very scary, but at the end  
42:53 of the day, Gemini is very good at design. So I think that using all these models  
42:59 and basically playing to their strengths  and mitigating their weaknesses by using  
43:04 other models is a game changer for me. So I'll  do peer review a bunch of times and I'll have  
43:10 other models review other models code and kind  of have them fight it out basically. Sometimes  
43:18 Claude Code will get really sassy and be like,  "This has been raised for the third time. And  
43:22 for the third time I'm telling you, this is  not an issue. This is by design." So it's  
43:26 just a really cool thing that I've added  and I haven't seen many people doing it.
43:31 That is such an incredible rant/way  to understand what's going on. Okay,  
43:38 awesome. So we just ran the review.  So show us what we saw there and let's  
43:42 actually try this peer review. I'm really  excited to see what you learned there.
43:45 Yeah. So basically Claude has reviewed its code  and it's found a bunch of bugs, a critical bug  
43:51 it found in the prompt, some high bugs, some  medium bugs, and now what I'll do is I'll do  
43:57 the same thing with the other models. So Codex has  a built-in code review that you can do, or I just  
44:05 like to say review all the code in this branch. Of course, branch is referring to the GitHub  
44:13 branch that we're working on. We're not working  on the live code base. And then I'll do this with  
44:18 Composer say with, let's do it with Composer One.  So I'll do /review here as well. And basically  
44:27 these are both going to run and do a in-depth  review similar to what Claude does. But again,  
44:32 because of the differences between the models,  they're all going to catch different things and  
44:36 they're all going to look differently,  and this is a really cool way to work. 
44:40 It's basically if you had other team leads  within the company review the code. Here,  
44:44 you can see how fast Composer is. I think GPT  probably will take a bunch of time. Like I said,  
44:49 it's in its own dark room right now reviewing  code and we'll come back in a few minutes.
44:55 Okay, cool. So we can let these  run in. We don't actually have to  
44:57 go through the whole process, but is  the idea once you get these results,  
45:00 you run peer review and you copy and paste  kind of these results. Is that the idea?
45:04 Exactly. I'll copy and paste the results. I'll  do peer review and then I'll say dev lead one  
45:08 and then paste from one of the models. And then  I'll say dev lead two and paste from the other  
45:12 model and basically have them fight it out  until I feel like we have no more issues.
45:17 Incredible. For me, this is super important  because I'm not technical and  
45:21 I'm not a developer. And I'll also  use /learning opportunity a bunch  
45:25 during this to learn about stuff that I  don't understand or don't fully grasp.
45:31 Incredible. What a clever solution to solving  this code review problem where it's like,  
45:35 I don't know what you ... I don't know how to  recode, so what am I going to even ... Yeah.
45:39 Yeah. Okay. Incredible. Let's wrap up this  workflow. Is there anything else that's  
45:45 important in this workflow? And again,  all this stuff is going to be available  
45:48 where people can just plug this stuff into  their Cursor account and use it themselves.
45:52 100%. The one thing I'll say is that I think just  like working in general with AI and even just like  
46:01 working on any product, doing constant  postmortems is critical. So a lot of  
46:08 times we'll find all these kind of bugs or  maybe Claude will fail to execute something  
46:13 correctly. And at the beginning when I  started vibe coding, I would basically  
46:17 just keep running at it like running at the  wall and until it worked. And once it worked,  
46:21 I was like, "All right, awesome. This works.  Let's keep going." But I've learned over time  
46:25 that updating documentation and tooling is  one of the biggest hacks for productivity. 
46:31 So when Claude will fail to do something or I'll  see this really bad bug that shows that Claude  
46:37 really didn't understand something, I'll ask it,  "What in your system prompt or tooling made you  
46:42 make this mistake?" And Claude will kind of like  go introspective and think of what made it create  
46:48 that mistake. And then I'll say, "Okay, great.  Let's update your tooling and documentation so  
46:53 that this mistake never occurs again." And  I do this every time I'm either building  
46:57 an internal tool or anything. And  I think this is just like working. 
