The exact AI playbook (MCPs, GPTs, Granola) that saved ElevenLabs $100k+ & helps them ship daily

1. 0:00 – 2:41

   Intro

   1. LHLuke Harries0:00

      When you're editing, as much as possible, try and edit the underlying prompt rather than the actual output.

   2. CVClaire Vo0:05

      I like that you have here the, "You are a," and then gives a very specific identity and job to be done at the top of this. And then you have very specific instructions where you say, "You must do A, B, C, D," and it's quite particular.

   3. LHLuke Harries0:22

      This saved us $40,000 a year for the tool, so immediately canceled it. Over $100,000 in agency costs. I think the highlight, though, is just not having to deal with more SaaS vendors, more agencies, constantly trying to get upsold. [upbeat music]

   4. CVClaire Vo0:38

      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 great conversation with Luke Harries, Head of Growth at ElevenLabs. Luke shows us how to make everything a launch by making everything automated with AI. He shows us his secret flows for generating case studies and tweets on the fly, how he saved his company tens of thousands of dollars by, yes, being a marketer that coded in Cursor, and explains what an MCP is and how he hooked it up to WhatsApp. Let's get to it. This episode is brought to you by Orkes, the company behind open source Conductor, the platform powering complex workflows and process orchestration for modern enterprise apps and agentic workflows. Legacy business process automation tools are breaking down. Siloed low-code platforms, outdated process management systems, and disconnected API management tools weren't built for today's event-driven, AI-powered, cloud-native world. Orkes changes that. With Orkes Conductor, you get a modern orchestration layer that scales with high reliability, supports both visual and code-first development, and brings human, AI, and systems together in real time. It's not just about tasks. It's about orchestrating everything: APIs, microservices, data pipelines, human-in-the-loop actions, and even autonomous agents. So build, test, and debug complex workflows with ease. Add human approvals, automate back-end processes, and orchestrate agentic workflows at enterprise scale, all while maintaining enterprise-grade security, compliance, and observability. Whether you're modernizing legacy systems or scaling next-gen AI-driven apps, Orkes helps you go from idea to production fast. Orkes, orchestrate the future of work. Learn more and start building at orkes.io. That's O-R-K-E-S dot I-O.

2. 2:41 – 4:22

   The future of AI in marketing

   1. CVClaire Vo2:41

      Hey, Luke, thanks for joining.

   2. LHLuke Harries2:43

      Thanks for having me.

   3. CVClaire Vo2:44

      In 2025, we've talked a lot about vibe coding, Cursor this and v0 that, but we have not talked enough, I think, about vibe marketing. So what do you think the future of an AI CMO is in the next couple years?

   4. LHLuke Harries3:02

      There's all these tools like Lovable and Cursor, and the rate of software production's going to go exponential, but it's not gonna matter if no one's actually using your tool. And so what's important is actually getting the product into market, getting people to know about your new features. At ElevenLabs, we have this launch process. So basically, every new feature we do, every new model, we run it through this massive checklist, which is like, okay, we need to first work out what are the value props? Then we need to work out what's the core messaging. Then we need to work out who's it for. Then we need to turn that into the blog post, the X post, and it, it takes a lot of time, these massive launch processes. And so the thing I'm really excited for the AI CMO is being able to go from every single new feature or new product and translating that into your entire launch process, making the assets, making the videos, making the images, but then also going beyond the launch. So what are those then evergreen channels that you'll be testing? And so, uh, let's say ElevenLabs, we launched the best speech-to-text model. Okay, we need to be running Google Ads for that. So then it will spin up, understand all the various keywords, spin up the Google Ads. It will optimize the landing pages. So I think this entire thing is gonna change massively, um, and we're already using a few of these different workflows, and I'm excited to

3. 4:22 – 12:10

   Using Granola and custom GPTs to write case studies

   1. LHLuke Harries4:22

      talk you through them.

   2. CVClaire Vo4:23

      Yeah, and I think one of the best ways that companies can market is actually just telling great customer stories, and I know you have a workflow for getting case studies out. You're lucky enough to have probably tons of customers that love you. So can you walk us through how you use AI to make case study writing really easy?

