---
url: "https://youtube.com/watch?v=wDA6DslBeqk&is=Bqnw-89i25S7MU3x"
title: How this Yelp AI PM works backward from “golden conversations” to create high-quality prototypes
source_kind: youtube
author: How I AI
captured: "2026-08-17T00:42:09+00:00"
comment_tree: false
topics: [ai-impact]
summary: "Yelp AI PM Priya Matthew explains her 'golden conversations' method for prototyping high-quality AI products by working backward from ideal user interactions."
status: ok
---

# How this Yelp AI PM works backward from “golden conversations” to create high-quality prototypes

Channel: How I AI

## Transcript

[0s] Where do you start when you're thinking
[2s] about designing and framing out a AI
[5s] product for what you're working on at
[7s] work?
[8s] >> What's different about managing products
[10s] that are powered by AI is there's the
[13s] interface of how a user interacts with
[16s] any product or product feature and that
[18s] still really matters. And there's also a
[20s] lot going on behind the scenes. There's
[22s] a lot also about how do you drive good
[25s] quality products because these
[27s] technologies produce different results
[29s] each time you use them. So we start with
[31s] golden conversations. What's the
[33s] experience that you're trying to drive?
[35s] And so this is just a way for me to
[37s] think about how to write that role
[39s] playing a little bit with AI. What
[41s] you're saying is actually write an
[43s] example conversation that can represent
[46s] what a real user might do. and you're
[49s] working backwards from that example
[50s] conversation which I have actually not
[52s] seen anybody do before.
[54s] [Music]
[57s] Welcome back to how I AI. I'm Clarvo
[61s] product leader and AI obsessive here on
[63s] a mission to help you build better with
[64s] these new tools. Today we have an AI PM
[68s] showing us how to AI PM. Pria Matthew
[71s] Badger is a PM at Yelp and is showing us
[74s] a completely new way to think about
[76s] product requirements, prototyping, and
[79s] how to build effective conversational
[81s] agents using conversational agents.
[84s] Let's get to it. This episode is brought
[87s] to you by GoFundMe giving funds, the
[89s] zero fee daff. I want to tell you about
[92s] a new product GoFundMe has launched
[94s] called Giving Funds. a smarter, easier
[96s] way to give, especially during tax
[98s] season, which is basically here.
[100s] GoFundMe giving funds is the DAFF or
[103s] donor advice fund from the world's
[105s] number one giving platform trusted by
[108s] 200 million people. It's basically your
[111s] own mini foundation without the lawyers
[113s] or admin costs. You contribute money or
[116s] appreciated assets, get the tax
[119s] deduction right away, potentially reduce
[121s] capital gains, and then decide later
[123s] where to donate from 1.4 million
[126s] nonprofits. There are zero admin or
[129s] asset fees, and while the money sits
[131s] there, you can invest and grow it
[133s] tax-free, so you have more to give
[135s] later. All from one simple hub with one
[138s] clean tax receipt. Lock in your
[140s] deduction now and decide where to give
[142s] later. Perfect for tax season. Join the
[145s] GoFundMe community of 200 million and
[148s] start saving money on your tax bill. All
[151s] while helping the causes you care about
[152s] the most. Start your giving fund today
[155s] in just minutes at gofundme.com/howi
[160s] ai. We'll even cover the daff pay fees
[164s] if you transfer your existing daff over.
[167s] That's gofundme.com/howi
[172s] ai to start your giving fund. Priya,
[176s] welcome to how I ai. I am so excited to
[180s] have you here because whenever anybody
[182s] asks me and they ask me a lot, how do I
[184s] do AI product management? I have to say,
[187s] wait, are you talking about product
[189s] managing with AI? Because I have some
[191s] ideas about that. Or are you talking
[193s] about product managing AI products? And
[196s] what's really great about the
[197s] conversation we're about to have is you
[199s] actually do both. So what in your mind
[204s] is really different about product
[206s] managing products using AI?
[210s] >> Yeah, I'm really excited to be here. Big
[212s] fan of the show and have learned a lot
[214s] about um AI, both managing AI products
[217s] and how to use it in my day-to-day from
[220s] the podcast. So it's exciting to be
[222s] here. For me, I think you know what's
[224s] different about managing products that
[227s] are powered by AI is there's the you
[231s] know interface of how a user interacts
[233s] with a with any product or product
[236s] feature. Um and that still really
[238s] matters with AI products. Um and I'll
[240s] show some of the tools that we use um to
[242s] explore that. Then there's also a lot
[244s] going on behind the scenes that
[246s] determines the product experience for
[248s] the consumer. So um the system prompts
[251s] and how that guides the conversation
[253s] flow is really interesting and I think
[256s] kind of a new challenge when you're
[257s] working on AI powered products and
[260s] there's a lot also about how do you
[263s] drive good quality products because
[265s] these um technologies produce different
[268s] results each time you use them. So
[271s] there's a lot of um interesting
[272s] challenges there too. Yeah. So, I'm
[274s] really excited to myself learn from your
[277s] flow because I'm building an AI powered
[279s] product as well. And so, let's dive into
[281s] it. Where do you start when you're
[284s] thinking about designing and framing out
[287s] a AI product for what you're working on
[290s] at work?
[291s] >> Yeah, absolutely. So, I thought a good
[292s] example would be to talk about building
[295s] a new feature capability into our Yelp
[298s] Assistant. So that's the product I work
[300s] on. And the way it works is a consumer
[303s] can come in for a service need. So let's
[306s] say you want to hire a handyman, a
[307s] plumber, an electrician, somebody to fix
[310s] your car, and you can describe the
[312s] problem in your own words, and then the
[314s] AI will understand what you're saying,
[316s] collect some project details, and um
[319s] help you get matched to pros and get
[321s] quotes. And so that's how the product
[324s] works. And we recently launched a
[326s] feature that allowed consumers to upload
[328s] a photo to help describe their need. And
[331s] that just makes sense, right? It it
[332s] helps for pros sometimes to be able to
[334s] see a photo along with the description.