47:03 If you end up doing a bunch of mistakes and  then end up releasing the feature to users,  
47:08 so you're like, "All right, it's a big success."  But going back and even when you've succeeded,  
47:13 looking and understanding what you did and  what you could have done better is critical.  
47:17 And also using AI, this is probably one of the  biggest unlocks. Going back to your prompts,  
47:22 understanding what was not good enough, iterating  on them and then seeing how AI's responses get  
47:28 better, I think that's probably one of the most  important things and one of the things that  
47:33 divides between people who are okay with using AI  and the people who actually know how to use it.
47:38 That is such good advice. So what  I'm hearing is when the models do  
47:43 something dumb, make a mistake, you ask it  to reflect on what the mistake it made was,  
47:48 and then you update the /command prompts  with that knowledge so that in the future,  
47:54 it's not making that same mistake,  and it just keeps getting better.
47:56 Exactly. These things just keep getting smarter  and smarter. So you're building up this  
47:58 really incredible prompt that  just gets better and better.
48:01 Exactly. Not always the /commands.  It will sometimes update  
48:05 different documentation or its tooling,  but basically it's understanding what  
48:09 the root cause of the mistake  that the AI made and fixing it.
48:15 Awesome. So the models are getting smarter  and then there's also the other parts of  
48:20 your workflow can get smarter as you  find flaws in the way it does stuff.
48:25 100%. Yep. Amazing. Okay. Is there anything else there  before I move in a couple other directions?
48:31 I think that's it. I think we covered pretty  much everything. Basically, just to wrap this up,  
48:35 what I do is I do a bunch of code review and  then update the documentation so that everything  
48:39 is documented. So the next time I try to build a  feature in this area, there won't be any mistakes.  
48:45 And then I'll do a bunch of testing. I'll do some  user testing as well before I release this to  
48:50 general availability. Obviously we're not going to  release this. This was just a show, but hopefully  
48:54 maybe by the time the podcast comes out, I'll  have done this correctly and release the feature.
48:59 It's incredible that this was not possible, I  don't know, two years ago, maybe a year ago,  
49:04 you are a product manager shipping a  product without knowing how to write code,  
49:11 barely knowing how to review code. You said you're  afraid of looking at code. As a product manager,  
49:15 you're building a product in Cursor using  all of these different AI models. You're  
49:21 making money with this product. We're so used to  this now, but it's insane what is now possible.
49:29 It's the best time to be live. 100%.  I think that I understand the fear,  
49:35 but AI just makes it so much possible. Just  a quick side note here, my brother, who I'm  
49:39 building one of the apps with is an entrepreneur.  He has a beautiful business that helps old people  
49:45 and seniors understand to use technology and  AI better. And he's basically doing the same  
49:52 kind of learning as me, and he's replaced  all of the tools he was paying for. I think  
49:56 he was paying for Zapier and Airtable, and he's  basically built a full-fledged CRM system and  
50:03 automation system for his business completely  alone. So for the people who are curious,  
50:10 optimistic, hardworking, this is  the best time to be a builder.
50:14 And what I love about this conversation  we're having here is it feels like the  
50:18 biggest barrier for a lot of people is like,  how do I get started? What exactly do I do?  
50:22 I open up cursor. It looks very intimidating.  I don't know how to write code. I don't know  
50:27 how to build stuff. I don't know about  databases. And so you're going to be  
50:30 sharing all these /commands and basically  this whole workflow with the audience.
50:34 Yeah. Okay. And like I said, just start at GPT.  Start on GPT, tell it what your idea is,  
50:40 tell it to explain to you what are  even the first steps of thinking,  
50:43 what are the decisions you need to  make? And just be inquisitive, learn.  
50:48 Don't rush things. It's very important to just  dive in and really spend the time to learn.
50:55 And you share this. One of your /commands is  learning opportunity, and it's how you learn  
50:59 a lot of these things. Just teach me this  thing and how this database issue works.
