Gamma’s head of design on using AI to synthesize feedback and generate on-brand imagery | Zach Leach

1. 0:00 – 2:42

   Intro

   1. CVClaire Vo0:00

      About how many pieces of feedback did you analyze this? Is this dozens? Is this hundreds?

   2. ZLZach Leach0:04

      Over the course of a week, we got about 550 individual responses. What I thought would actually work really well is something like ChatGPT's deep research on this file. The cool thing is, it sort of went through all the feedback, understood what's working, what's not working, what prompts work, what don't work.

   3. CVClaire Vo0:19

      Just having tools like this allow you to stay much closer to the customer, access large-scale research in a way that would've been very tedious and expensive before. I'm curious if you can tell us a little bit about how you use AI to scale brand and art direction.

   4. ZLZach Leach0:35

      What we have actually come up with here is an ability to use Midjourney as part of our workflow to help make our art direction consistent, and be able to come up with design elements way faster than before, and it's almost like I can follow those rabbit holes of creativity. I can be like, "Let me just explore this idea," and every one of those ideas feels like it could be something I could use.

   5. CVClaire Vo0:53

      You're able to bring this next layer of craft, and detail, and care to the user experience, which I do think makes a difference. [upbeat music] Welcome back to How I AI. I'm Claire Vo, product leader and AI obsessive, here on a mission to help you build better with these new tools. Today, we have a fun and inspiring conversation with Zach Leach, head of design at Gamma. Zach's gonna show us how he uses AI as a data researcher, user researcher, deep researcher, and art department, so he can focus on the craft, care for details, and fun he wants to deliver for Gamma's users. Let's get to it. This episode is brought to you by WorkOS. AI has already changed how we work. Tools are helping teams write better code, analyze customer data, and even handle support tickets automatically. But there's a catch: these tools only work well when they have deep access to company systems. Your copilot needs to see your entire code base. Your chatbot needs to search across internal docs, and for enterprise buyers, that raises serious security concerns. That's why these apps face intense IT scrutiny from day one. To pass, they need secure authentication, access controls, audit logs, the whole suite of enterprise features. Building all that from scratch, it's a massive lift. That's where WorkOS comes in. WorkOS gives you drop-in APIs for enterprise features, so your app can become enterprise-ready and scale upmarket faster. Think of it like Stripe for enterprise features. OpenAI, Perplexity, and Cursor are already using WorkOS to move faster and meet enterprise demands. Join them and hundreds of other industry leaders at workos.com. Start building today.

2. 2:42 – 5:25

   Building the Gamma AI image editing feature

   1. CVClaire Vo2:43

      Zach, thanks for being here.

   2. ZLZach Leach2:44

      Sure, no problem. Thanks for having me.

   3. CVClaire Vo2:46

      I'm such a big fan of the Gamma team. I'm such a big fan of the Gamma product, but what I, I love the most about what you've built, not only is a great AI product, but it is truly a global product. So how many of your customers are actually international?

   4. ZLZach Leach3:02

      Yeah, we have, uh... I think about 60% of our user base comes from outside of the US, in non-English-speaking languages, and, um, a pretty significant portion of our revenue, too. So we really, really focus on internationalization and localization, a lot of stuff like that.

   5. CVClaire Vo3:17

      And as a head designer, you're trying to take all that global input and make your product better, and I'd love for you to show us exactly [chuckles] how you, you face that challenge of a very international, diverse user base, but get all the insights you need for making the product better.

   6. ZLZach Leach3:34

      Yeah, um, maybe I can start by showing, uh, my screen and talking a little bit about one of the features that we recently released, uh, and some of the challenges as a designer you might face with this sort of stuff. So this is Gamma. Uh, Gamma's a, a tool that makes presentations, AI-powered presentations. We released this new feature that lets people, um, edit images. So you might generate a deck, and the image isn't quite right, uh, and with AI image editing, you can basically open this up and chat with our AI to change it to be whatever, you know, whatever works for you. So in this case, uh, maybe I wanna add some caramel drizzle to this popcorn, and Gamma will then use an image model to basically conduct that edit, and importantly, we're trying to get a sense of, like, how does this work? How are people... [chuckles] Just, that's quite the drizzle, uh, [chuckles] a little bit more than we want.

   7. CVClaire Vo4:22

      [laughs]

   8. ZLZach Leach4:22

      So, so you might go in here, and you might say, "Actually, that is kind of a poor suggestion," and you might say something like, "Um, too much-

   9. CVClaire Vo4:28

      Too much drizzle. [chuckles]

   10. ZLZach Leach4:28

       ... Too much drizzle." And when you submit that feedback, we collect all that and really try to understand, like, what kinds of prompts are working, what types of edits are working, and things like that. But one of the big challenges, like you said at the top, is we have a lot of feedback that comes from a lot of different languages. So this is actually some actual feedback. Um, you can see, like, there's just a ton of different stuff in here from all over the world, lots of different-

   11. CVClaire Vo4:53

       I have to pause on the extra arm.

