Brian Lovin - How to level up with AI as a designer

00:00:00I think that's probably a good place tobe right now as a designer is reallyunderstanding what's possible, trying topush the edges even if it sucks and youget stuck. [music] From that moment on,it's just changed the whole way youwork. Like as soon as you realize youcan't design half of this stuff inFigma, what you're really designing isthe harness for the agent to do longerthings and verify its own work. In thesame way, it would be crazy for acompany that did issue tracking or notetaking to not take a hard look atthemselves in the mirror and try andfigure out their role in the future. Itwould be crazy for designers today tonot be doing the same thing. Ourobsession with titles is what will screwpeople over. If you're like, "But am I adesigner or a PM or an engineer?" Like,"Oh my god, like which path?" No, no,no, no. These things are going away.Like all this stuff is getting very,very blurry. It's getting very, veryeasy to move between those disciplines.You want to be the kind of person thatcan move [music] between those types ofworking very fluidly and not get caughtup in,>> [music]>> "But what's my What's the box I'm in atthis company?" Right? Welcome to DiveClub. My name is Rid and this is wheredesigners never stop learning. Today'sepisode is with Brian

00:01:15Lovin, who's oneof those people that I pay very closeattention to both how he works and allof the tools that he's using. So, we'reabout to do a deep dive into his designprocess and [music] all of the ways thathe's iterating on both how he designs AIand also how he uses AI tools. Butbefore we get into all of that, I wantedto know what it was like when Brianfirst joined the design team at Notion.One of the things that ended up workingreally well is before I joined Notion,they invited me to their design offsite.Like basically when we wereworking on the deal to join Notion, Idon't even think we'd closed yet or wewere close to closing. They had theirdesign offsite in New York. I was in NewYork.So, I just went and hung out with theteam for a week and that was helpful. Ithink it was probably awkward foreveryone cuz they're like, "Who the hellis this guy that doesn't work at thiscompany that's just showing up? Likewhat are we allowed to talk about?" Allthis kind of stuff. But it was like,"Hey, Brian's going to join. I'm goingto join." And so, I just jammed. Andlike mostly I think I just sat back andlike absorbed and saw what everyone wasworking on.And one of my favorite sessions at theoffsite was I think it was the last daythey'd rented this co-working space andthey broke off into teams of like threeor four people and each team

00:02:30drew twopieces of paper where each piece ofpaper had like the name of a feature.And it was like, "Go design somethingthat combines these two features." Andit could beYou might get lucky and draw like AIplus chat, you know? [laughter]Or you could get unlucky and draw likeformulas plus permissions. And you'relike, "Oh my god, like [laughter] what?I don't know." And then everyone wentoff anddesigned for I don't know, it was likenot very long, like half an hour, anhour maybe. And then everyone presentstheir ideas.And the the quality of the presentationsblew my mind. I'm sitting there andpeople had like high-fidelityprototypes, like pulled together reallyimpressive Figma click-throughs,whatever, with animations. Everyone wasvery quick at pulling together likeproduction assets, so it felt veryrealistic.And some of the demos were everyone waslike, "Oh yeah, we should ship that. Ohyeah, we should ship that." Like reallycool. And so, that really was my firstexperience with the Notion design team.And it was just very impressive and Ifelt excited. Everyone was verywelcoming. I don't know. Like it wasjust good vibes all around. And so, thatwas in I think October '24. And then Ijoined January '25. I got thrown intothe deep end when I joined. The firstproject I worked onwas called

00:03:45internally we called it appbuilder. I don't think there's anysecrets here.And it was very complicated. It waslike, "Hey, Notion has all of theprimitives that we need for AI to buildany kind of tool that you might use atwork. We have databases and charts andviews andpages and AI chat and we have all thisstuff. Why not just take the next stepand let it write code and like generateany sort of of app that you might wantto live inside of Notion." And that wasjust like a really intense project forthe first four months and it was funnybecause I actually didn't really workthat closely with any I I worked closelywith one designer who ended up leaving>> [laughter]>> a few months in.Um he'll know who he is. Um and that wasthat was an awesome partnership.Otherwise, I kind of felt a little bitisolated from the team. We'd nevershipped app builder in that form, but wedid split it apart and ship it inchunks. So then, the first big thing Iworked on was Notion agent, which waslike one slice of that. And then theteam shipped custom agents and then weshipped now workers where you can likewrite code and deploy and host code onNotion that AI agents can use. And so,it's cool like now over this arc seeingthat initial thing I was working on. Ifeel like all the parts have been builtand they're working inside

00:05:00of Notion. Idon't know if that exactly answered yourquestion, but like I just had a reallyfun onboarding. It was like this wasgoing to be the place for me. Veryintense first few months where actuallythe first thing didn't really work.>> [laughter]>> And I had to like pivot and adjustand like learn on the fly. And then herewe are a year later and I don't know,I'm just [clears throat] nothing haschanged. Like the team is is awesome andI'm very excited about Notion's place inthe future. I think everyone here islike sufficiently AI pilled and that'sfun. Real quick message and then we canjump back into it. So, Jitter justreleased image to video and it's apretty big deal. All you have to do isupload an asset and Jitter willinstantly generate a short video clipfrom it with the help of AI. It's aneasy way to add motion and depth tostatic backgrounds, maybe create littleclips for brand visuals or just ingeneral make any scene feel more alivewith subtle movement. It's super fun toplay with and available today. Just headto dive.club/jitterto try it out.If you're like me, then you'reprototyping a lot in code lately. Butthe problem is you kind of have thischoice between a tool that isn't hookedup to your actual code base and designsystem or a local host prototype that'ssuper annoying to share. That's why Ilove what Dessen is doing. In one click,you and everyone on your team canprototype directly

00:06:16in your code basewithout ever opening an IDE. Dessenextracts your design language and givesyou the perfect sandbox to explore[music] without any of the technicalhurdles. And when you're ready, there'sa nice little share button top right andyou can send it to anybody on your team.It's a pretty big deal and you canconnect your code base and startprototyping today. Just head todive.club/dessen.That's d e s s n. [music]Now on to the episode. What's it likebeing a designer whose surface areapretty much entirely revolves aroundworking within shaping AI as a material?So, when I joined in January, I'd nevertouched AI. In fact, when I joinedNotion, I was pretty skeptical of AI.Like I remember when we were at the endof working on Campsite, which might wasthe startup before. At the very end,cursor tab completion was getting good.Just to like put a frame of reference,right? Like we didn't have codingagents. I think this was pre-Claude 3.5models and I just thought it was allkind of I'm like, "Nothing itmakes is good." And the people whowere paying attention well weren'tconcerned about the current quality.They were concerned about the trajectoryand how fast it was getting better. So,I joined that January and one of thereasons app builder was actually hardand probably I just made a millionmistakes.