51:04 Exactly. Okay. There's a couple directions I want to  make sure we touch on. One is coming back to  
51:08 a question I asked earlier about how this might  work at a larger company. Say it's not like Meta,  
51:12 but just like, I don't know, a thousand person  company, 500 people. How much of this can you  
51:17 plug and play into a workflow as a PM at a  larger company? What would be your advice for  
51:22 someone that may want to start trying to ship  code, at least showing people what's possible?
51:28 I think that first making your code base AI native  is a really important step, and I think this needs  
51:34 to be done by technical people. So basically my  codebase has a ton of just plain text in it. So it  
51:43 will have a bunch of markdown files that explain  to agents how to work in certain areas of the code  
51:48 base and high level structure so that the  agents navigate through the code base easier. 
51:54 And I think that if this is set up in  a really good way, I still don't think  
52:00 PMs should be shipping heavy database  chain migrations or any big project,  
52:06 but contained UI projects, especially if  you just build it, create the PR and send  
52:13 it to a dev to do the final finishes. I  think that's definitely something that's  
52:18 possible. And I think we're going to see  that a lot in the next coming years. I  
52:20 think basically everyone's going to become a  builder, so it should be really interesting.
52:25 Okay. So your advice here is as a PM, maybe don't  go right to Cursor, start building, shipping,  
52:31 trying to ship features to production, especially  complicated features. Do you think we'll get  
52:35 there? Do you think in a couple years, PMs will  be doing this and it'll feel less scary and crazy?
52:42 If there are PMs. Yeah, I think titles are  going to collapse and responsibilities are  
52:49 going to collapse and everyone's just going to  be building. I definitely think that the models,  
52:55 the context window is getting bigger, the models  are getting smarter and I definitely see how PMs  
53:02 or any other background can be writing. At the  moment, I wouldn't wait for that. I would use this  
53:07 as a collaborative learning opportunity to work  with your dev team. It's going to be difficult. 
53:14 A lot of developers are very, very  skeptic about the current state,  
53:18 and I think that it's going to be a lot  of sales work on your end to convince,  
53:22 but if you're able to convince, and I think  teams that are really sold on this and want  
53:27 to take the time to work on their workflow  about how can our team become more AI native,  
53:33 I think that these teams are going to probably  be a few years in the future and they're going  
53:38 to look back at the few weeks they spent  setting this up as the best time they spent.
53:43 Let me ask you another question around just the  job of a PM. One of the biggest fears people have  
53:48 with these AI tools for PMs for every function I  imagine is just you start to rely on these things,  
53:56 your skills start to atrophy, you're  producing all this slop that looks great,  
54:01 cool, amazing strategy doc. No, it's  actually not at all good. Are these Linear  
54:05 tickets or just products that are half-baked? What's your take on these two parts of just like,  
54:10 how has this impacted your craft as a PM?  Do you feel like this is weakening your  
54:16 skills because you're so reliant on these  tools and just how do you keep the quality  
54:21 of this stuff up and not just like, "Eh,  it's just a bunch of AI generated slop."
54:25 I have a very strong disagree to this and I've  heard it a bunch. I remember when I started using  
54:32 Tal Raviv has this whole course on building a PM  Copilot using projects, which is probably one of  
54:40 the best courses that you can take. And when I  started working with my own Copilot, I remember  
54:46 people at work looking and saying like, "Oh, so  you're basically outsourcing your thinking." And  
54:51 to me, that's just the worst way to look at it. And I think for some reason, these people usually  
54:59 have a high correlation with the kind of person  who doesn't like to show their presentation when  
55:03 it's only 10% done or doesn't want to ask for help  a lot. I think that there's a misconception with a  
55:10 lot of PMs that the job is always having the right  answers and being the smartest person in the room.  
55:16 And at least how I was trained and how I believe  the role of the PM is, it's the exact opposite. 