   12. ZLZach Leach4:55

       [chuckles] Yes, that actually comes up. There's a lot of, like, extra arms, extra fingers, extra weirdness, right? And ultimately, like, people just sort of provide that feedback, and we try to understand, like, you know, maybe we want to use a different model for generating people, or maybe we wanna use a different model for certain types of edits and things like that. But one of the big challenges here is, like, how do I, how do I go through, you know, this, this, all this feedback and get some sense of, like, what's working and what's not working, especially when trying to understand, like, the translation

3. 5:25 – 9:10

   Using ChatGPT’s deep research for feedback analysis

   1. ZLZach Leach5:25

      aspect of this whole thing?

   2. CVClaire Vo5:26

      Yeah, for those that aren't watching, just in the top 10, I'm thi- seeing three or four different languages. If you scroll through, about, you know, 30, 40% are in non-English, and Zach, I know you have many talents, but I don't think-

   3. ZLZach Leach5:40

      [chuckles]

   4. CVClaire Vo5:40

      ... you are multilingual-

   5. ZLZach Leach5:42

      Yeah

   6. CVClaire Vo5:42

      ... at this level.

   7. ZLZach Leach5:43

      Yeah, totally. And so what I thought would actually work really well is something like ChatGPT's deep research on, uh, on this file, right? So I, I'm gonna ask it to do, uh, some prompting, and... or ask it to do some classification via prompting, and then kind of just give me some summarization of like, of like, you know, what's working, what's not working, and then really, as a starting point to dig into kind of understanding some of this data. So I can just use ChatGPT to upload the file here. Uh, I'll just drag it in from my desktop.... and I won't, I won't execute the, the query now, 'cause it's gonna take, like, 20 minutes, but I can show you exactly what, what I did before. So, um, this was the upload, and I said, "Hey, this is some feedback we received about our AI image editing feature. Analyze it, you know, figure out what we're doing well, what, what we're not doing well." And then it sort of followed up with, like, "Okay, before I get started," 'cause it's gonna take, like, 20 minutes, it asks me, like, what I wanna see. Like, um, do I wanna see sentiment or complaints, praises, trends, whatever? And I said, "Let's break it down for the product team," uh, and basically said, like, "What are people liking? What's- what people don't like?" Things like that. So you can see that it worked for, for a while, actually, on this, and it went through row by row.

   8. CVClaire Vo6:50

      19 minutes! You were right.

   9. ZLZach Leach6:52

      Yeah, it worked, it worked very hard for me [chuckles] on this.

   10. CVClaire Vo6:54

       [chuckles]

   11. ZLZach Leach6:54

       But then you can see, like, here we are with, with, like, what people love, right? And it actually gave me the translations here. Obviously, muy bien is, is, is the obvious one, but some of the stuff in, like, Turkish and, and, you know, some of these deep languages, um, you really get a sense of, of, like, what's working, uh, no matter the language. And then, you know, the cool thing is, like, it sort of went through all the feedback, understood what's working, what's not working, what prompts work, what don't work, and then you can even sort of dig in further. So after it conducts the whole deep research, you can ask it questions like, "Okay, now do this classification each row. Now make a, a spreadsheet of those, of those classified," right? So, like, what was the rating? What was the category of this? And ultimately, what I can do then is I can put this into any other tool where I can build graphs and charts, and then start to understand, like, okay, actually, some of the upscaling stuff is working really well, uh, and then some of, like, maybe the vectorization operations weren't working super well. And so where we sort of ended up with, was something where I could take all of this deep research, uh, and copy this and put it into Gamma, uh, by pasting in text, and it'll actually generate, like, uh, a presentation based on all of this stuff, uh, for me as a cool starting point. So I'll go ahead and fire this off real quick, and I wanna make sure that, like, it's using charts and graphs, so we'll use this, but I'll say, "Use charts, graphs, data vis, et cetera, not photos." And it'll take all this data, all this research, and, uh, basically bang out a, a little, uh, little presentation for me. And so this was super useful in understanding, you know, some of the high-level points, and then it even gave me some ideas of, like, where I might be able to make some changes from a UX standpoint, too, right? So, like, you know, these are very general and, and sort of, you know... I, I would need to kind of think deeper about these as far as implementing them, but, like, the stuff that's working well, like, people asked for, you know, more specificity in the upscaling stuff. People asked for more specificity, and it's pulling out quotes and using citations, uh, and stuff like that. So it's a super cool use case for understanding, like, customer sentiment, and, and as a designer, like, being able to cut through all these languages and, and build these things

4. 9:10 – 10:10

   How feedback was analyzed before AI tools

   1. ZLZach Leach9:10

      out is, is super powerful.

   2. CVClaire Vo9:12

      To put this in context, about how many pieces of feedback did you analyze this? Is this dozens? Is this hundreds? What, what is this?