00:07:31But one of the reasons was Ikind of approached it like a typicaldesign project. Like,"Cool, we're going to build thisinterface where users are going to beable to like construct experiences withAI and I'm going to open Figma and I'mgoing to like design this in Figma."And I was designingchat experiences where it's like, "Okay,the user's going to ask for an app thatdoes this." And then the agent's goingto respond and it's going to show alittle preview of what it's making andit's going to be perfect. And then wewould test that out in in in engineersthat actually build it and it justdoesn't work that way. The modelsweren't that good. It was very slow.They would make mistakes. They wouldhave to like ask clarifying questions.>> [snorts]>> And so, there was this moment probablyin like February, maybe I'd been a monthin, where it was like, "None of this isworking. Like I can't keep making Figmamocks to try and simulate what it'sgoing to actually be like to use AI tobuild AI type things." So, that's whenlike, "All right, I just got to jumpinto the medium." And so, that was whenI started working on this internal tool,prototype playground. And really thegoal there was like, "I just need amedium where I can talk to AI and likeunderstand what it feels like to have itrespond. I'm going to test all themodels.Some of the AI SDKs, like from Vercel,were pretty good at helping you,whatever, execute tools, like returnstructured outputs. You could start to

00:08:46sort of simulate what it would be liketo have a Notion-shaped thing,but just in this totally isolated codebase." From that moment on, it's justchanged the whole way you work. Like assoon as you realize you can't designhalf of this stuff in Figma, what you'rereally designing isthe harness for the agent to do longerthings and verify its own work. And alot of that meansusing AI in your day-to-day work andjust getting like a feel.And then the other is talking to theengineers who are actually building theharness and being like,"Why is this slow? Why did this getworse? Why is this model better than theother model?" And luckily for us, wehave really, really good engineersworking on our harness and I just tryand learn from them and absorb theirbrains. And they're like all, you know,really on the next generation of modelsand I just try and absorb that into mydesign process. I think where designersneed to be designing right now is likeright at the the boundary of the nextthe current model, like hoping that thenext one solves it, right? Like we'reall shipping shitty vibe-coded slop justhoping like, "Ah, the next model'sdefinitely going to clean this up,right?" And I think that's probably agood a good place to beright now as a designer is reallyunderstanding what's possible,trying to push the edges even if itsucks

00:10:02and you get stuck and then justtalking to a lot of as many smart peopleas you can. Obviously, like working at aplace like Notion is very advantageousjust cuz there's a density of people andwe have have in-person days to learnfrom each other. We did that with thevoice models working on Inflight. Wespent like a good month trying all ofthese different interaction patterns forwhat it looks like to almost beinterviewed as a way to give feedbackand it was clear that it could be good.Yeah.>> most definitely was not good and it waskind of just like, "All right, we'regoing to put a pin in that. There'ssomething there, but like I guess we'lljust wait till the models get better."And it kind of feels like that line ofthinking exists in almost everydirection at the moment. If you are, youknow, especially pushing, you know,doing more frontier work. The people whoare really engaging with these toolsare, I think, by definition earlyadopters. Like there's the masses whoare using ChatGPT to, you know, ask likesome fairly basic questions, but ifyou're really trying to use AI right nowto automate knowledge work or simplifyyour job or like remove a process atyour company, you're a very earlyadopter and the tolerance for thosepeople for like kind of sub-parexperiences is actually quite high. Ithink people are still willing to crawlover glass to set up an open claw,right? Like that is not an easy

00:11:17thingand people are doing it. And so, I thinkthat's actually okay and and it has beenweird cuz you have to like readjust yourexpectations for your own quality bar.Where it's like, "I know that this thingisn't perfect, but the people who areusing it are very tolerant to thoseimperfections and are willing to liketrial and error their own prompts overand over again." One of the importantthings that I got to see at Notion and Ithink probably every designer everycompany is experiencing right now isevery 6 months everything that we didbefore becomes more or less irrelevant.So, like at Notion we've rewritten theagent harness, I think, three timessince I joined. Like roughly every 6months. Each time it's just strippingaway all these old assumptions that weused to have and like today we'redesigning arounduh harnesses that are roughly like giveyour agent the ability to write scriptsand search for things. Like that'spretty much all you need to write a goodharness right now. But I'm kind ofoperating under the assumption that thatwon't be true in 6 months. I don't know.Like are skills going to be useful in 6months? Are we going to need all of thislike You are prompt engineer. You are asenior engineer at Yeah, exactly. I Idon't knowif we need all of that stuff in 6months. So, I have founda little bit of like inner peace

00:12:32withit's just going to keep changing and myjob is to like understand what'shappening, why it's happening, like whyis it good at writing code and what arethe implications of that? Okay, well, itneeds like a safe space to execute thatcode. Okay, you look around, everyone'sspinning up sandboxes. Sandboxes areslow, let's make them faster, right?Like all these things are justdownstream consequences of the agentharness is changing and they've changedevery 6 months for the last couple ofyears and I assume that will continue.Clarifying question on the prototypingplayground that you built. I've seen alittle bit, but help me figure out whereit exists on this spectrum. Like on oneend it's like fresh, lovable build, nocontext. On the other it is a perfectcopy that you're like, you know,updating somewhat recently of the codebase. Likewhere does that exist? Like cuz I thinkthat level of fidelity is something thatI don't see that often. I'm assumingit's somewhere in the middle. Talk to meabout that for a second. It depends onthe prototype. It's just a big directoryof prototypes and then we have acomponent kit that kind of it's likeclose to Notion. It could be perfectlike if we just spent more time on it,but we've we've tried to do like the80/20 thing. Like get it pretty close.Like we have a sidebar that looks kindof like Notion sidebar, buttons anddrop-downs that are kind