[336s] But one of the things we wanted to do
[338s] was because we're doing this in our AI
[340s] assistant, think about, you know, how
[342s] can we leverage those AI capabilities?
[344s] Can the AI understand what's in the
[345s] photo and customize the conversation
[348s] from there? Um providing, you know, some
[351s] recommendations around what the consumer
[353s] should do next. as a a Yelp user, I can
[356s] imagine that the variety of services
[359s] that your pros are providing and um you
[362s] know with I don't run consumer
[363s] businesses, but I can imagine the the
[365s] variety of things a user puts into these
[369s] conversational or image upload
[371s] interfaces could be very diverse. So I'm
[373s] curious how you approach that from a
[376s] product development perspective.
[378s] >> Yeah, absolutely. Yeah, we certainly
[379s] cover a lot of different categories of
[382s] service needs at Yelp and one of the
[383s] challenges is yeah, making sure that the
[386s] experiences work across all those
[387s] different use cases that a consumer
[389s] might have. Do you want to jump in and
[392s] uh I'll I'll show you my workflow.
[394s] >> Yeah, let's do that.
[395s] >> Okay. So, I'm going to just open up
[397s] Claude. And here we're starting in a
[399s] totally new window. And you know, as we
[402s] talked about, like I think there's, you
[403s] know, two pieces to these AI products.
[405s] There's the behind the scenes part and
[407s] then there's the interface. uh user
[408s] interface that consumers see. Um and I
[411s] like to start with thinking about what
[413s] is that conversation flow going to look
[414s] like when we add this new functionality.
[417s] And so I'm going to show you here how
[419s] you can do that with claude. Um and you
[421s] can also use chat GPT or any other um of
[424s] these foundational models. So here I'll
[426s] say write a complete um sample
[428s] conversation between the consumer and
[430s] the AI assistant um where we want
[433s] consumers to be able to upload their
[434s] photo and then just add some scenario
[436s] requirements like we want the assistant
[438s] to analyze the photo maybe provide some
[441s] suggested replies and uh continue that
[444s] back and forth until they have enough
[445s] info to submit quotes. One thing I'll
[447s] call out on the prompting is I do like
[449s] to give a little direction on what the
[451s] output looks like. So you can see here
[453s] I'm saying like use assistant colon user
[456s] colon for labels, write it as one
[459s] continuous conversation. I think that
[461s] really helps make sure that you know you
[464s] get the output that you're looking for
[465s] and there's a little less back and forth
[466s] with the AI. So for the folks listening,
[470s] one of the things I want to call out
[471s] that I think is really interesting about
[472s] this approach is you're sort of using a
[476s] example conversation as your first pass
[480s] wireframe for building a conversational
[482s] AI. So instead of saying like show me a
[485s] chat window and show me messages that
[487s] show up in these buttons, what you're
[489s] saying is actually write an example
[492s] conversation
[494s] um that can represent what a real user
[497s] might do and um you kind of give some
[500s] some constraints about what that
[501s] conversation could look like and you
[504s] give it some of the capabilities that
[506s] might be available during that
[507s] conversation and you're working
[508s] backwards from that example conversation
[510s] which I have actually not seen anybody
[512s] body do before. So I think it's a really
[515s] unique approach that product managers
[516s] out there working on conversational um
[519s] AI products including myself can really
[521s] take a lot of inspiration from. How did
[522s] you come to this idea? I mean was this
[525s] your like are you just a genius and
[526s] you're like this is the first thing that
[527s] we need to do or how did you come to
[529s] this idea?
[530s] >> No, I mean I think this is part of um
[532s] our standard alumowered playbook at Yelp
[535s] where we start with golden
[538s] conversations. What's the experience
[540s] that you're trying to drive? Um, and so,
[542s] you know, I think, uh, this is just a
[545s] way for me to like think about how to
[547s] write that, um, roleplaying a little bit
[549s] with AI.
[551s] >> Yeah. And I just want to call this out.
[552s] We're going to take a little side uh,
[553s] detour to just some product management
[556s] ideas, which is I often tell product
[559s] managers to prototype their product as
[562s] close to the end product that a consumer
[564s] is going to consume, including the
[566s] content. So when I worked in dev tools,
[569s] um I would tell a lot of RPMs, don't
[571s] write a PRD, write a quick start and
[574s] documentation guide to the product.
[576s] Write the code snippets. Um and then
[578s] work backwards into what the product
[580s] should look like. And so I love this
[581s] idea of just from a general product
[584s] perspective, work with the artifact
[586s] that's closest to what the consumer is
[588s] actually going to experience and then
[589s] you can back into all the requirements
[591s] once you're kind of inspired by what
[592s] that end state is. So, what does
[595s] something like this get you?
[597s] >> Yeah, absolutely. So, let's go through
[599s] it. So, I'm actually going to upload a
[601s] real photo of a home service need. So,
[604s] here's like a picture with a cracked
[605s] porch. Um,
[607s] >> not your cracked porch.
[609s] >> It's not. No. Um,
[613s] yeah. And then we'll look at what um
[616s] what Claude comes back with. Um, I will
[619s] say one of the pictures I'm going to
[620s] test is from my bathroom renovation. So,
[623s] you will see my bathroom. And one thing
[625s] I'll call out is Claude now shows you
[627s] your thought process. And you'll see
[628s] this in a lot of AI tools. I really like
[631s] to read the thought process and it's
[632s] also something to do while you're
[633s] waiting. Um, but I think it really helps
[636s] because you can see how it's
[637s] understanding you. If it doesn't come
[639s] back with what you want, it also is
[641s] really good for troubleshooting. So,
[643s] definitely something I recommend doing.