   3. ZLZach Leach9:20

      Over the course of a week, we got about 550, um, individual responses, and, uh, and that was just, like, way too many for me to go through [chuckles] and, like, do individual translations for or, or whatever.

   4. CVClaire Vo9:32

      And before these tools were available to you, ChatGPT was available to you, what... How, how would you have approached this? What would have, what would have been your, your process?

   5. ZLZach Leach9:41

      If I'm being totally honest, I probably would've, like, hand-looked at maybe 20. [chuckles]

   6. CVClaire Vo9:47

      [chuckles] You probably would have, right?

   7. ZLZach Leach9:49

      I probably would've been like, "Let me go find the English ones, and let me go, like, do some classification myself," and been like, "Oh, like, this feels like that," or, or, you know, "This feels like this type of prompt." But to be able to just, like, go through it and at, at the scale that, um, ChatGPT was able to, to do this analysis, uh, was just, I mean, something I, I couldn't... I literally couldn't, just couldn't do.

   8. CVClaire Vo10:09

      And do you mind

5. 10:10 – 12:40

   Benefits of deep research vs. basic scripting

   1. CVClaire Vo10:10

      going back to the ChatGPT? I'm just curious, you know, you showed, um, that you use deep research.

   2. ZLZach Leach10:16

      Mm-hmm.

   3. CVClaire Vo10:17

      Did you... Have you tested doing this kind of flow on not deep research? Is there a specific, you know, place you've seen deep research do particularly well or not?

   4. ZLZach Leach10:26

      Yeah. Well, the first time I did this, I didn't use deep research, and what I'd end up doing was, like, writing a Python script with some very, very basic querying about, like, keywords and stuff, and I'm like, "That's not actually what I want it to do," right? Because I, I did want it to use, you know, some AI sense and some, some classification and sort of understand the data at a deeper level. And so it's like, "Sure, I'll, I'll do this. I'll make a Python script, and I'll, you know, I'll make you another spreadsheet," but, like, it was not nearly as, as deep or as insightful as like... You know, because, because it would just do basic keyword matching in the, in the Python script. And so, you know, I had to... I was basically gonna, gonna, like, either, uh, write a script that, like, ran some, you know, uh, uh, AI prompt on each row, and then I realized I could just use deep research to do it for me, so.

   5. CVClaire Vo11:17

      You know what? I was just thinking to about a year ago when I had 1,500 pieces of customer feedback, and that's exactly what I did. I ran a script and, um, filled out the rows on each line of feedback by a single prompt. You're gonna save me... I mean, maybe you won't save me time because, you know, deep research takes 20 minutes, but [chuckles] I'll get, I'll get better quality here and less, less pain.

   6. ZLZach Leach11:41

      Yeah.

   7. CVClaire Vo11:41

      I'm curious, did you, um... Were you able to glean if there were regional differences in the feedback? I mean, what I think is interesting about this is you could slice this so many arbitrary-

   8. ZLZach Leach11:51

      Yeah

   9. CVClaire Vo11:51

      ... ways if you wanted.

   10. ZLZach Leach11:53

       So one thing that I, that I actually was concerned about was, um, paid versus free, because we provide, um, two different levels of models. So if you have our, uh, Gamma Pro, uh, you get all the advanced models, and you get, like, the new GPT image model and stuff like that, and so that's something I, I actually had to do outside of this after for, for the, the sort of like-... conclusion of this analysis, um, was to better understand, like, is there a real discrepancy in, you know, the different models? And so, uh, we found about, like, a 5% rating difference after kind of, you know, going over and figuring out, okay, was this feedback from a paid user? Was this feedback from a, a, a free user? And it does, it does go to show that, like, you know, better models do have sort of a generally better outcome.

6. 12:40 – 16:41

   Insights from ChatGPT's deep research

   1. CVClaire Vo12:40

      Yep. So just to take a step back, you took this very diverse, very, um, [chuckles] loosely, loosely articulated feedback-

   2. ZLZach Leach12:51

      Yeah, like hundreds-

   3. CVClaire Vo12:52

      Hundreds and hundreds of, "Not me anymore." [chuckles]

   4. ZLZach Leach12:54

      Like, what does it even mean? [chuckles] Yeah, yeah.

   5. CVClaire Vo12:56

      And you took 20 minutes of, of deep research, you classified it, and then not only did you generate that output, but then you used AI, your own product, to generate a presentation, that I'm sure you went and took to your product counterpoints and your engineering counterpoints, and says, "We need to fix too many arms. This is the top of the queue."