00:13:47of like Notionbuttons and drop-downs. And then it's upto each individual designer to push thatas far as it needs to be pushed to provethe point of their prototypes. That'swhy I say it depends. So, for example,we have a designer here uh named WillDawson and he's been working on theeditor and like some of theinline stuff that you can do with AIwhile you're working on a Notion page.The only way to really feel what it'sgoing to be like [clears throat] tointeract with AI while you're typing apage and selecting text and likenavigating with your keyboard is toactually be in an editor. And so, inprototype playground he more or lessrecreated like a very simplified versionof the Notion editor with slash commandsand inline AI and the AI does actualthings and actually writes back andupdates the page in real time and selectstuff and it's very high fidelity and ifyou stare at it you're like, "I thinkthis is Notion." And it's not. Like thatyou have to push to really prove thatthis idea is good. So, anyways, all toanswer your question, it's a wide range.We try and give some components to letpeople get close enough,but then each individual designer canpush as far as they want. Now, over timeI don't know what's going to happen. Ithink two things will happen. One iswe'll continue improving thosecomponents inside of playground so thatjust by default they look better andbetter

00:15:02over time, but I've also seenmore and more designers just skippingplayground and they just go and iterateand design stuff in the production codebase. It's not appropriate for everytype of thing you're working on, butthere are an increasing number of thingswhere it's appropriate to just, I don'tknow, boot up the Notion code base,throw Claude at the exact same problemand you might as well get like a a realthing out the other side if you're goingto spend roughly the same amount of timeanyways. How much of the Notion designteam is actuallytouching prod, working in code at all?Is it still acceptable to be doingpredominantly Figma work? Talk to usI think it's totally fine depending onwhat you're working on, but I would saythe barrier to entry to try makingsomething in code is so low that I Ithink everybody has dabbled with it.>> What percentage of people go straight toprod? I'm not sure. Maybe like 10 to 20%and it's still task task-specific.There's still a lot of work that happensinside of Figma. There's a lot of peopleplaying with uh paper now. I don't know.I feel like I'm just surrounded by mypeople. They're all like we're alltinkering and trying the new things, butI think it's still about picking theright tool for the job and there aresome types of design work where it'sjust more helpful to be in a 2D canvas,taking screenshots and like masking overit and just like, "Ah, I think thislooks roughly right." And then the onthe other side, "No, I need the highestfidelity

00:16:18version of this in a productionenvironment." And we push theengineering team at Notion to make theNotion code base more legible to AI. So,like a lot of work that's happened underthe hood and I assume is happening atmost companies is like, "How do weenable this type of work from people whodon't have engineering backgrounds sothat the code that that they write withAI is not shitty, doesn't break theservice, is tested correctly, getsreviewed correctly." A lot of that islike code base infrastructure stuff,right? Everything from skill files tolike your actual CI pipeline to makesure that silly things aren't slippingthrough the cracks. As the companyfunnels more and more energy there, Ijust feel like we'll all keep climbingthis gradient towardstrying things in prod more often thannot. Um but there's always going to beexperimental confusing ideas where it'sjust way better to sketch on a napkin oron a whiteboard or open Figma and justdraw some rectangles. Actually, I'd saya lot of people now just open TLDraw andjust like draw squares. As you movealong this gradient, how does thatchange the way that you all collaborateand work together or, you know, piecesof the design process that werepreviously a staple that changed? Likehow much does the medium changecollaboration at Notion? I think it justmoves from, "Hey, check out this FigmaURL."

00:17:33to, "Hey, check out this deploypreview."I think that's pretty much it. Like whatwhat's the artifact that we're going topoint at together and try together andmoving that from one domain URL toanother, right? I think that's that'sbeen a big change. Prototype playgroundhas been cool because it's all just onecode base. You can like go and pokearound other people's stuff and likeyoink cool interactions from otherpeople's prototypes or just likeduplicate theirs and and riff on it.Like that type of interaction has beencool to see. But I don't know. I mean,we still just have design crits and weshow up in a meeting room and someonelike talks about the problem they'retrying to solve and maybe we share adifferent shape URL with each other.It's kind of an in-the-weeds question,but when I think about what I thought myworkflow was going to be like like 6months ago, like looking into the futurekind of thing, I assumed that I wouldprobably be doing a bunch of quickfunctional prototyping in code and thenreally sweating the details on thecanvas and fine-tuning all of thevisuals. And I found that it's likeactually the complete opposite.Where I'm doing a lot of like youmentioned paper, a lot of my go-wideexplorations are actually happening likethrough conductor in paper. But thenwhen I want to sweat the details, I'mdoing it all in code. Like I haven'tlike really specifically nudged pixelsaround on a canvas in in a lot ofmonths,

00:18:48actually. I haven't nudgedpixels in Figma in a long time. Thefrustrating thing is like AI is stillterrible at last mile fit and finish.And I think there's two ways to go aboutit. One is what you're describing, whichis you just get into code and like touchthe CSS cuz at the end of the day that'sthe thing that you're going to ship andthat's that's all that matters.And for that, I think AI sucks. Even allthe leading models, you can make themultra think and they're just bad at likevisual interaction design. Some peoplewatching will say skill issue, which Ican accept, but I don't know. I've beendoing this for a while and like I can'tget a hold on it to do good things.And so, so you do last mile inside of acoding tool, an IDE, or if you're reallycrazy and don't value your time, youlike prompt your way to a polishedthing. Like there's people promptinglike "Nudge that element four pixels tothe left." I'm like, "Don't Why why areyou spending money to nudge something incode? Just like go change the class,right?" Don't do that. So, that's oneway and then the other way is actuallyfunny enough, like having a higherfidelity system inside of Figma and justcrossing your fingers that the MCP toolsor paper like just cross your fingersthat the MCP tools are getting better.Figma's MCP, I think is actually gettingbetter. And the last time I used it, Ihad