[644s] >> Yeah. One thing that I'll do while this
[646s] is loading is call out, I too think that
[649s] reading the reasoning or the thought
[651s] process of the AI is interesting for two
[653s] reasons. One, it can often help you
[656s] improve your prompts because you
[657s] understand what the AI is understanding
[660s] or not understanding about your prompts.
[662s] As somebody who likes misspelled, no
[664s] sentence, low syntax prompts myself,
[666s] it's good good to see where I'm
[668s] misleading the AI. The other thing is
[670s] the thought process is often where the
[672s] AI reveals its personality. I think it
[674s] is so funny
[676s] >> to read like Gemini 25's thought process
[679s] versus 03 versus Claude is very nice.
[682s] Claude practices self-love. Um Gemini 25
[685s] does not. And so I just think it's uh
[688s] it's also interesting from just like a a
[690s] model understanding perspective. Okay.
[692s] So we got a we got a chat here.
[694s] >> Yeah. So then we can read through the
[696s] chat and it's, you know, it's saying
[698s] like, I can see you've uploaded this
[699s] photo of a front front porch stabs with
[701s] a significant crack running through the
[703s] concrete. So pretty good recognition of
[705s] the photo. And then it says, let's ask,
[708s] let me ask a few questions, you know,
[710s] how urgent is this? You know, are you
[712s] looking to repair this? Would you prefer
[714s] to replace the entire steps? And so I
[717s] could look through this, you know, and
[718s] maybe workshop it a little bit, giving
[720s] it some feedback. I also find it's
[722s] helpful to just create some more
[724s] examples. Um sometimes like when you see
[727s] a lot of examples, that's when the
[728s] trends come out and that's when you see,
[730s] you know, what you might want to improve
[731s] or change. And so I have a bunch of
[734s] images now. So now that I've tested it
[736s] with one and I've seen that, you know,
[738s] it works pretty well with that one. I'm
[740s] now going to test it with a lot more
[743s] images. And this is the prompt I'm going
[745s] to use. So I'm going to say now create
[747s] more examples based on these images. And
[750s] to your point earlier, you know, Yelp
[752s] covers lots of different um types of
[754s] service needs. So, this is where you can
[756s] kind of test and see how's it going to
[757s] do across a lot of different problems.
[759s] And so, here I have, you know, like a
[762s] appliance repair issue with an error
[763s] code. I have a hornet swap, a wasp nest.
[767s] Um, so you can see, you know, a larger
[769s] variety of things. And just because I
[772s] know you really wanted to see my
[773s] bathroom, I will also upload and add a
[777s] picture of my bathroom renovation in
[779s] progress. Um, and then I'm going to say,
[782s] um, you know, label each conversation
[784s] with a title and a number at at the top.
[786s] So, just another example of how just
[788s] that like little nudge on the output can
[791s] really help you get something usable.
[794s] Great. And so we're going to see here
[796s] how this AI thinks about potentially
[799s] framing responses to consumers on a
[802s] variety of as a homeowner total
[804s] nightmare scenarios. Everything from a
[806s] wasp to a bathroom renovation, which I
[808s] am also about to start um is just a
[811s] nightmare to me whether or not I want to
[812s] do it. Um and so you're getting these
[814s] example conversations and what are you
[817s] looking for? Are you are you looking for
[820s] patterns? Are you looking for product
[821s] inspiration?
[823s] what's kind of the thing that you're
[824s] seeking in these examples?
[827s] >> Yeah, that's a great question and I
[829s] think this like goes in with, you know,
[832s] there's the the a lot of people talk
[833s] about like evals are the new PRD and
[836s] this is like the very early step of of
[839s] getting getting to the eval process. Um,
[841s] you know, I think you you get a sense of
[844s] like what are the criteria that are
[845s] important for this capability. So, you
[848s] know, the first thing is like did it
[850s] actually recognize the image? Well,
[852s] right. So I can compare and see like in
[854s] this first one like the oven door lock
[856s] malfunction where I've uploaded this
[859s] picture and it is actually looking and
[862s] seeing that like it has the door locked
[864s] and it's trying to understand that
[865s] issue. You know maybe we would give it
[868s] feedback to go one step further like
[870s] pull that E3 error code you know look in
[873s] your LM see if you uh understanding to
[875s] see if you can guess what the issue is
[878s] and and diagnose it better. Um but I
[881s] think that's like the first step of is
[883s] it um doing that recognition right and
[886s] then after that you know we're we're
[888s] looking through the conversation to
[889s] first I just look at it qualitatively to
[891s] see like does this feel like it sounds
[894s] uh like it flows well is it concise is
[896s] it easy to understand um and then we'd
[898s] probably develop like more of a rubric
[900s] around what are the criteria that we're
[902s] looking for
[903s] >> okay so you have these different
[906s] conversations what do you do with them
[908s] next Yeah, and I'll just show one
[911s] example of refining these conversations
[913s] and why AI is really great for this. So,
[915s] you know, let's say I say I I think it's
[917s] good, but I don't think it's being as
[919s] opinionated as it could be about like
[921s] offering the user a recommendation and
[923s] maybe sometimes it's talking about
[924s] budget, which we think the consumer may
[926s] not know. So, I can ask it to rewrite
[929s] these conversations based on this
[930s] feedback and it will go through and
[933s] update all those conversations for me,
[935s] which I think is really nice. And um you
[937s] know then you can go through and see you
[939s] know do you feel like it's taking that
[941s] feedback well? Is it actually rewriting
[943s] it um based on that guidance? But
[945s] definitely you know you can see here
[947s] it's saying like this definitely
[949s] requires professional pest control.