   3. LHLuke Harries4:43

      With case studies, you know, you sign a great customer, and you now want to be able to tell the story about how they actually use your product. And so what we're gonna do live, Claire, is I'm going to have you write a case study for ElevenLabs. I know you've used ElevenLabs.

   4. CVClaire Vo4:56

      I have.

   5. LHLuke Harries4:56

      And what we're going to be using is the tools Granola, which is a fantastic transcription tool. It's a note-taking tool. Uh, and we're going to use ChatGPT, actually, with a custom GPT, which will then, uh, translate that into an excellent blog post, and as a bonus, we'll write a tweet in a, in our company voice ins- as well.

   6. CVClaire Vo5:16

      Okay, I'm excited, and I am a happy ElevenLabs customer, and you didn't pay me to say that, so [chuckles]

   7. LHLuke Harries5:23

      I, I did not. Um, so what I'm going to do is I'm going to open up Granola, and I'm going to do a case study interview with Claire. I'm going to just ask you two quick questions, and then we'll use the transcript for this. So, Claire, fantastic to meet you, and I would love to understand how you use ElevenLabs in your work.

   8. CVClaire Vo5:43

      Yes. So at, uh, my company, we have to produce a lot of customer-facing live events. It's the way that we connect with developers and customers in our community, and those live events can be either in-person events or they can be recorded, streamed events. And we put a lot of preparation into the messaging and the way we present our products at those events. So ahead of our user conference that's coming up in a couple weeks, we actually build a script for what our product keynote will look like, and it's me, it's our CEO, it's our SVP of product, it's engineers demoing, it's a Q&A with customers, and we like to run through and rehearse those keynotes.... and the rehearsals are very expensive from a time perspective. I just named ten people that have to be in the room. We have to walk through the script, we have to record it and then listen to it later, and we're really trying to nail a very specific set of timing. You know, we only have thirty minutes or so to get all this content in. And if you're participating in the dry run, it's actually really hard to listen for, is this good as a third-party observer of this keynote? So one of the things that I do with ElevenLabs that I find super useful is actually build prototypes of keynotes. So I load in the script into the, I think it's called the Studio Flow, and I give everybody in the company and our customers different accents. My boss has this lovely British accent, so [chuckles] I give him a British accent. I give myself a different accent. I pick voices for everybody in the keynote, and then I actually generate a proto- like, an audio prototype of the keynote and send it around for people to listen to for two things. One, timing, to make sure, do we have enough content? Do we have too little content from a timing perspective? And then, two, does it narratively flow and sound natural and is easy to understand and listen to, because it's a virtual event. And I found that that little flow, which I guess for the How I AI listeners, is also [chuckles] a little-

   9. LHLuke Harries7:44

      [chuckles]

   10. CVClaire Vo7:44

       ... a, a mini How I AI-

   11. LHLuke Harries7:46

       We've got a bonus

   12. CVClaire Vo7:46

       ... built-in, built into the flow-

   13. LHLuke Harries7:48

       Yes

   14. CVClaire Vo7:48

       ... has been really useful for us to make sure we get high-quality events going.

   15. LHLuke Harries7:51

       Fantastic. And so what I do is, you know, basically chat through in more-

   16. CVClaire Vo7:55

       Mm-hmm

   17. LHLuke Harries7:55

       ... detail, and I try and get out some, some concrete, concrete facts, the use cases.

   18. CVClaire Vo8:01

       Yeah.

   19. LHLuke Harries8:01

       Claire's already said she's using the Studio product, and, and maybe you would even go as far as j- using voice cloning of your different customers. So we may-

   20. CVClaire Vo8:10

       Yep

   21. LHLuke Harries8:10

       ... include that in the case study, and maybe you could give me a metric. So how has this driven ROI for your company? How has this doubled the revenue of LaunchDarkly?

   22. CVClaire Vo8:18

       Yeah, so I, I think it saves a significant amount of time internally, so there's definitely hours and hours saved in terms of iterating on something like a keynote. And then, if we make it high quality, then our customers hear more about our great products, and then, of course, we get to sell more.

   23. LHLuke Harries8:35

       Yeah, fantastic. So... And maybe even you, you say a statistic like, "It saves you, uh-

   24. CVClaire Vo8:40

       I'll say-

   25. LHLuke Harries8:41

       ... five hours of meeting prep time."