00:20:03a design in Figma with named layersand all the components are roughly namedthe same as they are in the code base.And I don't know. I can like paste a URLof a Figma frame into Claude code and itspits out a pretty damn good thing onthe other side. That's cool.>> And so we we have like gone full circleinto this funny world where yeah, itturns out if you name your layers likethe model can kind of figure out whatyou're trying to do with the thing. Andso I I I see teams approaching that onboth sides like invest in the Figmathing, invest in whatever naming schemestructure there will make your designsmore legible to AI.And then on the other side, obviouslythat has to end up in in code at somepoint. And I think honestly that's justlike getting in there and touchingpixels and making buttons feel good orwhatever. And then once the button feelsgood, then the AI knows how to reuse it.So Yeah, I think that's actually thekey. Cuz like people talk about how AI'sreally bad at design. I'm like you'retotally right.But if you spend>> the time Yo, I missed that one, man.[laughter]Like if you if you spend the time reallyreally really sweating the details on agood set of primitives once, AI's sogood at not only reusing, but alsoextrapolating. Like I might be like,"Hey, I really like this tertiarybutton, but I'm using it on this card.The height's much bigger. It's not goingto be a noticeable effect. So I'm justlike

00:21:18take that general idea and justextrapolate it." Oh man, I've beengetting such good results from that. Theother thing that I'm curious if you'reusing at all, do you use agentation? Ihave played with it. I That's myantidote to the nudging. I think I thinkagentation is is awesome. Um but Ihaven't used it very much. I'm okaysaying nudge this four pixels to theleft if I can bundle that with like sixother prompts like that and then do itall at once. And then I'll just likespin up a new tab in conductor. So Imight have like five tabs in conductorall doing four or five different promptssimultaneously when I'm like doing likemore of the visual refinements. At leastI convince myself it's it's faster, butmaybe I'm also just I don't know. Idon't know I don't open the CSS anymore,I guess. I don't know. I don't know. Youknow, there's there's the the studieswhere it's like people self-report beingway more productive with AI, butobjectively on a measured basis we'reless productive. And the counter to thatargument isit's okay if you're like accomplishingthe same number of tasks in less time,but it seems like more people areaccomplishing more tasks at like a lowerquality in the same time. So we're inthis weird zone where it's like, "Yeah,you can do more. Each thing needs morefollow-up and fixes." And that makes youfeel more productive overall.

00:22:34I'm nottotally sure.But agentation, I mean the whole pointof that is and and tools like that,right? Like what's the other one thatJosh Puckett's making? Oh, Dial Kit.Dial Kit. Like all these things, right?Like we just need to make our intentlegible to AI. And your intent is or orthe fidelity of that intent is muchhigher if you can say "This Reactcomponent in this position with this IDshould look like this." instead ofhaving to like describe visuallysomething on the screen.Like of course the AI will be better atperforming operations when it knows thename of the thing or the underlyingcomponent or knows what file it's in. Soall of those tools, it's all about justmaking your your intent legible. Andthat's the kind of stuff where it'schanging all the time, right? Likeagentation came out like a month ago andthen now Chrome has all that stuff likebuilt into Chrome. Any agent can use anybrowser now. Like this stuff is changingso fast, but all in the pursuit of agentlegibility. On the topic of thingschanging so fast, can you walk usthrough some of the key evolutions ofyour tool stack over the last 3 6months? We're in this really funnytimeline where people are tribal aboutthe

00:23:49tools that they use. You have peoplewho use Claude code, you have people whouse Codex, you have people who useCursor.And everyone kind of thinks the otherpeople are dumb or they try and likejustify why their thing is better.And then you get into like very veryniche arguments about "Well, Claude isbetter at like creative planning andCodex is better at following specificspecific instructions." Like all thesethings might be true. So for me whatthat means is I just want to try all thestuff try not to be tribal. I Igenerally prefer the the Claudeecosystem, but you know, I had given uponWell, to to answer your question, I wasusing Cursor a lot last year. I stoppedusing Cursor in favor of Claude code inthe terminal, then switched from Claudecode in the terminal to Conductor.Conductor to me was like the perfectmiddle ground of like I can see thecode, but I can multitask and I'm mostlyjust chatting. But then you know, Cursorcomes out with Composer 2fast.And I'm like, "Okay, let's go try itout." And you open Composer 2 fast andit's pretty good and it's fast and I getto look at the code and I can tweak itat the same time like, "Oh, this is theperfect tool for me to do front-endpixel polishing type work."I don't want to be like

00:25:05sitting around waiting for Opus tochange the corner radius of things andlike then I have to switch from theterminal to an a browser and like Idon't want any of that. Like CursorCursor's actually a really good tool forthat. So right now I have no allegiance.I just pick the best thing for the job.And then if it's at work, I get tobenefit from unlimited free tokens. Andif it's personal, I just try andoptimize for tokens per second or someratio of intelligence, right? Like umand I pay for the Claude Maxsubscription, the $200 one. So I'm like,"Ah, I kind of want to get my my mymoney out of that thing." Um so I biastowards Claude in Conductor. Is theprimary value prop other than speed?Like I I get the the speed thing withCursor, but is that becoming your choicebecause you are actively getting in andwriting like handwriting like a cavemanCSS in theOkay. I've handwritten more code in thelast 2 weeks than I did in the last 4months. The current meta right now, ifyou pay if you're on Twitter a lot likeI am, is nobody's writing code anymore.What are you talking about? Liketouching code, that's so 2025. And I'mdoing that inLike I use agents to yolo PRs,especially in prototype playground. LikeI yolo stuff all the time and I don'tneed