[951s] Don't attempt a DIY removal of this
[953s] nest. Um which I think is probably good
[956s] advice. Um,
[959s] and then to your other point about like
[962s] how do we get um an artifact that is
[965s] closest to the ex what the consumer will
[968s] experience that is the next step that
[970s] I'm going to show you and something I
[971s] think that is pretty unique to Claude.
[974s] Um, so Claude has a special
[976s] functionality built in where it actually
[978s] can create an artifact that uses the LM
[981s] that powers Claude to produce those
[983s] responses. And that's very unique to
[985s] Claude. If you did this in another
[987s] prototyping tool, you would typically
[989s] have to set up a API key and um
[993s] integration which just takes a little
[994s] bit more work and with pod you can do it
[997s] out of the box. So here you can see I'm
[999s] asking it to create an assistant app as
[1001s] an artifact have a chat interface where
[1004s] the AI responds using the LLM that
[1006s] powers Claude and then also create
[1008s] system in uh prompt that is based on
[1011s] these example conversations and then
[1013s] analyze these upload loaded photos and
[1016s] include a camera um icon in the input.
[1019s] And then I'm actually going to upload
[1021s] some um screen grabs of our current Yelp
[1023s] Assistant and indicate that it should
[1026s] use these attached screenshots as an
[1028s] example for what the front end should
[1030s] look like just so that it feels a little
[1032s] bit more real.
[1033s] >> Got it. So you really are using example
[1036s] conversations and just reference designs
[1040s] as your PRD here. And then what you
[1043s] called out that's unique about quad
[1045s] artifacts is it has fully integrated
[1047s] quad AI. So you can actually generate
[1050s] artifacts that do make native LLM calls
[1053s] to the anthropic API. So if you are
[1055s] prototyping little AI product out there
[1058s] um check out Claude because it just
[1060s] makes it a little simpler and you don't
[1062s] have to pass it a bunch of API keys.
[1065s] >> Yeah, absolutely. And you can see that
[1067s] it's writing the code here and at the
[1069s] top it actually wrote the system
[1070s] instructions. And I think this is also a
[1072s] really good way to learn because you can
[1075s] see that based on these example
[1076s] conversations, how is Claude translating
[1079s] that into system instructions. Um so
[1082s] it's, you know, mirroring some of my
[1084s] initial prompting and redirection around
[1086s] providing suggested replies, um not
[1089s] asking the user about budget. And so I
[1091s] think that's um really helpful. And then
[1094s] you can see it gives some examples from
[1096s] my examples as part of how to guide the
[1099s] um assistant around photo analysis as
[1102s] well. All right. And so I'm going to
[1105s] test it out and we'll see if it works
[1107s] out of the box. Um it does sometimes
[1110s] require a little back and forth.
[1112s] Um so you can see here I have uploaded
[1115s] the photo of my issue and Claude is
[1118s] thinking.
[1119s] Okay, great. Um so here you can see it
[1121s] worked pretty well. So it said, you
[1123s] know, I can see it's showing F2 in red
[1126s] and the door locked and this is a common
[1129s] error code relating to the oven lock.
[1132s] You know, typically you want a repair
[1134s] technician. It's asking about the
[1136s] urgency. So it is, you know, simulating
[1139s] pretty well this conversation. And one
[1142s] of the reasons why I think it's helpful
[1144s] to simulate it in this kind of artifact
[1146s] is you can also get a real feel of how
[1149s] this would be for the user. Like you can
[1150s] see like sometimes a response that looks
[1153s] fine when you have it in a doc feels
[1155s] really long when you see it in like the
[1157s] little chat bubble and the mobile
[1158s] interface
[1159s] >> and you know that waiting period of like
[1162s] the three dots and then the response
[1163s] comes back when you play out the full
[1165s] conversation
[1166s] >> can feel very different. So I think this
[1168s] is also a really good step to do
[1171s] >> and then you can of course share this
[1173s] with your team or your designers or your
[1175s] engineers and they can also start to get
[1177s] a sense of how does this feel? Can we
[1180s] actually do this? How can we refine it
[1182s] or make it even operate better? So I I
[1186s] just have never thought of this low. I
[1188s] have to repeat it again for folks. You
[1190s] know, kind of starting inside out with a
[1192s] conversational agent, prototyping
[1194s] example conversations first, getting
[1196s] them
[1197s] um refine getting a good set of example
[1200s] conversations that you can then put into
[1203s] a um prototype generating tool in this
[1206s] instance claude to then back into the
[1209s] chat experience including the system
[1211s] prompt that would best serve those
[1213s] conversations as such a great flow. I'm
[1216s] so impressed. This episode is brought to
[1218s] you by Persona, the B2B identity
[1221s] platform helping product fraud and trust
[1223s] and safety teams protect what they're
[1225s] building in an AI first world. In 2024,
[1228s] bot traffic officially surpassed human
[1230s] activity online. And with AI agents
[1233s] projected to drive nearly 90% of all
[1235s] traffic by the end of the decade, it's
[1237s] clear that most of the internet won't be
[1239s] human for much longer. That's why trust
[1242s] and safety matters more than ever.
[1244s] Whether you're building a next-gen AI
[1246s] product or launching a new digital
[1248s] platform, Persona helps ensure it's real
[1250s] humans, not bots or bad actors accessing
[1253s] your tools. With Persona's building
[1255s] blocks, you can verify users, fight
[1257s] fraud, and meet compliance requirements.