55:22 It's basically harnessing anything that can get  us as quick as possible to delivering the right  
55:29 solution to users. And I just think this is  like that really smart person that has context  
55:34 or your mentor or whatever, but is just always  available and doesn't judge you and can really  
55:40 help you. So if you're using it to just create  your outputs and then putting them out there,  
55:46 yeah, that's AI slot, but it's also human error. I think it's really important that you own your  
55:52 own outputs. If you put anything out there or show  something in a product review and you say, "Oh,  
55:58 sorry, that was built by AI." That's your mistake.  I think if you use these intentionally and really  
56:06 take the time to understand how to use AI in  the correct way, it's one of the biggest game  
56:10 changers that will make you much better as a PM. And another thing here is that, especially for  
56:17 more junior PMs, it allows you to play at such  a higher level than you would normally. I think  
56:23 that at Wix, I wasn't thinking of what's the  marketing strategy of the company and how will  
56:29 the onboarding be completely revamped within  the whole product. But on my side product,  
56:35 I can just do whatever decisions I want and think  of the strategy and marketing and the messaging. 
56:39 And this is basically just getting me reps,  which is one of the most important things  
56:44 at the beginning of your career. So I understand  the fear that how do you outsource certain stuff  
56:51 and you're not owning 100% of everything, but I  think the upside is so much more valuable. And I  
56:58 think the only way that AI makes you worse  at your job is if you're using it wrong.
57:03 Is there anything that you've learned about  reducing the sloppiness of the output,  
57:08 just like a tip for keeping the quality  high of the stuff that it produces?
57:13 Similar to people, setting up AI for success,  for the task at hand. So if I just brought in  
57:23 a junior to write a deck or something and I  didn't give it any guideline, I just said, "Give  
57:30 a strategy deck." He would probably just go online  and find top strategy deck and just reproduce  
57:36 that, which is basically what AI is doing.  It's basically just fed all of the internet. 
57:40 So instead of that, guiding it and giving it  context on what your style of writing is and  
57:48 what you're trying to solve and all these  different things, I think that's probably  
57:51 one of the biggest unlocks. So that's just a quick  tip. And also Cursor has a /command called deslop,  
57:58 which is basically going back over the code.  I don't know if this is integrated into the  
58:03 product yet, but it's on Twitter. Their  founders have been talking about this,  
58:07 so that's definitely something I would run after  just to make sure that no slop is left behind.
58:12 That is so funny, deslop. Okay. One more question,  
58:17 which may lead to something else, but kind  of going in a whole different direction.  
58:21 You used AI to help you actually interview for  the job that you got at Meta. Talk about how  
58:28 you did that, because a lot of people right now  are struggling to find a job reading about all  
58:32 these people using AI to help them interview. You  actually did it. What did you use? What worked?
58:37 I feel like the analogy here is I have 12 nieces  and nephews and you can see how people who have  
58:44 grown up in a different world, how their mind is  formed differently. So if you ask me, how do you  
58:50 answer a phone, I'll do this. But a child now,  when you say, "How do you answer the phone?"  
58:55 They'll do this. They'll do the iPhone answer. And  I feel like people who are growing up now in their  
59:04 professional lives were the same just with AI. So every time I'm faced with a new challenge  
59:08 or problem, I think AI first how to solve  it. So Meta reached out and said they'd  
59:14 like me to interview. Straight away, I opened  up a project within Claude. I started looking  
59:19 online for all the best information out there,  things that I resonated with. I took a ton of  
59:24 frameworks and stuff from Ben Erez who has  written a guest post for you, who I think  
59:30 is one of the best minds out there right now. And basically I created a project which was my  
59:35 coach, which I would come and consult what to  do at each phase. I would mock interview with,  
59:42 and this was amazing. Also, I created  a game in Base44, which helped me ... I  
59:48 was really struggling with segmentation  within the product questions, so thinking  
59:51 of the correct segments. So I basically just  created a quiz game, which creates questions  
59:56 and different segmentations and I have to choose. So this was like, I spun this up. It's a web app  
60:02 that I would play sometimes when I was on the  bus to work. So basically, I think Ben talks  
60:08 about this a bunch, so I don't just go read Ben's  stuff, but just creating a project and feeding it  
60:14 with all the best information on the internet and  then mocking a bunch. I will say that the biggest  
60:19 game changer for me was doing human mocks. So cold outreaching to people on LinkedIn and  
60:26 having them do actual mocks for me, I think that  at the end of the day, especially for the Meta PM  
60:32 prep, which is super competitive and difficult,  I think there's no way to get around that.