00:26:20to look at the code. I think themodels are still bad enough that if youreally care about the final quality ofthe thing that you're shipping, I thinkyou got to look at code. And I justhappen to be working on some stuffrecently where I likereally care about that or that that partmatters, the final fit and finish.And so yeah, I'm opening an editor andwriting code like it's 2023 2024 and itfeels pretty good because now I get toreach over to the side and like grabthis really fast or smart model to justtake some of the the annoying parts offmy plate while I'm in the middle ofthat. I think it's a good time to nothave an allegiance and not get tribalabout tools. Just try it all. Just tryit all, see what works, and try andunderstand when people say like Codex isbetter at following instructions andCursor is better at better at creativeplanning, like what that actually means.Like it's one thing to hear that and belike, "Okay, cool." But then you go andtry it and you like try and reallyunderstand why you would pick adifferent model for a different type oftask. The better sense you get of whichtool's good at which job, I think thebetter you can be as a designer or asoftware creator. Cuz it'd it'd be like,I don't know, knowing when to reach fora frame versus a group or auto layoutversus free form moving of things. Likeyou just have a better mental model ofwhat to reach for to accomplish thething that's on your mind. There's onequestion that I can't stop asking

00:27:35myself. What if companies applied totalk to you rather than the other wayaround? And that question is thefoundation for the all-new Dive TalentNetwork. And it's working. Like rightnow, I'm helping many of the mostexciting startups that I know to hirethe designers and builders who listen tothis show. So if you're curious whatmight be out there and maybe you want toget on my list or maybe you're evenlooking for your next design hire, headto dive.club/talentto join today.I've started doing little tests inConductor just to compare Opus and Codex5 4 in different environments. And solike [music] maybe I'll have Opus createa plan and then I'll have Codex reviewthe plan and then make them go back andforth. I get that's been really helpful.And then I've also just, you know howthey have like the built-in reviewbutton in Conductor? Yeah. Yeah. I'vejust been I've been hitting it. Like Iliterally did this this morning. I hitit with Claude and then I went into thesettings and I changed it to Codex and Ihit it again. I was just like, "What'sYeah.>> What does the output look like?" I know.I do the same thing. Yeah. I think Ilike Codex for reviews more, actually.I'm like I have a few instances nowwhere I'm like, "I think it's actuallybetter. It's catching things." I do thesame thing. Because it catches things,like I think it's the exact same reasonyou would do a design review with morethan one person,

00:28:50right? Like everyone'sjust going to have a slightly differentperspective. The thing that I findreally frustrating about that experienceis that it's very inconsistent. Like I Ido the same thing. I will plan withOpus, have Codex review, and then afterthat I actually don't really care whichone implements. And then after that I'llhave one of the two review theimplementation.I actually get very inconsistent qualityof the review. Um for example, I'mworking on this side project, Shiori,and I was like, "I should like run aprompt and like do like a securitycheck."And I gave it to Opus or Codex, I Ican't remember, one of the two. And itwas like, "I found 10 like securityissues." And then I gave that thing tothe other model and the other model waslike, "This is security theater. Likeliterally none of these things matterfor your use case or your code base.This is all Like they'rethey're wasting your time." And then youcould give the same prompt and theywould say the opposite thing. I findthat isuh the pretty exhausting part of usingAI today is the inconsistency butfeeling like you have a grip on it likeoh if I just get my skills in the rightorder if I just review thingsin the right right flow and like havethis model do this. I think it's a lotof cope for the fact that like themodels are still inconsistent. Theystill

00:30:06generally like are bad atoutputting consistent code. They stillwrite really stupid that if you'renot reading the code and then 3 monthslater you're like I don't know why myapp doesn't work or it's really slow.It's like yeah cuz you've just been likeslop cannoning thinking that you'redoing this genius like two modelsreviewing each other. They don't knowyet. Who knows? In a year it might besolved. It's not right now I don'tthink. Taking a step back fromtools and getting super granular intoliterally thetypes of prompts and text and commandsor questions like what you are literallyusing language wise to interface withthe models. Are you able to see clearchanges or progressions in terms of likehow that part of your practice hasevolved over the last whatever 6 monthsof really trialing all this stuff? Ifyou think about what the AI is doing istaking some number of like upstreamcontext tokens that you give it andtrying to come up with the bestprediction of what follows, right? Thatmeans if you have really high qualityinput tokens, you're going to get higherquality output tokens but those inputtokens are where you get to steer whereyou want the thing to go. So forexample, I mean this sounds so basic butlike if you want it to pay attention toa certain thing,

00:31:21you have to say thecertain thing. Like you actually have tosay the thing that you want it to focuson. If you just tell the model, hey whatare some edge cases we've missed? Nowall of a sudden the words edge case,they're in the context and it's going tolikeyou know,interpret from that forward and be inthe mindset of looking for edge cases.If you understand how that worksyou start to do these weird they almostfeel like it's like astrology orhoroscope type of prompts where it'slike how could this be simpler orwhat are edge cases we might have missedor is this the most elegant way we couldhave done done this or I'm going to showyour work to my boss. Can you make surethat it's like really good? Like theseare>> that one.These are stupid these are stupidprompts but they work because it justflips the right bits so that the modelwill extrapolate from there. One of myfavorite ones is is from Simon who's theco-founder of Notion. It's uh I have itI have a snippet for it. It's um let'sstep back and think really hard. How canwe make this simpler and dumber whilestill achieving our goals?I think that's a good prompt and I runthat I don't know 20 times a day.Really?

00:32:36Wow. Yeah like after everything.Every plan like make it simpler anddumber dumber. Every time it fixes a bugis there a simpler dumber way to dothat? Every time I'm likeit makes a new button I'm like is therea simpler way to make that button? I betyou made it way too complicated. So thaton the language side is why anyone whoknew how to code before AI is having away better time using AI than someonewho didn't know how to code becauseguess what? They know the right words tosay. They know what goes in that thatupstream sequence of tokens. They knowhow to say things like I don't knowdurable workflow or background job orqueue parallelization like like thesethings that I think if you don't knowhow to program like you just don't havethe vocabulary for but if you've beenprogramming for a while you have thevocabulary and it feels very natural todescribe those kinds of systems. Thosepeople are obviously getting betteroutcomes. That's why like I have to bevery open-minded to everything that I'mbad at is skill issue because in theoryif you can describe your intent moreclearly in English with the right set ofwords that the model can extrapolatefrom, it will do the right thing intheory. Right. So I think a lot of timeswhere it does bad things I have to likeself-reflect like I think I wrote a badprompt or I wrote a