[1260s] All through identity flows tailored to
[1262s] your product and risk needs. You may
[1265s] have already seen Persona in action if
[1267s] you verified your LinkedIn profile or
[1269s] signed up for an Etsy account. It powers
[1271s] identity for the internet's most trusted
[1273s] platforms and now it can power yours
[1275s] too. Visit withpersona.com/h
[1280s] how I AI to learn more. You know, now
[1283s] what I have to call out is this looks
[1285s] pretty good, but it doesn't look quite
[1288s] like Yelp. So, how do you take this how
[1291s] do you take this to that next step of,
[1293s] you know, really um designing out what
[1295s] the real product might look like?
[1297s] >> Yeah, for sure. And I will say like I
[1299s] think this is all just a starting point
[1301s] and it's a part of a conversation with
[1303s] your larger team, right? With the
[1304s] engineers and with the with designers
[1306s] like I think this is really something
[1308s] that helps me clarify my own thinking
[1310s] and ideas and like refine what is that
[1312s] ideal conversation look like and and
[1315s] also just you know be a better
[1316s] collaborator because I understand system
[1318s] instructions better um as uh as we're
[1321s] going through features. Um but yeah, so
[1324s] I think um you know, it still goes
[1326s] through our our our usual like design
[1328s] and engineering pro uh processes once we
[1330s] have a good idea of you know where we're
[1332s] headed and it really has been a
[1334s] collaborative process for us between
[1336s] design, product and engineering where
[1338s] we're all writing these conversations
[1340s] together. We're giving each other
[1341s] feedback on them. Um so now we're going
[1344s] to I'm going to talk about you know how
[1346s] do we how do we think about the
[1348s] exploring ideas on the other side? So we
[1350s] we went pretty deep on like what does
[1353s] that conversation flow look like? How
[1354s] can we use cloud to um explore ideas
[1357s] there and the other piece is like how do
[1359s] use what does the interface look like?
[1360s] What are the user flows? How does a user
[1363s] get into these assistant experiences?
[1365s] And I have seen that a lot of those
[1366s] little details matter as well. You know
[1368s] what are the prompts? How how does a
[1370s] user understand the capabilities of the
[1372s] assistant? And so here with uh I'm going
[1375s] to show another tool which is magic
[1376s] patterns. And I think magic patterns is
[1379s] really great for when you want to
[1380s] explore something visually and like kind
[1383s] of consider what that flow would look
[1384s] like. I know Colin Matthews was on this
[1387s] show earlier and he showed how you can
[1389s] recreate a you know an existing product
[1391s] using component library or screenshots.
[1393s] So I'm not going to cover that in
[1395s] detail. So here I've recreated our Yelp
[1398s] Assistant um with that kind of approach.
[1400s] But I'm going to show you how you can
[1402s] then move on um to actually explore
[1406s] features within u magic patterns which I
[1409s] think is a lot of fun. So here I'm going
[1411s] to actually ask it to add a prompt
[1413s] suggestion at the top for start with a
[1416s] photo which allows the user to upload a
[1419s] photo. And you know you can see here
[1421s] it's it's thinking and it's saying I
[1424s] will start um add this prompt suggestion
[1426s] for start with a photo. um this will
[1428s] likely require these things. Um for
[1431s] styling, I'm going to consider this. So
[1433s] again, like reading those thinking
[1434s] instructions, I think is super helpful.
[1436s] So what it's doing now, now that it has
[1438s] those instructions, it looks like it's
[1441s] sort of doing this thing that you see in
[1443s] a lot of these prototyping tools, which
[1445s] is it's creating or updating new
[1448s] components, updating components. It's
[1450s] going to kind of insert those design
[1452s] elements
[1453s] into into this design for you to give
[1456s] feedback and test with. And I just have
[1458s] to say, you've been a PM for a little
[1460s] bit. I've been a PM for a little bit.
[1462s] Have you ever had access to this kind of
[1464s] like ondemand
[1466s] design and code? Like is has this
[1468s] totally like changed the way you think
[1470s] about working through designs,
[1472s] wireframes, stuff like that?
[1474s] >> Yeah, it absolutely has. Yeah, I think
[1476s] my mind was kind of blown to be honest.
[1478s] the first time I use these like natural
[1480s] language prompting prototyping tools
[1482s] just because yeah it's just so magical
[1485s] for you as a PM to be like hey I can
[1487s] just describe what's in my head and
[1489s] actually have it you know come to life
[1492s] um in a prototype. So it really has uh
[1495s] you know I think the core of the of the
[1497s] PM job and the earliest part of the
[1499s] workflow hasn't really changed and that
[1501s] you're still trying to understand deeply
[1504s] the user problem figure out what to
[1506s] prioritize. Um but I think it really
[1508s] helps in the phase after that where as a
[1511s] team you're exploring the solution
[1513s] space. What can really solve that
[1514s] problem for a user? How do we make them
[1516s] aware of it? How do we make sure it's
[1518s] easy to use? And I feel like it's just
[1521s] really fun to be able to like play
[1523s] around in these tools and explore ideas
[1525s] um myself visually and and find better
[1528s] ways where I can communicate something
[1529s] that's in my head.
[1531s] >> Amazing. Okay. So now we have a start
[1533s] with a photo.
[1534s] >> Okay. So yeah, we have a start with a
[1535s] photo. As you can see here, it's got
[1537s] this UI where I can start with a photo.