60:38 That is so cool they use that post. I wasn't  aware. We're going to link to it. And in that  
60:42 post, Ben shares all these prompts you can feed  ChatGPT to help you prepare for interviews,  
60:48 do mocks online. It's a really important point  to say that those take you to a point, but it's  
60:53 actually better to use humans. I actually have a  post coming out soon in collaboration with Noam  
60:59 Segal about how everyone is using AI to interview. And one of the most interesting ways I've heard  
61:05 people and that we've found in this research was  that people use it to get feedback. They record  
61:08 the interview and then it gives them feedback.  Here's where you could have done better.  
61:12 Here's what you missed because the feedback  loop is so missing. No one ever tells you,  
61:16 here's what you did badly in this interview.  No one tells you that, and AI can do that.
61:21 So I'll add two things to that. One, which  is exactly this. So I'll mock with AI. Also,  
61:27 I did something really cool where there's  a question bank online free by Louis Lynn,  
61:33 which basically is an always updating bank  of questions that people are asked in real  
61:37 interviews. And I basically used Comet, which  is the Perplexity's browser. And I had the  
61:44 agent run all kinds of analyses on what the most  asked questions are. And that's how I knew how  
61:49 to prioritize what questions I would mock. And then at the end of these mocks, I would  
61:56 tell Claude within the project, "You're my coach  and I don't want you to make me feel good. I want  
62:02 you to make me as ready as possible for these  interviews. So give me feedback, like you said."  
62:07 And the other thing that I did was really cool  was some questions where I didn't have time to  
62:11 mock. I would ask Claude to play the candidate,  and then it would just give me a really good  
62:16 answer. And I could also learn from that, like  learning from someone who does a perfect answer.
62:20 Oh, man. I really love the way you phrased  it, that people kind of in your generation,  
62:24 the default is, "I have something I need  to do. Let's go to AI immediately and help  
62:30 me prepare for this thing, help me figure it out." Yeah. And this comes back to this quote that I always  think about, which I think everyone is always  
62:36 hearing, but it's such an important quote that  it's not that you will be replaced by AI at  
62:42 least for a long time. It's you'll be replaced  by someone who's better at using AI than you.
62:47 I agree. And that's what these conversations are for  to help people keep up with all that and to  
62:51 learn some of these skills. And again, see  where the future is going and start to learn  
62:56 how to get there yourself. Okay. Zevi, before  we get to our very exciting lightning round,  
63:01 I'm going to take us to a recurring segment  on this podcast I call Failure Corner. 
63:05 And why I love this segment is people  come on this, just even this conversation,  
63:09 it's like all these amazing things  you figured out, everything is going  
63:12 so well. People rarely hear the things that  don't go well, and those are often the most  
63:17 interesting and impactful stories. So  the question is just, what's the story  
63:22 of a time you failed in your career and  what did you learn from that experience?
63:26 Yeah, I love this. I love this. I love Failure  Corner, big fan. So I'll tell a story about  
63:34 when I started at Wix. So basically I started  within Wix as a student program and straight  
63:38 out of the student program, you get put into a  certain team. So I was in the editor, which is  
63:45 the core product of Wix. And the other  PMs were just the best PMs almost at Wix. 