00:33:51lazy prompt or I wastired when I wrote that. That ischanging how I I relate, right? So likeI've stopped writing late night prompts.Like if I'm tired I should not ask it todo a thing cuz I'm going to ask it in alazytired Brian sort of way. Let's talkabout Shuri in a little bit. I've beenseeing some tweets things you'rebuilding. So maybe give us a little bitof context but then the lens that I'mparticularly interested in using is youknow at least for myself anytime thatI'm doing something from a blank canvasand you know just no constraints I cando anything I want it's an opportunityto try new things or to tinker like dothings like me a little bit differently.So like given the context of what it is,how are you then taking advantage of theblank canvas this time around? Yeah it'svery funny and almost embarrassing cuzit's like we can make anything. AI canmake anything and I'm like I'm going tomake a bookmarking tool.>> [laughter]>> It's like it's like weather apps stockmarket graph chart to do and findObsidian replacement.>> [laughter]>> Yeah.Yeah. I I started working on Shuri Iguess a year a year ago just to solve aproblem that I have which is like nomatter what system I have for read itlater app, I just never go back

00:35:06and readthe things later. I haven't quitefigured out why that is but maybe it'sthis idea of ownership like having madethe thing makes me feel more connectedto it or want to like get juice out ofit, you know? Um so I made my own littleread it later app and basically all itdid was sync stuff into to Notion andthen I used my Notion as the like listof stuff. I was I was basically Notionmaxing all of last year's like how do Iput my whole life on on Notion?And that actually worked. Um it wasworking pretty welluh but it's kind of hard to distributethat as an app. It's a Notion databaseand I have some like functions runningin the cloud over here but I don't knowhow to distribute that. Having things inNotion is great except it's sort offighting the rest of Notion forattention. Like Notion has all my otherstuffbut sometimes when I want to be in likea I want to sit down on the weekend andread for half an hour, I just want avery simple focused experience thatshows me what's what stuff on my readinglist. So that's when I started workingon Shuri which is like the standaloneversion of that. The cool thing about AIis you can stand up a product like Shuriin 2 seconds. Like literally the dumbestprompt will get you therebecause there's so much training databecause so many people have madebookmarking and read it later apps.>> [snorts]>> The interesting part for me has beenlike figuring out how deep the iceberggoes. Like on its surface

00:36:21it's a list oflinks. It's not complicated but underthe hood and especially when you haveusers saving links, you realize peopledo weird things with their bookmarks.People have saved links that redirect tolocalhost. Like how is this possible?Who like somebody somewhere set up aredirect for a live domain on the web onthe internet that redirects to alocalhost route.Usually those people are trying to hackyou so you have to like consider that.I've had people import like 10,000Wikipedia articles.Okay, 10,000 Wikipedia articles like howhow should we process that 10,000Wikipedia article import happeningsimultaneously? Oh I need to like golearn about distribution of systems andparallelization and what happens if oneof those fails? Do we retry it?What happens if one of those Wikipediaarticles is 30,000 words long anddoesn't fit into a column of a database,right? Uh so you get to like hit allthese little edge cases and so now whatI'm having a lot of fun doing with Shuriisjust going really deep handling allthese edge cases and at each step tryingto write the test or the verificationframework or add the logging necessaryso that AI knows about

00:37:36all these edgecases and can design around them in thefuture. So this is very basic but I'mI'm kind of proud of this like the flownowis when someone reports a bug on Shuriwhich yes bugs still happen. All I haveto do is take their email addressand paste it into Claude or their userID whatever it is. I paste it intoClaude and I have Claude hooked up toSentry for error reporting super base soit can like figure out information abouttheir accountand I use Axiom as a log drainand I can just paste their user ID andpaste the content of their emailand it will go and replay exactly whathappened leading up to the moment offailure because every single event islike logged and tracked and all theexceptions and stuff. It knows whatdevice they're on all this kind of thiskind of thing. So bug reporting and bugfixing has become quite fast. So to methat is the the fun part now. I feellike Shuri, you know the rock tumbleranalogy from Steve Jobs? You put allthese rough rocks in the thing and youtumble it around and if you tumble itenough what comes out is like a very setof polished stones. I'm having funfiguring that part out like what is thetumbler that lets me get a polishedstone out the other side? Something thatis resistant to edge cases but if anedge case does come

00:38:51up I'm able to findit very quickly and easily or the AIfinds it very quickly and easily on mybehalf and just fixes it and then addsthat so it's like a self-healingself-reinforcing system. Not all of it'sautomated yet. In fact very little ofit's automated but that's sort of thelong-term goal here is like can I putmore of this on rails so that I can justfocus on building cool and fun newfeaturesand the like infrastructure, thescaling, the edge case handling, thebugs, all that stuff is a self-healingloop. The cool part for me about workingwith AI now is I get to stretch justbeyond what I would have beencomfortable doingas an engineer writing code before AI. Idon't know. I want to build an API forShuri. Like2024 version of me as a programmer wouldhave been like ah I don't really knowwhere to even start. Like how would Ibuild a CLI? How would I build an API?How would I do rate limiting, how wouldI do abuse detection, how would I doall these kinds of things. I don't know.And now all of a sudden that type ofwork is very accessible. I can justguide the the agent in that direction.That has been cool for Shuri. Like API,CLI, MCP, just pushing a little bitbeyond my comfort zone. And I'm probablynot pushing enough. As a designer whohas been writing code for a while, it'shard to shake out

00:40:07of this mindset oflike, oh, but how would I build this?And get into this new way of working,which ishow would I describe my end vision orthe end goal in a clear enough way thatthe AI can do it? It honestly made me goback to what you were talking aboutearlier in terms of like productivityand speed and everything. And I I thinkI actually agree. It's weird. Like thedominant measuring stick that we use forAI is so often speed and efficiency, youknow?But I'm like, I don't really care aboutthat that much. Like that's not theimpact that it's had for me. It's notefficiency. It's likegiven me the ability to play in a set ofadjacent ballparks that I previously hadno business even walking into the doorfor, you know? So it's like it's likemore about like the raising the ceilingand like increasing what I can bring tothe table. I'm not necessarily doingthat much stuff faster, but I'm doingmany more different types of things thatwere previously inaccessible. Shuri haskind of become a little bit of aantidote for me cuz I spend my week atwork like do more faster, write morecode, build more prototypes. And then Iget to sit down with Shuri and thinklike, what needs to be true for this tostay interesting and stay a project thatI would work on for the next severalyears. Like what