[1539s] Um so you know that's you know one
[1542s] option. And then of course like you know
[1543s] we did something simple when you launch
[1545s] this feature where there's just a camera
[1547s] icon but I'm showing this example as a
[1549s] way that you know you can explore like
[1551s] what would other ways be that we could
[1553s] make this experience um as you're
[1555s] thinking about iterating. And so here
[1557s] I'm going to show you this really cool
[1559s] feature within Magic Patterns which is
[1561s] called inspiration mode. Um and
[1563s] definitely recommend digging into this
[1564s] menu in general. Um, they have like a
[1566s] lot of nice little shortcuts, but this
[1568s] inspiration mode is my favorite because
[1571s] you can quickly explore lots of
[1573s] different options. So, here I can say,
[1576s] "Give me some options on how the start
[1578s] with the photo flow could work to make
[1580s] it feel more guided for the user." And
[1582s] this part of the prompt I workshopped a
[1584s] little bit, but I think works to help
[1586s] have the inspiration mode come up with
[1588s] different ideas. I say like think
[1590s] expansively and make each option
[1592s] differentiated and then explain in in
[1595s] your response which option um what each
[1598s] option is. Um and so I'm going to go
[1600s] ahead and submit that and it will
[1603s] generate for me four different options.
[1606s] And you'll see that um once it goes
[1609s] through this process, it will actually
[1611s] have four different boxes on the screen.
[1613s] And as you want to explore those
[1615s] options, you can click through those
[1616s] boxes and it'll update what's on the
[1618s] left side. So you can really quickly
[1620s] explore and see the different ideas and
[1623s] you know decide what you like. Um and I
[1626s] like doing this because I think
[1627s] sometimes we come in and we feel like we
[1631s] need to have a whole PRD before we can
[1633s] start prototyping. And that's definitely
[1635s] one approach and use case for AI
[1637s] prototyping tools. But I've also found
[1640s] that they're helpful even earlier when
[1642s] you you do understand your you know your
[1644s] user problem, what you're trying to
[1645s] solve for, but you may not know really
[1648s] what those solution looks like and you
[1650s] want to explore and maybe get some ideas
[1652s] from AI as well. Yeah, this just makes
[1655s] me think I don't know if designers are
[1656s] going to love this or hate this. I
[1658s] remember this experience when I was a
[1659s] designer where somebody would give me a
[1661s] purity or a feature like this and I
[1664s] would give them back a design like what
[1665s] we see on the left and they'd be like
[1667s] great but can we like try it over here
[1669s] and try it over there and move it up
[1671s] there and make it this button and like
[1672s] make it a link and that like manual
[1675s] iteration where it wasn't really um
[1679s] moving the product forward. It was kind
[1682s] of getting our own minds around what the
[1684s] problem space and the solution space
[1686s] could be so that we could move the
[1687s] product forward just took a lot of time
[1689s] and so I think it's really interesting
[1690s] to compress the time for ideiation so
[1693s] that you can get to the ultimate product
[1695s] a little bit faster.
[1696s] >> Yeah, absolutely. And like some of our
[1698s] designers are also using using magic
[1701s] patterns or even other AI prototyping
[1703s] tools like Figma has it Figma make and
[1706s] and so I think it's really just part of
[1708s] the conversation. you know, I'll ping a
[1709s] designer, hey, I was thinking about this
[1711s] and, you know, was thinking maybe we
[1713s] could go in this direction and send them
[1714s] a link and they'll be like, oh, I was,
[1715s] you know, exploring something similar
[1717s] and we'll just trade notes. So, to me,
[1719s] it's a replacement for what I was doing
[1721s] before, which was really hacky Figma
[1723s] mockups and like not so great
[1725s] wireframes. Um, and so I I think it's an
[1728s] extension of that like wireframing hacky
[1732s] Figma prototype process where it just is
[1735s] easier for someone to understand because
[1737s] they can actually click through and see
[1738s] the flow.
[1740s] >> Yeah, it's just more interactive I think
[1742s] is really it might not be higher
[1743s] fidelity, but it's a richer kind of
[1747s] prototype experience than you would get
[1748s] from sort of a flat design.
[1750s] >> Okay, we at least have three successful
[1752s] generations. We can click through
[1755s] >> with with with all AI, you know,
[1757s] sometimes you get errors, but you know,
[1759s] here it says it's like a guided category
[1761s] selection flow. So, we'll click through
[1762s] and see what they did. So, you can see
[1764s] here it's like kind of customizing it a
[1767s] little bit for the category of um of the
[1771s] service. So, I'm going to go back and
[1773s] maybe select another category and see
[1774s] how it's different. So, it's like, you
[1776s] know, kind of customizing some of the
[1778s] tips um in this one. Let's see. I might
[1781s] need to actually select a photo to see
[1783s] what it does. Um, so you can see it's
[1786s] like going through an analysis.
[1789s] You know, this is not using the LLM
[1791s] behind the scenes. So, you can see it's
[1793s] not uh not making sense, but I think the
[1796s] idea here makes sense where it's like,
[1798s] okay, it's going to do this like kind of
[1800s] real time detection. Um, and then in
[1802s] this one, it looks like it's like
[1804s] multiple photos. So you can see here
[1806s] it's you know showing like you know you
[1808s] could um prompt the user to maybe take
[1811s] multiple pictures. I will just click on
[1814s] this to show that you know this is how
[1816s] AI works or sometimes sometimes you get
[1821s] errors and you need to fix them. Um you
[1823s] know usually there's that like shortcut
[1825s] to like try to fix it. Um, if it doesn't
[1828s] work, um, there is also like a debug
[1831s] command within magic patterns, which I
[1833s] found pretty useful, which just tells it
[1835s] to like look through your code, try to
[1836s] come up with what's wrong to fix it.
[1839s] >> Um, let's see if it did fix it. For our
[1843s] listeners that are not wa not are not
[1845s] watching, I will spare you reading the
[1848s] uncaught react errors about um
[1851s] incompatible React versions. But that is
[1854s] what we are looking at right now, which
[1856s] is we are looking at a compatibility
[1858s] issue between 18 and 19.
[1861s] >> Yeah.