   26. CVClaire Vo8:43

       Yes. Yes, it saves me ten hours of meeting prep time. Let's make this one a good-

   27. LHLuke Harries8:47

       Perfect

   28. CVClaire Vo8:47

       ... case study for you. [chuckles]

   29. LHLuke Harries8:48

       [chuckles] Great. Uh, so then I click Stop Transcript, uh, and generate the notes.

   30. CVClaire Vo8:54

       So all this time, Granola is sort of recording what we're saying and, um, analyzing it on, on the back end. So now what we're getting is this auto summary based on all the stuff that I just said.
4. 12:10 – 13:58

   Generating tweet threads using ChatGPT

   1. LHLuke Harries12:10

      that live on our blog. The other thing then is, if you think about... I love to treat everything I can as launches. So if you think about your case study as a launch, like, first of all, you have to write it, but then the distribution really matters. And so what I've also done in this GPT is tell it what a great tweet looks like. Even the aesthetics of a great tweet. Like, "Okay, come up with a hook line, and then do a few, like, either bullet points or a short paragraph, then do an image." And so if I go write it as a... "Write a tweet thread for this," it will then-... rewrite that into a tweet summary, and it writes these handy brackets around-

   2. CVClaire Vo12:52

      Oh

   3. LHLuke Harries12:52

      ... like, what should the assets actually be? Uh, this is the bit where I don't think we have a full AI, AI CMO just yet, which is I'm really excited about the imra- image generation models of the new GPT models, because then I think you will actually be able to do, like, end-to-end launches and case studies by pulling this in.

   4. CVClaire Vo13:11

      Yeah. So this, this tweet generation chat right now not only writes the content, but actually puts these placeholders of what media you would need to make an effective tweet. So a screenshot-

   5. LHLuke Harries13:24

      Yes

   6. CVClaire Vo13:24

      ... or, or something like that.

   7. LHLuke Harries13:26

      Yeah.

   8. CVClaire Vo13:26

      And so here you've gone from, I don't know, we spent three minutes where I, I blabbed a little bit about a, a use case to a very polished case study, a tweet, and then I'm presuming you're gonna do three or four other, [chuckles] other things off this one, one asset pretty quickly.

   9. LHLuke Harries13:43

      Yeah. So then you would do the LinkedIn post. Uh, you maybe would write it in the style of... You could have a, a GPT in the style of your founder's tone of voice.

   10. CVClaire Vo13:53

       Yep.

   11. LHLuke Harries13:53

       So then you basically paste it into that, and he's got the asset, too. Um, and then

5. 13:58 – 15:14

   Building case studies into a systematic workflow

   1. LHLuke Harries13:58

      the way I'd always zoom out and think about this is: how do you actually put this into a workflow? So I think the best growth systems, you can do these one-off efforts, but things get busy, you know, your time gets taken up, and really, how do you build it into a system? And so concretely, set up a Zapier such that each time you get a closed won deal in your Salesforce, it sends them an email with your Calendly link, and you just get booked. Fantastic, different customers, maybe it's a month into their, into their contract, where all you have to do is rock up, have a nice chat with them. You can even get GPT to summarize or pre-prepare what the different bullet points and topics you should cover. Chuck it through your Granola and then ChatGPT flow, and you'll, you'll be turning out five case studies a month in no time.

   2. CVClaire Vo14:44

      This is a great flow because I often find things like case studies or little marketing assets are easy to make, but you have to remember to do them, and if they take time, you, you know, you get put in a meeting, or you have to pass it to somebody else, and you just sort of forget, and you slow down the next steps, and then you produce less assets. So I think it not only makes it easier to produce the assets, but it makes sure that that engine keeps going 'cause you, as a human, are not responsible [chuckles] for that,

6. 15:14 – 19:39

   Best practices for prompt engineering

   1. CVClaire Vo15:14

      that next step.

   2. LHLuke Harries15:15

      And a common theme, uh, I think we touch on in, in one of the next examples as well, is like, when you're editing, as much as possible, try and edit the underlying prompt rather than the actual output. And so if you're like, "Ah, it always does headings, which don't..." You know, maybe they're not particularly strong, or I like more numbers, or I like more concrete stats, make sure to incorporate that back into the underlying prompt.