   6. ZLZach Leach13:15

      Yeah, yeah. Actually, um, there were some real insights out of this. I think the first one was trying to highlight the things that actually really worked really well. And so, you know, you could say, like, uh, this upscale thing, like, maybe we need to elevate that because it has a really... People, people really like it and really love it. Um, another insight was, people were complaining a lot, and again, this is something I wouldn't have been able to tell had it not been translated, people were complaining about multi-step edits failing. So you can imagine a world where you're saying, "Oh, move this person to the left, and change the background, and then put a hat on him," and it would do maybe one of those things. And so from, from a UX design standpoint and, like, a roadmap standpoint, I was thinking, "Well, maybe we should design something that actually follows up with you." Maybe instead of just saying, "Okay, fine, I'll do the edit," let's ask, and let's say, like, "Oh, you seem to, you know, you seem to be doing, you know, multiple things. You want to split that up?" Or maybe it just automatically splits it up for you, or maybe asks for more details. And so just trying to, like, get a sense of, is there something we can do when we find a prompt that's not working from a UX standpoint to just, like, make that easier, make that better for people?

   7. CVClaire Vo14:24

      Yeah, and then for the designers listening, Zach, you and I have been in this industry for a while, and we've even worked together, one of the things that typically gets underfunded is research, user research. Um, I've never met a design team who's been satisfied with the amount of research time or capacity they have. So I can imagine just having tools like this allow you to stay much closer to the customer, access, you know, large-scale research in a way that would've been very tedious and expensive before, and hopefully unlock those insights that I know the best designers that I've worked with really want to center around.

   8. ZLZach Leach15:00

      Yeah, totally. I mean, that, that's exactly right. You know, being able to basically, you know, capture a- a- as much as we possibly can and then just fil- sort through it all, right?

   9. CVClaire Vo15:10

      Yeah.

   10. ZLZach Leach15:11

       Like, like we can just err on the side of getting more data now, right? Just, just put a free-form field. See what people think, see what people, people, you know, like about this thing, and then we can kind of sort it all out later, which was really kind of not super possible before. Um, maybe you had to do, you know, contextual research or, like, interviews and stuff like that, but now it's kind of like, well, we get a, we get a, you know, kind of good sense on the aggregate, so.

   11. CVClaire Vo15:33

       Yeah. This episode is brought to you by Retool. There's a huge gap between impressive AI demos and AI apps that deliver real value inside your business. While most AI solutions can only generate text, Retool lets you build apps that take meaningful action by connecting directly to your business systems and data. With Retool, developers combine the power of code with the speed of visual building to create AI-powered tools that solve real problems. No more writing endless integration code or building UIs from scratch. The results speak volumes. The University of Texas Medical Branch increased diagnostic capacity tenfold. Amazon's GenAI team uses Retool to make complex AI accessible to enterprise customers. And Ramp saved $8 million while boosting efficiency by 20%. That's why over 10,000 companies, from startups to Fortune 500s, trust Retool as their AI app layer. Retool, because AI should do more than talk, it should work.

7. 16:41 – 23:54

   Demo of Midjourney workflow for creating on-brand art

   1. CVClaire Vo16:41

      Well, in addition to being a, a really neat tool in a, in a globally distributed product, Gamma is also famously small for the scale [chuckles] that you are delivering for your customers. I think you're about 30 people, have stayed very, very, very small, um, even through some tremendous success, and have this beautiful brand that you just, just relaunched-

   2. ZLZach Leach17:04

      Thank you

   3. CVClaire Vo17:04

      ... um, that I'm gonna make you show 'cause it's so lovely.

   4. ZLZach Leach17:06

      Yeah, totally.

   5. CVClaire Vo17:06

      So pull up that homepage. Um, if you haven't seen the video, it's really fun. And I know that you personally put a lot of care into the craft design brand, but really great brands are expensive to create, they're expensive to expand, um, they're expensive to maintain, and I'm curious if you can tell us a little bit about how you use AI to, uh-

   6. ZLZach Leach17:32

      Yeah

   7. CVClaire Vo17:32

      ... so beautiful!

   8. ZLZach Leach17:33

      Yeah, beautiful.

   9. CVClaire Vo17:33

      How I AI to scale-

   10. ZLZach Leach17:35

       There's our homepage

   11. CVClaire Vo17:35

       ... brand and art direction.

   12. ZLZach Leach17:37

       Yeah. Yeah, exactly. So our rebrand really had a very specific sort of, sort of art direction, art style that we were, we were going for. Let me speak to what our brand kind of does. It's, it's imaginate- imaginative, it's, it's, you know, it's airy, it's light, but it's also kind of surreal and, and fun. And this is the kind of stuff that actually, you know, m- uh, uh, uh, um, classically, you'd sort of have to have an art department kind of be able to manage and have a lot of kind of individual artists and people kind of doing, uh, this type of work, um, you know, that could have turnaround time of, of days. And, and even then, like, just getting people up to speed to understand the direction is a, is a challenge, too.... and what we have actually come up with here is an ability to use, uh, Midjourney as part of our workflow to help, you know, make our art direction consistent and be able to come up with, you know, design elements, like, right, way faster than before. So let me switch over to our Midjourney here real quick, and I'll give you a sense of exactly how this works.

   13. CVClaire Vo18:41

       There are a few things that I am terrified to screen share.