[1862s] >> All right. So like all good AI demos,
[1866s] this one did not work. But I do want to
[1869s] say just stepping back what I wanted to
[1871s] just call out is you have demoed for us
[1874s] a completely new way of thinking about
[1877s] product management prototyping and
[1880s] product requirements
[1882s] in a way that is very different than I
[1884s] think what classic product management
[1886s] has looked at. And so you're starting
[1889s] from a kind of example consumer
[1892s] experience first. you're backing into
[1896s] kind of a rough prototype of what could
[1898s] support that experience. You're using a
[1901s] AI prototyping tool, in this instance,
[1903s] magic patterns, to then put that
[1905s] experience in your brand and design
[1908s] guidelines. And then you're using that
[1911s] as a jumping off point to fork and
[1913s] inspire a couple different versions of
[1916s] what that ultimate user experience could
[1919s] look like. And then I'm presuming you're
[1921s] going to take one of these and you're
[1922s] gonna say I think we want to start here
[1924s] for our MVP or our V1 and then that you
[1927s] know you get the team together and then
[1928s] and then that's where you start. And so
[1931s] I think for the product people listening
[1933s] what I like about AI is it's not just
[1936s] multimodal and that you can put any sort
[1939s] of um file type or data type in. It also
[1943s] allows you to approach problems from the
[1945s] front door, the back door, the side
[1946s] door, the window. Like, you know, you
[1948s] can come at your product problems in a
[1950s] much less linear way. And in fact, you
[1953s] can start at the end, go back to the
[1955s] beginning, come to the middle, fork off,
[1957s] go back to the beginning, and
[1958s] reprototype. And it's not expensive,
[1960s] it's fast, and it's interesting. And so
[1963s] I think what you've inspired me to do is
[1965s] actually think a little bit differently
[1967s] about what the starting point of product
[1969s] management could be not just for AI
[1970s] products but for product in general. And
[1973s] then of course you showed some great
[1974s] ways that AI can help with that.
[1976s] >> Yeah, absolutely. Um and I will say yeah
[1979s] to your point you know you can pick
[1980s] which one you like the best um which you
[1982s] think fits your you know where you are
[1985s] um in your in your product journey and
[1987s] your user needs. Um, you can also like
[1989s] if there's one that feels like, hey,
[1992s] this like AI assisted one seems really
[1994s] interesting or this multifoto one seems
[1996s] really interesting, but maybe not like
[1997s] where we're going to go right away, you
[1999s] can fork this design and it will create
[2003s] um a totally separate window and chat
[2005s] for you um of just that variant and then
[2009s] you can just run off with that, you
[2010s] know, maybe on the side um while you're
[2012s] continuing down the original path you
[2014s] were in.
[2015s] >> I I love that. So we have seen your AI
[2019s] powered AIPM
[2021s] process and usually I would bump us to
[2024s] lightning round but part of our
[2026s] lightning round is going to have a
[2027s] couple demos in it. So as my first
[2030s] lightning round question can you do a
[2032s] quick world tour of a couple
[2035s] nonworkreated AI use cases that you
[2037s] think our listeners would really get a
[2039s] lot of value from?
[2040s] >> Yeah absolutely I can share a few
[2043s] personal examples also. Um so um one is
[2047s] you know I have started this um you know
[2051s] talk AI channel that was at Yelp which
[2053s] was actually inspired by a talk AI
[2055s] channel in Lenny's community and um I
[2059s] wanted to create a monthly newsletter
[2061s] that gets sent out that just summarizes
[2062s] all the great discussion and content
[2064s] that was being created there. And so um
[2067s] I'm just going to show an example of how
[2069s] to do that using Lenny's community. Um,
[2072s] and so here I have this um, set of
[2075s] project instructions that say, you know,
[2076s] I'm a community manager writing a weekly
[2078s] newsletter. Um, use these Slack
[2081s] conversations and format them just like
[2083s] the community wisdom newsletter. And
[2086s] then I think what's really cool is I can
[2089s] just come in here and I can say, you
[2092s] know, I want to just make a version of
[2094s] this community ver uh wisdom using this
[2097s] slack chat and I can upload the file of
[2102s] all those slack chats and I did
[2105s] randomize the names or um replace the
[2107s] names for privacy also using GPT. Um,
[2111s] and then you can see here it's going to
[2114s] make a version of that community wisdom
[2117s] newsletter just using those Slack chats
[2120s] and um, reuse that same format. And by
[2123s] using a project, I can, you know, save
[2125s] myself some time on the prompting.
[2128s] >> Great. So, you're copying and pasting
[2131s] um, like a week's worth of Slack
[2132s] conversations. M
[2134s] >> you're putting it into this cloud
[2136s] project which you've been given a um
[2139s] you've given a template and then you're
[2141s] having it generate on a weekly basis or
[2143s] whatever kind of a summary of what's
[2145s] going on in that community and other
[2149s] kind of like content that's being
[2151s] shared.
[2152s] >> Yeah, absolutely. And then you can see,
[2154s] you know, kind of follows that community
[2155s] with some uh format and pulls out what
[2158s] the top threads are. And so you might
[2160s] want to make some edits to this
[2161s] afterwards, but it really, you know,
[2164s] gets a really good first draft that you
[2165s] can then edit.
[2166s] >> Amazing. And you're probably everybody's
[2168s] favorite community member.
[2171s] >> Yeah, it's definitely a lot of fun um to
[2174s] yeah, see what people share. And then
[2176s] I'll show a couple other examples. So,
[2178s] you know, I showed the example of
[2180s] creating the Yelp Assistant and I
[2182s] actually used the same workflow to
[2185s] create this parent pal to explain how
[2188s] artifacts work to my husband and he was
[2190s] really excited about it. He was like,
[2192s] "Hey, like let's try it out with, you
[2194s] know, Tommy where Tommy throws toys down
[2196s] the stairs." So, you know, I did like,
[2199s] you know, my two-year-old um throws toys
[2202s] down the stairs and uh it's some the
[2206s] same kind of artifact where it's powered
[2208s] by Claude's LLM and it's going to ask me
[2211s] some clarifying questions like what's
[2213s] the trigger and it's like always at
[2215s] dinnertime when we are cleaning up. Um
[2218s] and then you can, you know, see how the
[2220s] AI will provide some parenting guidance.