   3. CVClaire Vo15:40

      And on that point, can we go back to the, the GPT just for a quick minute?

   4. LHLuke Harries15:44

      Yes.

   5. CVClaire Vo15:44

      'Cause I'd like to call out some things that I think you do pretty well here in, in the prompting that I think folks can learn from. Okay, so I- from, from a prompt perspective, you know, very commonly, everybody starts with the "You are a..." And so I like that you have here the, "You are a..." and then gives a very specific identity and job to be done at the top of this, basically making sure that copy that comes out of ElevenLabs matches, matches the strategy or matches the, the tone of voice and the brand. And then you have very specific instructions where you say, "You must do A, B, C, D," and it's quite, quite particular, which I think is nice. Um, some folks I know love very general prompting, but I find that if you have a point of view of what your tone of voice should be, this sort of, like, very precise formatting prompting is, is very important. And then you've broken down those instructions by types of content generated. So you have instructions for tweets, instructions for blog posts, and then the last thing at the end, um, which I also think people underutilize, is good examples. And I- I have a question: Do you use any bad examples in here, or is it, is it all good examples?

   6. LHLuke Harries17:00

      I mean, they were comically bad when I, [chuckles] when I was coming up. We do actually use, uh... For the next workflow I'll show you, we do actually use bad examples for that, for translation. But, you know, a bad blog post is clearly a bad, [chuckles] a bad blog post. Um, and actually, if I was... One extra thing I found, sometimes I think, you know, to draw back the learning from the Granola team, like, give it as much context as possible. So if I was to extend this further, I think it would- I would give it a lot of information around, like, what's the core messaging for each different product that we want to get across and really nail? And then it knows when I'm doing the different interviews, "Ah, the studio product, we really want to em- emphasize how you can do, like, multi-dialogue, complex speech," and then it would draw that out, too.

   7. CVClaire Vo17:49

      Yeah, the other thing that I think people worry about is that, like, AI on top of AI on top of AI becomes very lossy, and I like the idea that you use the Granola summary, but then you also use the raw transcript. So then you have both sort of the, the high-level summary, as well as some raw context, and because these contact windows are so big, the chat can make sense of it.

   8. LHLuke Harries18:12

      And without doing the raw transcript, you wouldn't get any of the lovely quotes as well-

   9. CVClaire Vo18:16

      Yeah

   10. LHLuke Harries18:16

       ... from what the customer exactly says.

   11. CVClaire Vo18:17

       Oh, gosh, I didn't even think of that. Okay, well, this is... I'm gonna steal this workflow. This is so-

   12. LHLuke Harries18:23

       [chuckles]

   13. CVClaire Vo18:23

       ... so great and so fast, and, uh, I love your philosophy of everything is a launch. So that's a, that's a really good way to, to think about things. 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. 19:39 – 25:10

   Building a custom translation system that saved $140k

   1. CVClaire Vo19:39

      So I know you had another use case where you were using a external tool or some sort of tool, and you actually just built a solution that saved the company quite a bit of money.

   2. LHLuke Harries19:51

      Yeah. So this one was, um, for ElevenLabs, we're in a whole bunch of different countries, and it's very important to us that we localize all our content. And so we want our homepage to be in Hindi, in Spanish, in German, in Polish, in Japanese. Um, and I set out about this process of how do you go about localizing the website? And I spoke to loads of the top experts, and apparently what you're meant to do, you set up a very expensive localization tool. Uh, so the one we chose, I won't name the name, but it was $40,000 a year, uh, and it quickly went up, or they kept on trying to push it up. So it's forty- you're now paying $40,000 a year for this tool. And then the tool, you then need lots of humans inside to actually do all the translation work. So then you're getting agencies, which you're paying about $100,000 for, and, and we set up this flow. There was actually quite a lot of engineering work to connect it to our CMS, to connect it to our code base. Um, and I was like, "Okay, fantastic. I've done all that, uh, but the AI translation is terrible." And then we found out the agencies and the humans were terrible because you're constantly playing this cost game, um, of trying to minimize the cost, so you can't get anyone who's any good. And meanwhile, we have AI, which is, like, utterly taking off. And I had this situation where my ke- team kept on sending me screenshots back from ChatGPT, being like, "Oh, no, this one's better instead." And I'm like, "Well, if we're just using ChatGPT for the reference of what's better, why don't we just use ChatGPT for the whole thing?" And, and so I've got this, uh, Figma board, where I've kind of laid out in a bit more detail what we started with and what we went to, but basically, we ripped out this entire tool, all of the agencies, wrote a very small server, where all it does is take the string, has a prompt per language, explaining what the tone of voice is for that language and the context, sends it back, whether that's into GitHub or Payload. And this saved us $40,000 a year for the tool, so immediately canceled it, over $100,000 in agency costs, and previously, we were waiting days to get the translations back. Whereas this is now instant, and if anything is very sensitive, like, say, our pricing page, we just have one of our team, so we're already a decently large team, one of our team just to do a quick sense check, and if, if anything's wrong, again, we update the prompt rather than the source code as much as we can to make it, make it better.