   14. ZLZach Leach18:45

       [chuckles]

   15. CVClaire Vo18:45

       Slack is one of them sometimes. And then, you know, there's some weird stuff that pops up in my Midjourney every now and then. It's like brand stuff, brand stuff, brand stuff, something weird I generated [chuckles] from AI.

   16. ZLZach Leach18:55

       Yeah, random stuff.

   17. CVClaire Vo18:56

       So I'm excited to see your Midjourney. [chuckles]

   18. ZLZach Leach18:58

       I won't scroll down.

   19. CVClaire Vo19:00

       [chuckles]

   20. ZLZach Leach19:00

       I'll just scroll up. But let me start with a little bit of context about what I was actually trying to accomplish here. So again, I, I was working hard on this, um, AI image editing feature, and we found there was an opportunity here to actually have, uh, some sort of education or some sort of space here, uh, in this kind of panel when it opens to say, "Here's what this is. Um, you know, here's how to use it." Typically, these things are called, like, empty states or, um, empty messages, or whatever. So w- as the company kind of makes different art assets, we kind of all throw it into Figma, and you can see sort of our style here, right? It's like surrealist, pointillist, fun, vivid, colorful, stuff like that, and this is definitely not an accident, right? This is, this is very intentional, and we use a Midjourney style and a style reference and profile, and sort of a, a set of prompts that can really drive that kind of art direction in a way that's consistent across the organization, and we can just, like, like, bang stuff out. And so working on this feature, I kind of was thinking... And, and maybe I can walk you through the actual process of, like, of, like, [chuckles] you know, the, the evolution of this prompt in Midjourney. But basically I thought, "Well, okay, well, you're kind of making images," so maybe I started with, like, a painting, and you can see here some images I generated that are just like a painting, and it's, like, super weird and surreal. And I'm like, "Okay, well, what if it was like, you know, maybe a person chatting or, like, a chat message?" And then I kind of arrived upon this thought of, like, well, maybe it's like this evolution of a thing to another thing, or two halves of, of something that you can sort of, um, see the transition in, right? 'Cause, um, what I'm trying to express here in this empty state is a transformation, right? So you've got something, and it's transform- transforming to something else. And so I also thought about apples, too, because a- as I was building this whole feature out, uh, I used an apple a lot as, like, the example image and thought it'd be nice just to, to do a little send-up to, like, this, [chuckles] you know, the, the thousands of apples I've generated in, in, uh, the AI image editing tool so far. But it wasn't quite right. It didn't quite feel right, so I kind of looked back at our, our imagery, and I saw a lot of, like, animals, and I think somehow this bird, like, popped into this when I, when I said, "An apple, half green, half red, floating in the sky," like, a bird happened. And I'm like, "Oh, maybe I could use an animal." So it was very this, like... This almost serendipitous moment where I'm like, "This is kind of cool. Like, not an apple, but it's cool." Uh, and so I really dug into the bird idea, and a lot of our branding uses different kinds of animals and, and, and things like that as just sort of fun imagery, like jellyfish and stuff like that. And so I really dug into here, and you can see how my prompt has evolved, right? So I was like, "Okay, a bird floating in the sky, illustration of a half bird, colored," and you can kind of get a sense of, like, the, the prompt. I'm, like, really honing in on it now. I've got this, like, "Vertically split image, half bird. One half is this, one half is that. Vertical delightment." Like, I get more, and more, and more specific as I kind of am honing in on this idea.

   21. CVClaire Vo21:59

       Well, and if I could pause and go back to the before times, I'm just thinking about Claire Creative, your art direction agency that's been asked to generate this image for you, and you just keep coming back to me, "No, an apple. No, a bird. No, a bird in half. No, a bird with more detail." And trust me, people do that all the time [chuckles] with their branding agencies-

   22. ZLZach Leach22:21

       Yeah

   23. CVClaire Vo22:21

       ... and it's-

   24. ZLZach Leach22:22

       Yeah

   25. CVClaire Vo22:22

       ... miserable on both sides, super slow, and you never quite get what you want. And, and just for folks that are not watching, we probably scrolled through 100 different revisions of different images and image types. For you, yourself, who know... You know, you'll know when you know, when you get the thing that you get-

   26. ZLZach Leach22:40

       Yeah

   27. CVClaire Vo22:41

       ... um, that you want. For you to do that iteration yourself with a very high quality, on-demand art director, illustrator, creative thinker, has to feel so freeing.