00:41:22what has to be true?And like part of that isI have to understand the code base. Ihave to care about who's using it andthat they're taking care of. I have tocare about good infrastructure. Andthose things actually force you to slowdown. I quite enjoy that. And what AI isallowing me to do is slow down and focuson the hard parts. And let the computertake care of the rest. So for example,my current workflow right now is duringthe weeks I plan and during the weekendsI execute. So a typical day might be Iwake up, check and see if there are anyerrors overnight or bugs or userfeedback. And then I just have a simpleskill that's like fix issues and I canjust paste that into Claude and let itgo and fix all the sentry issues or goand debug things.Cool. That's off my plate. Now I can golike talk to AI. And like, I think Iwant to do this feature next, but I'mnot quite sure how it's going to work.Here's roughly the intent or the theuser goal. Submit. And then I go towork. And then I come back at night andI read the plan.And I'll chat with it for a little bit.And then maybe like do one more round,submit, and then I go to bed. Okay, sothat's like one day. And then by likeFriday, Saturday morning, you've gotlike a pretty good plan. Like here'sOkay, here's how I think I'm going totackle this thing. And then Saturday allthe bullshit's off my plate. There's notlike a pile of bugs to worry about.

00:42:37I just get to go and look like reallystudy that plan and then get to workbuilding it. I think that works for sideproject. But I also like trying to beintentional about pacing myself. Iscrambled a little bit like to get itout the door as you have to do to likeget all the features in place. But now Iwant to get into like methodicalrhythm, planning, shipping, polishingthe stones, handling the edge cases. Andjust over time like maybe start slicingoff little bits that can be automated.That's the interesting part. Coming upon time here, I want to take a hard leftbecause there's a fewpotentially unrelated things that Ihonestly just kind of wanted to hearabout. And one of them is I'm hopingthat you can heal some of my FOMO fromTokyo Design Conference because you yougave a talk there. And obviously wecan't go do the whole thing, but whendesigners left that room after you weredone, what was like the core thing thatyou were hoping that they took away withthem? The thing I wanted people to takeaway from that talk is that AI is notmagic. The underlying systems are verycomplicated. I don't think anybody wouldclaim they know what an LLM's doingunder the hood. But it is doing thingson your computer like writing files andwriting scripts and like invoking bashcommands that were invented in the 70s.

00:43:52And those are knowable things. Andthere's two ways you can sort ofnavigate that world. One is you can notpay attention to what AI is doing onyour computer, the kind of code it'swriting, how it writes the code, how itnavigates your code base. You can justnot pay attention and let the model takethe wheel. Or you can be curious aboutit. Like what does that mean? Like wait,what is grep? What's grep?And what's grep in relation to? Likesed? Sed, what's that? And you're goingto find out about awk. And you're goingto be like, what the Like some guyin the 70s just like made this thing upand it has stood the test of time. Andnow these like three little commands arewhat power like all modern agenticcoding. It's insane.>> [laughter]>> My manager at Notion, his name's MaxStoening. He's very very good at nerdsniping me. And he nerd snipes me to bemore curious about how computers work.He's sent me on a journey to play withLinux. He's sent me on a journey tounderstand what's happening when youboot up a terminal. And he's doing thesame for for many other people atNotion. Like how do youdemystify AI so that hopefully you canfeel in control and wield it moreeffectively. Or just like have a betteropinion or like kind of see where thepuck is going.

00:45:07Like again, AI is usingthings that have been around for 50years.And the people who know about thesethings are getting a lot of mileage.They're like, oh no, I know howcomputers work. I can like point AI atthis computer shaped problem and and getcool things out the other side. So Ithink that's what I wanted people totake away from my talk. And and the talkwas really I don't know, I shared someof the the prompt examples I gave here,which is really about like how do youjust get the right upstream tokens thatthe AI can can extrapolate from there.How can you nerd snipe yourself to likeunderstand what's happening when you hitsubmit on a prompt. I don't know if itwassuccessful.But I hope I hope at least one personbecame more curious after that. Thecomment about Max and nerd sniping makesme want to go down another little rabbithole because Max is someone who I guessI've always looked up to from adistance. Like I've never met Max, neverworked with him. And I think it isrepresentative of almost the way that Iview Notion as a whole. Where I'm likethere's obviously a bunch of reallytalented people there building reallyincredible software. I've never been inreally any environment like that before.So you're now over a year in. How do youthink either that culture or the peopleor the way of working, like how else hasthat

00:46:23shaped you as a designer over thelast year plus? The cool part aboutNotion right now is from top to bottomwe are very comfortable revisiting oldassumptions. Does Notion really need tolook like this? Will it be this shapeforever? And we're starting to poke andprod at that. Like the the some of thedesigners on the team just redid the theleft sidebar making it much more of achat forward experience. A lot of usinternally pretty much has used Notionthrough through chat or through agents.The willingness to embrace that andchange and like question all theassumptions is very cool. II think that is a hard thing for acompany at this sizeto do because it's scary, right? It'slike things are going well, but theworld's going to look very different ina few years and we need to be be readyfor it.And then I think the second part of thatand what Max is good at and what theleadership here is good at doing isencouraging people to think bigger. Ithink right now OpenAI and Anthropic,they're way more YOLO about likewouldn't it be crazy if we did this? AndI think there's a lot of softwarecompanies that are a little bit scaredto ask that of themselves. It feelsoutside of their like permissionboundary. They're like culturalpermission boundary of being