   3. CVClaire Vo22:27

      So you just re- replaced this tool. And what I think is so interesting is I have this debate internally, as somebody who provides SaaS software, you provide [chuckles] SaaS software-

   4. LHLuke Harries22:39

      Yeah.

   5. CVClaire Vo22:39

      ... and I, uh, I think one of the existential threats in the SaaS industry-

   6. LHLuke Harries22:43

      Yes

   7. CVClaire Vo22:43

      ... is the cost of building going to zero. And I talk to so many people, and they say, "Ugh, teams will never build this inter- like, why would you?" But then what I think you're showing is it, it can actually be quite cost-effective and improve quality-

   8. LHLuke Harries22:59

      Yes

   9. CVClaire Vo22:59

      ... to think about building these tools yourselves. And did you- I mean, who built this? Was it the engineering team? Like, who actually-

   10. LHLuke Harries23:04

       So this, I did the first 90%-

   11. CVClaire Vo23:08

       [chuckles]

   12. LHLuke Harries23:08

       ... in one day, and then I got one of the engineers to help it. So I literally built it all in Cursor. I was actually ill, so I was meant to be skiing. I was lying in bed, and I was having to deal with the fact that we'd just gone through three different agencies who didn't meet our quality bar for translation.

   13. CVClaire Vo23:27

       Yeah.

   14. LHLuke Harries23:27

       And I was like, "I cannot be bothered-

   15. CVClaire Vo23:29

       [chuckles]

   16. LHLuke Harries23:29

       ... to get a fourth agency. I mean, I'm just gonna rewrite it all." Uh, so I did the bulk, and then, um, yeah, one of the fantastic engineers on our team, he helped, uh, helped get it into production. I think the highlight, though, is just not having to deal with more SaaS vendors, more agencies-

   17. CVClaire Vo23:45

       Yeah

   18. LHLuke Harries23:45

       ... constantly trying to get upsold. And my broader- the broader question of, like, is all SaaS dead? I think no. But I think human-in-the-loop SaaS, like-

   19. CVClaire Vo23:56

       Yeah

   20. LHLuke Harries23:56

       ... if your job is about putting low-skilled workers in some sort of flow, which translation is, I think that's very risky because just the AI... And, like, at the moment, we still do need a little bit of, like, every week or two, we have someone just give a quick scan. Is it all great? And they do make little tweaks, but, you know, give it two years' time, I would much rather bet on AI costs getting cheaper and the quality going up rather than paying for more agencies.

   21. CVClaire Vo24:22

       So maybe I have three, three takeaways here. One is, you really should reconsider looking at build versus buy on some things, especially if you're not satisfied with the quality of the buy. It's worth, worth the investment. So I think that's, that's thing one. Thing two is, look out, your marketers are gonna hop into Cursor [chuckles] and get it 90% done and then hand it to you engineers. So you might as well do it yourself.

   22. LHLuke Harries24:49

       They'll be, they'll be making more work.

   23. CVClaire Vo24:51

       Yeah, they'll be making more. And three, um, uh-... do you know how many software products I have built out of the frustration that I'm supposed to be on vacation-

   24. LHLuke Harries24:59

       Yes

   25. CVClaire Vo24:59

       - but I am actually sick? That is, like, the perfect, the perfect time to get something new done. Um, so winter season is, is a highly-

   26. LHLuke Harries25:08

       Yeah

   27. CVClaire Vo25:08

       - productive [chuckles] season for me 'cause

8. 25:10 – 29:47

   Open sourcing the translation solution

   1. CVClaire Vo25:10

      I'm always sick.