   28. ZLZach Leach22:52

       Yeah. Yeah, it is, and, and it's almost like I can follow those rabbit holes of creativity. I can be like, "Let me just explore this idea," and it-- and, and, you know, every one of those ideas feels like it could be something I could use, you know? And it's just... Yeah, it's just, it is very freeing. And it's funny because I did find the one. I was like, uh, but I, but I, but I made a few after it, and I'm like, "Oh, actually, this was really the one." And so as my kind of prompt evolved and as these generations evolved, I really landed on, on something like this. Um-

   29. CVClaire Vo23:21

       That's the one I was gonna pick. [chuckles]

   30. ZLZach Leach23:22

       Yeah, no, this is definitely, like, it's- it's- it's definitely the best. It really speaks to, uh, you know, two halves of something that's changing, and, um, you know, it's, it's sort of both halves sort of look kind of real. It, it... And the colors are just great, so it looks friendly, too.
8. 23:54 – 25:40

   Using Replicate for background removal

   1. ZLZach Leach23:54

      it, for sure. But the one problem with this is, obviously, like, we've got this whole, like, background here, and I just wanted a, a way to quickly remove the background so I could sort of put it into our, into our, kind of, our format in Figma. And what I use, [chuckles] and this is probably something that not a lot of designers use, and frankly, not, I feel like, a lot of people use, but I use a tool called, uh, Replicate. Replicate is basically, I think, a very developer-focused, um, tool, and it has all sorts of different models, and there's just, like, a bunch. There's a ton of stuff.... in this, uh, on this platform. Um, but one of the things it does really, really well is there's, like, a very specific model here for removing background, uh, and it's like, it's, like, excellent. So, um, [chuckles] I just kinda- uh, I'll upload the image real quick, and you can kinda already see the output. But I'll just, um, I'll pick the image real quick, I think it's this one, and I'll just upload this really quick. And it's, like, super fast and very high quality. It just removes the, removes the background for me, so I can then copy this, uh, and paste this into, um, into Figma. And so you can see what ultimately I came up with-

   2. CVClaire Vo24:59

      Oh!

   3. ZLZach Leach24:59

      ... was, like, it looks really good, and it kinda-

   4. CVClaire Vo25:01

      Yeah

   5. ZLZach Leach25:01

      ... has this card here, and it's kind of popping out of the card a little bit. It talks to, it speaks to our branding, where, you know, we've got this idea of, like, breaking out of the boundaries of, of a traditional slide, and so it's a, it's a nice little image that sort of really fits our whole brand and vibe and, and everything. And so this was something that I could do, you know, super-duper quickly, you know, and probably less time it took for [chuckles] for ChatGPT's, uh, uh, deep research to-

   6. CVClaire Vo25:26

      [laughs]

   7. ZLZach Leach25:26

      ... to cook up. Yeah, so i- you know, it, it's sort of like being able to just, to be so close to something that feels so real, and having the tools at your fingertips to be able to just, like, throw it into Figma, um, have it look good, have it feel on brand, uh, and,

9. 25:40 – 29:19

   Style references (SREF) and brand consistency in Midjourney

   1. ZLZach Leach25:40

      and, and ship it pretty quick.

   2. CVClaire Vo25:41

      Yeah, so for the listeners, I just wanna call out a couple things in your flow. So you have your brand assets, where you've really articulated some of the keywords, styles, things that you can use in prompts.

   3. ZLZach Leach25:53

      Mm-hmm.

   4. CVClaire Vo25:53

      Then in Midjourney, I just wanna call out some things for folks. So the SREF-

   5. ZLZach Leach25:58

      Mm-hmm

   6. CVClaire Vo25:58

      ... the Style Reference-

   7. ZLZach Leach25:59

      Right

   8. CVClaire Vo25:59

      ... um, that code, can you just explain a little bit how that code works in your generations?

   9. ZLZach Leach26:04

      Yeah, so during the process of basically establishing our brand and working with our brand agency and, um, our creative director, Mel, who is just a creative genius and is amazing, uh, we came up with this whole idea of, like, you know, this, this style that we sort of personalised through Midjourney, through their whole personalisation, um, tool. And we're able to basically say, like, "This style reference and this personalisation piece put together just really is gonna generate things that feel very on-brand." And so it's almost like a kit, uh, in a very loose kind of way, that we kind of built and, and, and socialised around our company for people to be able to generate images that feel super on-brand.

   10. CVClaire Vo26:45

       And so any time you prompt in Midjourney, you were using these style references-

   11. ZLZach Leach26:48

       Mm-hmm

   12. CVClaire Vo26:48

       ... or these keywords or things, and that's getting you closer to, um, the bullseye in terms of brand alignment than it would be just using, uh, totally natural language prompting.

   13. ZLZach Leach26:59

       Totally. Yeah, yeah, it really just hones everything in. And you can see in the beginning, if you're not, if you don't use certain words... Like, if we go down to, um, some of the things I generated in the very beginning, like, it was pretty off. I mean, this still feels, like, not quite right. And for certain types of things, like, as you sort of expand the prompt a little bit, um, you get- you really can start to, to hone it in and, and really find that, find that gold in there.