63:51 There's four other people had much more experience  than me and they were ridiculously good. And I  
63:56 remember coming in and thinking like, my first  product review, I'm going to blow these people's  
64:00 socks off. They're not going to believe how  good of a PM I am. And I basically didn't  
64:05 really share what I was thinking. I would  work tons of hours alone and I was like,  
64:11 "I'm going to kill this product review. They're  going to be so impressed." And I ended up failing  
64:16 miserably. My product review was not good. It wasn't the format they expected. They  
64:21 had a ton of questions that I missed and  I felt awful when it was over. I was like,  
64:26 "Ah, you're such an idiot." And I saw that  everyone was like, "All right, cool. Yeah,  
64:29 just come back in two weeks and we'll keep getting  at this." And I understood in that moment that  
64:36 they had zero expectation of me being a 10X PM,  but the expectation of me was being a 10X learner. 
64:42 And the second I understood that, my whole mindset  shifted. And I think this is probably the best tip  
64:48 that I give now to junior PMs is basically be the  best learner you can be at the beginning. No one  
64:55 expects you to know all the answers and no one  expects you to be good. So basically what I did  
64:59 was I took each person on the PM team, there  was four other PMs and I assessed what their  
65:05 strength is and used them as a mentor for that. So Neri who's still my mentor till today,  
65:11 he has the best product sense of  anyone I've met. Oya is super,  
65:14 she's like a methodology expert. She just thinks  in frameworks. Yahra, who is the head of product,  
65:21 basically can look at a product and then instantly  understand the third and fourth order effects  
65:25 of them, the system thinking. So every time I  had an issue with one of these areas, I would  
65:29 come to one of them and consult them, and this  does two things. First of all, I learned a ton. 
65:35 And the second thing is that when the  next time, the next product review,  
65:39 my success felt to them like their success  because it wasn't this kid who's trying  
65:44 to show us up how cool he is. It was like our  mentee kind of making us all proud. And it was  
65:51 such a great shift for me. And basically at the  end of the day, I really excelled through this.
65:57 That is an awesome story. And this idea of  learning is such a good thread throughout  
66:02 this whole conversation that AI is good at  getting stuff done, but it's also really  
66:06 good at helping you learn how to do the thing  and to be this partner, this thought partner,  
66:11 the way you talked about the interview process  you went through and this learning opportunity  
66:16 /command. So awesome. Great story. Zevi, okay.  Before we get to our very exciting lightning  
66:22 round, is there anything else that you wanted to  share? Anything you want to leave listeners with?
66:25 Yeah. So kind of to tie back into the first  thing I said where if people walk away thinking,  
66:33 "Zevi's so cool." Then I've failed here. I  think that it's just the best time to be alive,  
66:40 I think. It's the best time to be a junior  contrary to what a lot of people are saying how  
66:46 there's no more junior roles out there and people  get out of school and you can't find a role. 
66:52 Yeah, that's true. But also when else in  history could you get out of school and  
66:55 just build a startup on your own with a couple  of friends completely bootstrapped. And I see  
67:02 more and more people towards the end of my  time at Wix, I was interviewing and I saw  
67:09 more and more people building their own stuff  with AI. And I think contrary to what a lot of  
67:15 people think, it's the best time to be a junior. It's the best time to be a learner. And I think  
67:19 if any listener is listening to this and you're  a curious person, you're a hardworking person,  
67:26 I want to say kind, I'm not sure, but if you're a  kind person and a good communicator, you have such  
67:31 an unfair advantage and you can give more value  to companies than most people who have 20 years of  
67:37 experience. So I really hope people get inspired  by this and start killing it with their projects.
67:44 Amazing. So many ways to be inspired  from this conversation. Zevi, with that,  
67:47 we've reached our very exciting lightning round.  I've got five questions for you. Are you ready?
67:52 Yep, let's do it. What are two or three books that you find  yourself recommending most to other people?
67:58 So I'll take one from each kind of genre. So in  fiction, I love The Fountainhead by Ayn Rand,  
68:06 one of my favorite books. Really makes you  think, really makes you feel. Business books,  
68:12 I'm a big fan of Shoe Dog, the Nike  story, one of my favorite books.