00:47:38able toplay in that space. And so yeah, I Ithink Max and and everyone else is goodat pushing people to like, why not? Likewhy can't we I can't we Here's here'sNotion's box, right? Like people thinkof Notion is like a wiki, a note-takingtool,document collaboration tool. Like what'sright outside that box that's reallyinteresting. Let's go play there.And [snorts] that is a very fun place tobe. Yeah, we're like two days fromLinear or Karai publishing the articlesaying issue tracking is dead. Which I'mlike, how many companies on the planetwould would do that type of thing? I dothink Notion is potentially one of them.But it's like such a weird moment inSaaS where like you said, everything onpaper could be going well, but if you'renot looking I don't know how many yearsahead, it's just likeif there's one thing that's certain iswe know that like the amount of changein two years is going to beextraordinary. Nobody knows for certain,but it seems like a good idea for everycompany right now is uh understand howto make your systems legible to AI sothat you can participate in thatecosystem in the future. Seems like theecosystem won't go away. We don't knowexactly what the agents will be doing orhow they'll want to work in the future.So you just need to design your yoursoftware in somewhat of a flexible way.Which is why we come back to like

00:48:53APIsand CLIs. It's like these[clears throat] things have existed fora long time. Computers know how to usethem. They're not going to go away for along time. So like let's invest in thosethings, right? I thinkwe just shipped a Notion CLI. Notion'snot historically a developer tool. Wehave a developer platform you can buildintegrations and and build stuff foryour team. But hey, Notion has a CLI sothat everything you can do through theNotion API is now legible to an AI. Andof course we were very early with MCPand all that kind of stuff.Like these are the things that I thinkyou have to be doing to poke at theedges and and just make sure you're in agood position to move quickly as the themodels change shape or or newcapabilities come out. So yeah, I meankudos to Karai and and the Linear teamand any company that's not doing thatand like really taking an honest look attheir current business in this new worldwe're living in. If they're not doingthat, they're in trouble, I think.Final question, maybe we could shiftfrom company strategy to personal morecareer strategy for a second because Ohboy.the other thing that that blew up theinternet at least our little communityof Twitter was Lenny's tweet aboutdesigners being the only role that isnot experiencing growth right now. Andactually there's the correlation betweendesign growth and PM growth is

00:50:09not inour favor currently. Do you have anythoughts on like what that all means fordesigners and also just like what itlooks like to respond well in this timeas a designer who maybe is a little bitearlier in the career. Maybe they're notworking at a company like Notion and youknow, there's a there's a pretty heftyuncertainty cloud above us all at themoment. In the same way it would becrazy for a company that did issuetracking or note taking to not take ahard look at themselves in the mirrorand try and figure out their role in thefuture. It would be crazy for designerstoday to not be doing the same thing.I think that our obsession withtitles is what willscrew people over. If you're like, butam I a designer or a PM or an engineer?Like, oh my god, like which path? No,no, no, no, no.These [snorts] things are going away.Like all of this stuff is getting very,very blurry. It's getting very, veryeasy to move between those disciplines.And you want to be the kind of personthat can move between those types ofworking very fluidly and not get caughtup in, but what's my what's the box I'min at this company, right? Um so youwant a designer who can ship

00:51:24code toprove that an idea is good.>> [snorts]>> You want a PM who can design aroundtheir spec to see if their spec holds upto reality.You want an engineer that really caresabout building great visual componentsystems so that they can prototypebeautiful things without needing adesigner to come in and tell them tonudge stuff around, which is reallyannoying. Like you want people thatbreak out of these made up walls we'vebuilt around ourselves and our roles andour titlesand are just focused on shipping goodsoftware. So I don't know, that's hardto say to someone who's new becausethere's just so much to learn and somuch to do.I would say if you can just focus onmaking things for real people, thatsolve a real problem, iterate quickly,try and make the thing as good as youcan and stop worrying about but I don'tknow how to code. I don't know how towrite a PRD. Like who cares? Just likeship the thing and solve the problemand learn how to do that faster. Uhyou'll be fine. I don't know. I mean Idon't know. I see I think sometimes I Ilike I'm like I don't know how long I'mgoing to have a job. Like is all thisgoing away? Um and the only thing I knowhow to do is just keep making the bestsoftware I can make, solve real problemsfor

00:52:39real people, pay attention to thethe the tools, the patterns, try andunderstand what's going on that the labsare shipping, try and understand thecurrent meta of skills and contextmanagement and prompt engineering. Likejust try and understand it. And it doesfeel incredibly noisy. I would say thisis the most distracted I've ever been.It is the most context switch exhaustionI've ever felt. The other thing here isyou got to figure out a way to make thissustainable. I will say here's myprediction. This is going to age verywell or very poorly in the next 3months.When Opus 3.5 and and Sonnet 3.5 cameout last year, I think it was in likeMarch or April or something. I went on areal bender of like AI pilled everythingAI. I am doing a gigantic coding 24hours a day. Me and my my fiance who isalso a designer, we were taking avacation. We were like sending promptsand like sessions from our phones. Likewe had a hard time disconnecting,unplugging. We were just like, oh mygod, we have to if we're not prompting24/7, we're wastingtime. We're wasting valuable time. Andif that sounds familiar, it's becausewe're in the exact same moment rightnow. Everyone's feeling the same wayright now a year later. And what I cantell you is that in

00:53:54the middle of lastyear,ooh boy, I hit a wall. You hit a wallbecause you just can't sustain thatlevel of being plugged in and tryingevery tool and trying to be productiveevery waking hour of the day for morethan a few months at a time. I think atleast my brain can't handle it. And so Ipredict that we're going to have a comedown really soon. I would argue it'salready happeningum but in the next month or two. And theonly reason that that wouldn't be trueand that this will age poorly is ifthere's some big model capability thatdrops in the next month or two.But I think people are going to getexhausted. What that means is peoplewill come back to how do I spend my timemore effectively? How do I stay calm inall of the chaos and just focus? How doI turn off Twitter for 2 hours a day andlike actually get work done withoutbeing distracted? I don't know. Thesethings are going to be tried and true.Like I think the ability to focus and beundistractable for 2 hours a day is likea meaningful competitive advantage rightnow, which is insane.Brian,appreciate you, man. This is always fun.You're a joy to have on>> [laughter]>> and you always bring the the energy andthe laughs and this has been a greattime. [music] So thank you. Thanks forhaving me.

00:55:04>> [music]