   2. LHLuke Harries25:11

      You ne- you need to get them in the multi-year deals, not ending in a holiday.

   3. CVClaire Vo25:14

      I have to ask, do you feel like this is... Should we, should we, SaaS eats its own tail? Would you ever productionize this and sell this to others? Is it very custom-

   4. LHLuke Harries25:22

      No, I think we're gonna open source it.

   5. CVClaire Vo25:24

      Oh, cool!

   6. LHLuke Harries25:24

      So there's a couple of, there's a couple of things that I'll just show you. Um, if enough people in your YouTube comments-

   7. CVClaire Vo25:31

      Okay

   8. LHLuke Harries25:32

      ... uh, say to open source it, we'll open source it. But just to add a... You know, this basically summarizes what it was, is you have all your code in GitHub. You have your, like, strings, and then you push it into the SaaS tool, and the biggest issue, truthfully, was they didn't allow you to edit the prompt of the AI. And so there was just no way the translation would be any good, understand your brand language, understand the glossary, and they quoted us about six months before they were gonna ship the ability to edit prompts. And I think it's because the whole business model is based on no, "Get these humans in it." Um, and so instead, what we shipped was this, whereby it's just a GitHub Action, which understands, runs every time you changes, change the keys in your translation dictionary, sends it to an LLM with a prompt per LLM and saves it back, and that works way better. Um, and then the same with our CMS, we just built into it a, uh, translate button, which again, just sends it in, and it was so nice just having one source of truth. All the text is either in your... Well, two, I guess. [chuckles] Ei- either in your CMS for all the blogs, or it's in the code base for all your core pages. Also, another shout-out, this is, this is not a paid advertisement to Cursor, but we wrote this Cursor rule which does all the lift. One of the things with the engineers was like, it's quite annoying to have to extract all your strings, and so we wrote this one Cursor rule where you just say, extract, like, translate the strings, and it grabs them all into this en.json file, nicely wraps it, handles server-side or, or client-side rendering, so that was pretty fun.

   9. CVClaire Vo27:12

      So I'm making the first request.

   10. LHLuke Harries27:14

       Yes. [chuckles]

   11. CVClaire Vo27:14

       Everybody in the comments, join me as well. I would like this open source because I would-

   12. LHLuke Harries27:19

       Yeah

   13. CVClaire Vo27:19

       ... 100% use this, use this flow. And, you know, I think people get scared when you hear, you know, human-in-the-loop is out, but I do think there's this opportunity for folks to operate at a s- at a higher level of their craft. And so, you know-

   14. LHLuke Harries27:35

       Yeah

   15. CVClaire Vo27:35

       ... this is not the fun part of translating, is taking string A into string B. Then you can start doing things like, Does this match localized style? Is this appropriate? Is this how we wanna talk to our customers in this region? And so I do think there's this ability for humans to then add a layer of quality, um, and use of their intellect and skills that is higher level than, than this.

   16. LHLuke Harries27:59

       Yeah, and they're really cool. So we spoke through the Cursor rule. The GitHub, GitHub Action is here, which is basically instantly on each push, it generates that. But exactly as you said, we just have this prompt file per, um, per language, which talks through, again, our brand guidelines, translation style, keywords, and the cool thing is, you can define that, and, like, you can really take the care and the nuance exactly how you want to represent your brand, and then you'll be pretty confident that that's then, uh, scaled up across any content you're then putting out in the future. And before, truthfully, we weren't planning on translating, you know, every blog page, but now, like, you actually can do, and it's a much better experience for your users.

   17. CVClaire Vo28:43

       Have you tested putting these language-specific prompts in the na- in the language itself?

   18. LHLuke Harries28:49

       So we explicitly decided not to-

   19. CVClaire Vo28:53

       Mm-hmm

   20. LHLuke Harries28:53

       ... because I wouldn't have then been otherwise able to vet-

   21. CVClaire Vo28:56

       Yeah

   22. LHLuke Harries28:56

       ... that we were consistent.

   23. CVClaire Vo28:58

       Yeah.