68:15 I just finished reading that. So funny. Amazing. This is great. It was great.
68:18 Yeah. I love Shoe Dog. And then more on  the psychology side Mindset by Carol Dweck,  
68:25 who coined the term growth mindset.  It's just such an amazing book. It  
68:30 kind of sounds like a self-help book, but  then you understand that it's completely  
68:34 psychological and is based on research  and that book completely changed my life. 
68:39 Really, I was always with a fixed mindset, and  then after reading that, I kind of understood  
68:44 that it was something holding me back.  And since then, I've been really,  
68:47 really trying to cultivate a growth mindset.  So I really recommend everyone reading that.
68:51 Again, connects to that thread of the way you  described it, being a 10X learner versus a 10X  
68:56 doer. Okay. Next question. Favorite recent  movie or TV show you have really enjoyed?
69:02 Yeah, actually, my wife is really into film, so  we watch a lot of TV. It's probably our favorite  
69:08 together time. I just finished watching The Pitt,  
69:11 which was amazing. It was really good.  And my first recommendation to everyone  
69:16 is if you haven't seen Severance run to  see Severance, one of my favorite shows.
69:21 Is there a favorite product that you have  recently discovered that you really love?
69:24 This is a good question. I'm always trying new  products. I'll always have three or four browsers  
69:30 installed on my computer and all this different  kind of stuff, and I recently discovered a Loom  
69:36 alternative. I was kind of disappointed with Loom.  They were taking so much money and the product,  
69:43 I don't know, I just didn't love it. And  there's an open source alternative called Cap,  
69:47 which is just really well-crafted. You can  see that the person was really sweating the  
69:53 details and it's just a really, really great  alternative. So I've been using that recently.
69:58 There's also a product called Supercut  that I love that's also a Loom alternative.  
70:02 Shout out. Okay. Two more questions. Do  you have a favorite life motto that you  
70:06 find yourself coming back to in work or in life? Yeah. I'm kind of between two right now. One,  which has become a Twitter meme basically,  
70:14 which is you can just do things. I feel  like that is basically going always in my  
70:19 head every time I do something that I'm  just shocked at the speed and ability to  
70:24 do things now so you can just do things, and  the second one I stole from my brother. His  
70:30 motto is nobody knows what the fuck they're  doing. And I just love that. And I think it  
70:34 kind of makes you take life more lightly. So  yeah, nobody knows what the they're doing.
70:39 I think people see these companies on the outside  and it feels like everything they've got all  
70:43 figured out. And if you're ever on the inside of  a company that's doing really well, you're like,  
70:47 how is this staying on the rails? How is this  still a thing that is working? Doesn't make any  
70:51 sense. It's all about to fall apart. Yeah. Okay. Last question. You've had a long entrepreneurial  
70:59 thread throughout your career. There's a  couple other real world businesses you've  
71:04 started in the past. You did a thermal  clothing business and then like a hummus  
71:08 delivery thing. So maybe pick one of those  and just tell the story of what that's about.
71:13 Yeah, I'd love to. Really fun that you asked about  this. So I'll tell the thermal clothing because  
71:19 I think it's really cool. So in high school,  I was selling thermal clothes in 10th grade  
71:24 for one of my sister's friends or something,  and basically it was just packs of thermal  
71:30 clothing, shirt and pants. I grew up in  Jerusalem, so it's a bit chillier there. 
71:34 So it was perfect for the weather. And in 10th  grade, when I was selling them, they were like 20,  
71:40 $25 a piece and I was making like $4 a sale,  and if you look in the food chain, I was like  
71:47 sixth or seventh down the line. So this was like  crazy margins. So during the summer I thought  
71:52 about it like I should just go straight to the  importer. So throughout the summer I called the  
71:57 importer and at first he was really, really mad. He was like, no, you have to work for me for years  
72:02 to get to this state. And I said, listen, man,  I'm finishing school soon. This is not going to be  
72:06 my career. Either do it or not. And we basically  negotiated throughout the whole summer. And this  
72:13 was also like how I did things before ChatGPT.  So he would throw out something, he'd say, "Oh,  
72:20 the import tax has gone up." And I'll just search  Google, like Import Tax Israel and start reading. 