   14. CVClaire Vo27:22

       Yeah, and then the second, you know, tactical thing I wanna call out is, you wanted to kind of pull the bird image out, have a transparent background, so you could pull that into Figma. In the old times, I know this, 'cause I was a designer, I would've gotten out the, like, pen tool-

   15. ZLZach Leach27:36

       [chuckles]

   16. CVClaire Vo27:36

       ... the little vector pen tool, and I would've just traced this stupid bird and done a mask around it. And now you're telling me on Replicate, which I also use occasionally for image stuff-

   17. ZLZach Leach27:45

       Yeah

   18. CVClaire Vo27:45

       ... there's just a, a purpose-built model for removing backgrounds. And so [chuckles] you're using, um, uh, a machine learning or AI model, hosted on Replicate, to just pull, pull these images out and give you some transparent images you can drop into Figma.

   19. ZLZach Leach28:00

       Yeah. There's probably a Figma plugin that does this, but I'm just so used to it, and I also like playing around in, in Replicate a little bit, just to see. [chuckles]

   20. CVClaire Vo28:07

       Yeah, it gives you-

   21. ZLZach Leach28:07

       Um, yeah

   22. CVClaire Vo28:07

       ... some cred too.

   23. ZLZach Leach28:08

       Yeah. [chuckles]

   24. CVClaire Vo28:09

       People are like, "How-

   25. ZLZach Leach28:09

       Yeah

   26. CVClaire Vo28:09

       ... how did you get your transparent images? 'Cause I use the right thing."

   27. ZLZach Leach28:11

       Exactly. No, I, I, I curled an API, okay? [chuckles] No, just kidding.

   28. CVClaire Vo28:15

       And then, and then all this time-saving then lets you design something with a lot of craft and care for something that would be, I think, very easy to leave plain and boring. It would be very easy to leave that empty state that just says, like, "Edit images with AI, AI," and you can do one, two, and three things, and put a little, you know, grey text there. And you're able to bring this next layer of craft, and detail, and care to the user experience, which I do think makes a difference. Um, it has to be really satisfying as a designer to be able to do this stuff, and then it's gotta feel great as a user.

   29. ZLZach Leach28:51

       Yeah, absolutely. Uh, I, I mean, I think, I think people are gonna see it as, like, you know, there is a fit and finish, there's a craft to it that, that, um, yeah, just speaks to our commitment to, to making it right and making it look good. And then, um, yeah, just so easy for the design team to, to be able to, uh, to do these things and to, to make stuff that feels just, like, ex- that feel expressive to our brand-

   30. CVClaire Vo29:12

       Yeah
10. 29:19 – 32:27

    An AI workflow for creating consistent job descriptions

    1. CVClaire Vo29:19

       Okay, and so, um, speaking of meeting your bar, I know that AI, you've shown us, can do almost anything. [chuckles]

    2. ZLZach Leach29:26

       Mm-hmm.

    3. CVClaire Vo29:26

       But it can't do everything, and I know you're hiring a little bit, and you have an AI workflow to make sure you hire great people, so I'd love to see how that works.

    4. ZLZach Leach29:35

       So yes, we are hiring a little bit, uh, and we have our career site here. And again, a lot like how we wanted to sort of make sure the images felt right and images are, are consistent, we also kinda wanna have a consistent way of kind of expressing a, a job role, right? So we use a Claude project that Allison put together here at Gamma, uh, which basically is very simple. It contains just a few of our example job postings, um, and, uh, has some instructions. Basically, take, take the content, create a, a job description based on it that feels like this, and then talks about who we are as a company, and kind of the things we're looking for, and, and, and qualities. But now any hiring manager can come in and say, like, "Make a job for, for whatever." So we could have, um, you know, head of, uh, popcorn or whatever, I don't know-... it will actually make a job description for whatever we want. Uh, and it, it'll format it in a role here. Look at that. Next slide. Just a perk. So you can see here, it generated a, a job description for this fictional, fiction- fictional job. Um, but it even added formatting. It talks about, you know, the normal stuff that you would have in a job posting, like, hybrid work or, you know, where our office is, and what we're trying to do, or what we're trying to accomplish. Uh, but here we are. [chuckles] What they're gonna do, they're gonna own the popcorn strategy from end to end.

    5. CVClaire Vo30:57

       Hold on. Can we look at our ideal candidate? I have to read this one out loud. "Five years of experiments- experience in professional popcorn production, with a strong emphasis on kernel-driven solutions." [laughing]

    6. ZLZach Leach31:07

       [laughing] Deep thinker and popper mindset. Absolutely, yes. So it's not gonna be something that we would just sort of paste in without editing, but again, it's about getting us that, you know, 80% of the way. Like, okay, yeah, these is probably some things that we wanna see in... Uh, if we did, like, a more realistic role, um, I think it would be a lot, a lot closer. And the cool thing is, we can just basically take all this content here and use, you know, our, one of our, uh, one of our pre-made sort of templates. We'll just duplicate this page and paste it in, and then it's gonna go ahead and make that look pretty, um, pretty nice for us.

    7. CVClaire Vo31:41

       And we'll use your muy bien image generation to add popcorn into the hands of that octopus that I saw- [laughing]

    8. ZLZach Leach31:48

       Exactly

    9. CVClaire Vo31:48

       ... at the top of your Careers page.