[2222s] And I think the really fun thing for
[2224s] this is that, you know, you can build
[2226s] something that's just really for your
[2227s] own personal use case. Um, and it's a a
[2231s] really fun process to do that. I'll show
[2233s] one other one, which is um my siblings
[2236s] and I like to play this board game,
[2237s] Settlers of Katan. But the bad thing is
[2240s] it kind of takes a long time, especially
[2241s] if people don't go fast. So, I'm working
[2244s] on this Settlers of Katan timer where um
[2247s] I actually have a timer for me and my
[2250s] siblings and both for the setup and the
[2252s] main game play. But this one I actually
[2255s] built in Lovable because my siblings had
[2257s] a lot of feature requests about tracking
[2260s] the future uh you know who who's won
[2263s] over time and having a leaderboard and
[2265s] handicaps and all sorts of other ideas.
[2268s] So, I definitely think it's a lot of fun
[2270s] to prototype with AI for your personal
[2273s] use cases. And I know some PMs are like,
[2276s] "Hey, I really want to work on AI
[2277s] products, but I don't have that
[2278s] opportunity right now." I think the fun
[2281s] thing about these prototyping tools is
[2282s] you can build a use case that's just for
[2284s] you or just for you and a family member.
[2287s] Um, and learn a lot as you're doing it.
[2289s] You just gave me such a good idea
[2291s] because I don't play a lot of board
[2293s] games, but my kids get like 10 to 15
[2297s] minutes of Minecraft every day, but we
[2299s] only have one
[2301s] >> like uh time timer. Um, and so so I need
[2305s] an iPad where they can like both click
[2306s] their button and have it have it
[2308s] countdown. And then they're also really
[2310s] worried about fairness. So I will also
[2314s] use a uh relational database to store
[2317s] all their time
[2318s] >> and say I promise every week you are
[2321s] getting an equal amount of Minecraft.
[2324s] There is no no lack of fairness and then
[2326s] when they fight about it I'll use your
[2328s] parent pal GBT.
[2331s] >> I love it. Yeah, you can just direct
[2333s] them to check the dashboard.
[2336s] >> Amazing. Okay, last question and then I
[2338s] will get you back to all your
[2339s] prototyping and all your AI building.
[2342s] >> When AI is not listening, other than
[2345s] clicking that debug button in magic
[2347s] patterns, what is your tactic? What do
[2350s] you do?
[2351s] >> I I think that when AI is not working
[2354s] and you've already tried some of the
[2355s] debug um methods, I think it's helpful
[2358s] to actually think about the ways that AI
[2360s] is different than a human. Like often we
[2362s] just get in this chat and we're like,
[2364s] this is just like talking to someone.
[2366s] Um, but when you're hitting the wall, it
[2368s] it helps to like take a step back and be
[2370s] like, "This thing is actually not a
[2371s] human. Like, what could be going wrong?"
[2373s] And think about AI's limitations. And,
[2376s] you know, the ones that I try to keep in
[2377s] mind are it tends to lose context as you
[2381s] go through many different turns. And it
[2383s] has a limited context window. And so,
[2386s] when you start having a really long
[2388s] conversation with AI, sometimes it just
[2390s] goes haywire. And so the um methods I um
[2395s] recommend are if you're doing AI
[2396s] prototyping, you can use that fork or
[2399s] you know a remix to start a new chat
[2402s] with the context of that code and that
[2405s] actually resets the context window. Um
[2407s] so that's a good idea if you're going
[2409s] really far and deep with a prototype. Um
[2412s] and the same thing applies to a chat.
[2413s] Like if it's going haywire and you've
[2415s] had like a hundred back and forths, you
[2416s] can ask the AI to summarize the chat and
[2420s] the context and start a new chat.
[2422s] >> You gave me such a good idea with your
[2424s] last two answers because I am going to
[2426s] prototype a parenting pal for the
[2429s] relationship between me and my a my AI.
[2432s] >> Be like, AI parenting pal,
[2435s] >> my my 4-second old AI is no longer
[2438s] listening to me. What do what do I do?
[2442s] Um, that's that's really great really
[2444s] great feedback. And yes, reminder, AI is
[2446s] not human until the AI overlords take
[2448s] over and then you can be whatever you
[2450s] want.
[2450s] >> All right, Priya, this was such a
[2453s] practical, super useful, inspirational
[2455s] conversation. Where can we find you and
[2458s] how can we be helpful?
[2459s] >> Yeah, you can find me on LinkedIn and
[2461s] then I also have a Substack called
[2463s] almostmagic.substack
[2465s] where I share some prototyping tips and
[2468s] other tips about building AI products.
[2470s] >> Amazing. Well, thank you for sharing and
[2472s] joining How I AI.
[2474s] >> Awesome. Thanks so much for having me.
[2477s] Thanks so much for watching. If you
[2478s] enjoyed this show, please like and
[2480s] subscribe here on YouTube, or even
[2482s] better, leave us a comment with your
[2483s] thoughts. You can also find this podcast
[2486s] on Apple Podcasts, Spotify, or your
[2489s] favorite podcast app. Please consider
[2491s] leaving us a rating and review, which
[2493s] will help others find the show. You can
[2496s] see all our episodes and learn more
[2498s] about the show at howiipipod.com.
[2501s] See you next time.