72:25 And I'll be on the phone with him and I'll be  like, "Hey, I would just basically stall." And  
72:29 then I'd somehow come back with a challenge.  And I ended up getting a really great price,  
72:36 like 12 and a half dollars a piece. So I was  making 100% profit and I spread throughout a  
72:42 bunch of different schools. Each school, I had the  coolest people in school selling for me. And then  
72:49 a really fun thing that I did was we had a really  awesome basketball team and our basketball team  
72:55 would basically be 30 points up within the first  half and it kind of got boring for the crowd. 
73:01 So I wrote a song, like a basketball  chant about Thermal Clothes that  
73:05 basically has my number within it. And  the end of it was if you join in now,  
73:11 we'll give you a discount. And it was with drums  and everything. And still when I go to Jerusalem,  
73:16 I know some people who I don't even know  my number by heart because they know it  
73:20 by the tune. And sometimes when I walk in  Jerusalem, people stop me and say like,  
73:24 "Hey, it's Thermal Zevi." So that was  just a really cool experience as a kid.
73:29 This explains so much just the marketing  genius of that move. Oh man. Okay. Zevi,  
73:36 this was incredible. Two final questions.  Where can folks find you if they want to reach  
73:40 out and maybe follow up on some of the stuff?  We'll link to the scripts and prompts and all  
73:44 that in the show notes so you don't have to read  that, and then how can listeners be useful to you?
73:50 Awesome. So I've been helped throughout my  whole career a ton, so I love helping any  
73:56 way I can. So reach out on LinkedIn or on X.  I'd really love to help whoever I can. How can  
74:03 listeners be useful to me? So if you're a student,  try StudyMate, tell me what you think. If you're  
74:09 in Israel and you are not using dictation  yet, try Dibur2text. Tell me what you think.
74:14 Amazing. I just love how much you're giving  away and how useful that's going to be to so  
74:18 many people. So again, we'll link to that in  the show notes. Zevi, you're awesome. Thank  
74:22 you so much for being here. Thank you so much for  sharing so much. This is going to help, I think,  
74:25 a lot of people and I think it's going to  help people get over the hump on. Okay,  
74:28 I see all these people doing cool stuff.  Here's how I can actually do this stuff. So  
74:33 thank you so much for being  here and for sharing so much.
74:35 Thank you for having me. And  if you build something cool  
74:38 with some stuff that I learned here,  hit me up, send me. I'd love to see.
74:43 Amazing. Zevi, thank you so much for being here. Thank you. Bye everyone. Thank you so much for listening. If  you found this valuable, you can subscribe to the  
74:51 show on Apple Podcast, Spotify, or your favorite  podcast app. Also, please consider giving us a  
74:57 rating or leaving a review as that really helps  other listeners find the podcast. You can find  
75:02 all past episodes or learn more about the show at  lennyspodcast.com. See you in the next episode.
$

How a Meta PM ships products without ever writing code | Zevi Arnovitz

@LennysPodcast 1:15:12 16 chapters
[product development and MVP][AI agents and automation][developer tools and coding][productivity and workflows]
// chapters
// description

Zevi Arnovitz is a product manager at Meta with no technical background who has figured out how to build and ship real products using AI. His engineering team at Meta asks him to teach them how he does what he does. In this episode, Zevi breaks down his complete AI workflow that allows non-technical people to build sophisticated products with Cursor. *We discuss:* 1. The complete AI workflow that lets non-technical people build real products in Cursor 2. How to use multiple AI models for differ

now: 0:00
// tags
[product development and MVP][AI agents and automation][developer tools and coding][productivity and workflows]