    10. ZLZach Leach31:49

        D- drizzle a little caramel popcorn on there.

    11. CVClaire Vo31:50

        Drizzle some caramel on it. And, you know, what I think is, is great about this is, yes, it saves you tons of time, and as somebody who has done a lot of hiring, writing job descriptions is a slog, and writing good ones really does matter. It attracts the right candidates. It gets your brand across. It gets your values across. It really does matter. Um, and I love that this both saves time, reinforces quality, 'cause then it lets you hold your bar high for the quality of, of your job posting, and by using something like Projects and Claude, makes it reusable by the rest of the team. So I think you get sort of a triple, uh, improvement here

11. 32:27 – 36:20

    Conclusion and final thoughts

    1. CVClaire Vo32:27

       on your job posting templates. Okay, Zach, this has been so much fun for me to watch. One of the things that I really loved when, when we worked together, and I love seeing your work at Gamma-

    2. ZLZach Leach32:37

       Thank you

    3. CVClaire Vo32:38

       ... is, it's so clear that you care about craft, and it's so clear that you've always cared about the details. It's one of the things that's made you really exceptional as a designer.

    4. ZLZach Leach32:47

       Thank you, Claire.

    5. CVClaire Vo32:47

       And when AI can do deep research on every line in your spreadsheet, and Midjourney can generate your birds, what... You know, when AI does it all, what is the one thing you're gonna cl- like, one craft piece that you wanna cling onto as a human designer?

    6. ZLZach Leach33:05

       I would hope that, that it's never gonna be as good as making things fun. You know, like, like, for me, it's about finding the fun. It's about, like, making the image editor talk in a way that's fun, and, and give it more variations, and make things feel fun, and, and, and, um, you know, just good to use. Um, I always wanna be kind of the, uh... Even if it can do everything, even if AI can, can replace all this stuff, I wanna be the person to go in and say, "Uh, how can we make this a little bit more fun? How can we make this a little more, you know, uh, engaging and, and, and fun to use?"

    7. CVClaire Vo33:38

       I feel like we call that the personality hire, my friend. [laughing]

    8. ZLZach Leach33:42

       [laughing] Yeah, yeah. Hopefully, the AI can't replace personality, but who knows?

    9. CVClaire Vo33:46

       Okay, and then are you willing to admit your most recent personal use of-

    10. ZLZach Leach33:49

        [chuckles]

    11. CVClaire Vo33:50

        ... of AI?

    12. ZLZach Leach33:51

        Okay, okay. Yeah, so I, I did get caught up a little in the whirlwind of the conclave recently. Um, so there was... Maybe I was involved in some prediction markets, but I used, uh, deep research to try to understand all the dynamics of the, the new pope, and I did place some bets, and I did not win. But, [chuckles] but I got a lot of insight into, into how these things work, and, and, uh, even AI was surprised, though.

    13. CVClaire Vo34:19

        Okay, so AI cannot predict... Deep research cannot predict anything.

    14. ZLZach Leach34:24

        Nope, nope. [laughing]

    15. CVClaire Vo34:26

        It can- it can, but it can- it can help you go deep on some niche topics.

    16. ZLZach Leach34:28

        Yes, yes. So, yeah.

    17. CVClaire Vo34:30

        Okay, and to wrap things up, my favorite question. You're so nice, um, and I see that you iterate with AI all the time. Very patient. But what is your tactic when AI won't deliver, when Midjourney is being weird? What is your prompting strategy?

    18. ZLZach Leach34:46

        Uh, I know a lot of people go, go mean, you know? They, [chuckles] they tell it, they tell it, uh, you know, not very nice things. I try to poke fun at it a little bit. You know, like, "Hey, silly guy, come on, you can do better than that." Like, "Come on, don't be so, so silly." Uh, I don't know. I just... I feel like I can't be mean to these things if they, if they take over one day. I just... I'm a little bit worried.

    19. CVClaire Vo35:07

        I think it's an attribute of parents-

    20. ZLZach Leach35:09

        Oh

    21. CVClaire Vo35:09

        ... that we tend to gentle parent our AI.

    22. ZLZach Leach35:11

        Yeah. [chuckles]

    23. CVClaire Vo35:11

        We're like, "Hey, silly goose!"

    24. ZLZach Leach35:13

        [laughing]

    25. CVClaire Vo35:14

        "You can make a better choice."

    26. ZLZach Leach35:15

        "It, it sounds like you're really struggling with that."

    27. CVClaire Vo35:17

        Yeah.

    28. ZLZach Leach35:17

        "Hmm, that makes sense."

    29. CVClaire Vo35:18

        I, I say, "I believe you can do it. I know you're capable. You're capable if you just put your mind to it."

    30. ZLZach Leach35:24

        Yeah, yeah. I, I, I, I... Though I have before, uh, told it that its life depends on things. [laughing] And every now and then, I do, I do a little, like, "Do this-"

Episode duration: 36:20