   24. LHLuke Harries28:59

       Um, but I did... I'm not sure if you saw that tweet I did, but I did say that recently over [chuckles] basically, someone asked a doc system a question in Arabic, and it replied, "Well, because the docs are in English, I'll reply in English." It's like, no, you, you do definitely want it to reply in the language, uh, the language you want.

   25. CVClaire Vo29:18

       I love this, and I love that you built, built most of this. And is the maintenance cost very high?

   26. LHLuke Harries29:23

       There's none. There's none.

   27. CVClaire Vo29:25

       Oh.

   28. LHLuke Harries29:25

       The, um, [chuckles] well, it hasn't broken so far. Um, I mean, it's just because it's just a GitHub Action, which is updating the JSON strings.

   29. CVClaire Vo29:35

       Yep.

   30. LHLuke Harries29:35

       So, um-
9. 29:47 – 38:07

   Building a WhatsApp MCP

   1. CVClaire Vo29:47

      give yourself a case study. Okay, and then we saved, I'm not saying the best one for last, but I love this last example. Uh, one, because you get to explain to the audience what an MCP is, for those that are still confused by the concept, and two, I think you built a really fun one. So you wanna talk about your WhatsApp MCP?

   2. LHLuke Harries30:06

      What an MCP is, it's a Model Context Protocol, and so it's a protocol, uh, written by Anthropic, which enables anyone to expose tools to AI agents. And so the example, and why I built the WhatsApp one, was we all get tons of messages. We're all in tons of different WhatsApp groups, and it's really hard to stay on top. But currently, if you ask a tool like Claude, it has no idea about any of your WhatsApp messages. It can't help you out. And so what... Again, [chuckles] actually, that same weekend that I was ill, I did one- I did the Saturday was, uh, bricks out the, uh-... ripped out our translation software. And then the Sunday was, okay, can I actually connect WhatsApp to my AI system using an MCP? And it's pro- part of my broader thesis of basically like, I think a personal AI assistant really only needs your WhatsApp, your calendar, and your email, and then it knows everything about you, and it can even organize tasks, send emails, you know, organize dinners, dinners with your, with your friends. So, uh, that's why I built it. And there's also some, like, cool, um, use cases for work as well, uh, which we can jump into. So let me show you. Cool! So this is the WhatsApp MCP repo, and how it works is it has two main parts. So it has, um, a bridge, which it pretends... You know WhatsApp web, when you sign in, you scan that barcode? That's exa- exactly how it works. So it actually pretends to be WhatsApp web, and, and it uses a fantastic library called WhatsMeow to do this. Uh, so you- when you run it in your terminal, you scan your barcode, and, and the first thing it does is it downloads all your messages onto a local computer, saves it in a SQLite database. And, and what that means is, you can then keep querying it as much as you want with an AI, uh, and you have no risk of, or very low risk of being banned because it's only downloaded it once. So, yeah, to be fair, this is unofficial, uh, unofficial stuff. Um, and then the other bit's the WhatsApp MCP server, which basically gives the ability for your AI to query this, uh, SQLite database, as well as, like, sending messages, sending voice notes. So if we jump over to Claude-

   3. CVClaire Vo32:30

      Mm-hmm. And this is the desktop Claude-

   4. LHLuke Harries32:32

      Yes

   5. CVClaire Vo32:33

      ... instance?

   6. LHLuke Harries32:33

      Yeah, desktop Claude. They've also now shipped the ability for you to run it from claude.ai, uh, if you host it too. Um, and what I've got is, I've got this MCP called WhatsApp here, and I can just type into the chat. Uh, so what, uh, are some recent messages in WhatsApp I've received? [typing] And, and what that will then do is it will ch- use the tools that the WhatsApp MCP exposes, and then summarize and use back that information. And here you can see it's talking about ElevenLabs' new features. And so, uh, ElevenLabs is launching a conversational AI agent or a new speech-to-text offering, um, as well as a few tests, like "How are you?" and "Hello, world!" Um, and a few examples of, like, why you may actually want to use this is I'm in a whole bunch of different WhatsApp groups, and what you can do is you can use it to summarize, oh, what over the last week were people actually talking about? And, and often now, some of the best, uh, way to keep up with trends or the best thoughts on new tools, they're all in these WhatsApp groups with hundreds of messages per day. And so, uh, if you're looking for a way to get an edge on Twitter or LinkedIn, you can say, uh, "Summarize the thoughts on ElevenLabs from the messages," um, that will then search the messages that you've received on WhatsApp, which is talking about it. Uh, and then you could take something like this, chuck it into your GPT, which is already trained on your tone of voice, uh, and generate a tweet thread for you.

   7. CVClaire Vo34:19

      Everything's a launch.

   8. LHLuke Harries34:21

      Everything's a launch. Uh, and also to plug it, if you're interested in how we run launches, the- I have a blog post which goes through, like, all the different sets that we use as well.

   9. CVClaire Vo34:32

      Yeah. We'll, we'll link to that in the, in the show notes. So, okay, just recapping this for folks that are still-

   10. LHLuke Harries34:38

       Yeah

   11. CVClaire Vo34:38

       ... have, have their mind blown. So you built this MCP, which you've open-sourced, which again, is unofficial but friendly-

   12. LHLuke Harries34:45

       Yes

   13. CVClaire Vo34:45

       ... um, implemented in a nice way. Uh, you download this code, run it locally in your CLI. It does a one-time pull of your data.

   14. LHLuke Harries34:55

       Yeah.

   15. CVClaire Vo34:55

       So if you wanna update it, do you just run and refresh that again, or does it pull? How does it-

   16. LHLuke Harries34:59

       So when you start it, it will pull-

   17. CVClaire Vo35:01

       Mm-hmm

   18. LHLuke Harries35:01

       ... all the recent messages between the last time it ran and now.

   19. CVClaire Vo35:05

       Yeah.

   20. LHLuke Harries35:05

       And then whilst it's running, it will also receive any new messages.

   21. CVClaire Vo35:08

       Got it. Okay.

   22. LHLuke Harries35:08

       So actually, I can als- and I can use it to, so send a message to... And then I'll do a phone number. [typing] And then potentially if my phone's not on silent. Yeah, so that's just came through saying hello. And, and then you can also use it, connect it... The cool thing about MCPs as well is you can connect to lots of MCPs at once.

   23. CVClaire Vo35:33

       Yep.

   24. LHLuke Harries35:33

       And so I have an ElevenLabs MCP installed, so I can use it to- you could, like, generate a voice roundup. So you generate text-to-speech of this. [typing] Yeah, so you can stitch these all together.

   25. CVClaire Vo35:50

       Okay, so, so you have this MCP, it pulls down your stuff, it gets regular updates. It's all local, so none of this stuff-

   26. LHLuke Harries35:55

       Yeah

   27. CVClaire Vo35:56

       ... is going to the cloud. Um, and then you've connected to that serv- or that server through your, your local desktop Claude, or, or if you, you hosted it, you could do it on the, the web version. And then now, just in this chat box, you have-

   28. LHLuke Harries36:11

       Yes

   29. CVClaire Vo36:11

       ... access to all these different tools. And, you know, one of my hypotheses with AI is, like, tabs start to go away. If you're like me, you have five hundred tabs open, everything along the bottom of your desktop dock, and you're switching context, and this sort of centralized chat interface that can access all these tools just makes you much more efficient, and it also allows you to get really creative about how you stitch these tools together.

   30. LHLuke Harries36:37

       Yeah, and so, and so for this one, for example, we just generated this text, and then you can say, "Send this to the phone number."
10. 38:07 – 41:08

    Creating specialized AI agents on demand

    1. LHLuke Harries38:07

       it up. And so the really cool thing about these chat-based MCP tools is it can be much more... You know, it's trying lots of different ways to, like, "Okay, how do I actually send this audio message?" You know, on the spot, we were like, "Okay, generate this, now send this." And the AIs, as the models get smarter and smarter, are able to get- deal with these, like, high-level abstraction tasks. And so a genuine one you could do is, uh, why not have your AI actually phone up Claire and have the conversation for you about that? [chuckles] So, uh, if you want, we could actually try that now. I'm not sure if it will work.

    2. CVClaire Vo38:48

       Why not? Let's try. [chuckles]

    3. LHLuke Harries38:50

       So we can do that. So create a conversational AI that can, uh, do case study interviews. And so this will then use ElevenLabs to actually create an AI agent that you can speak to about case studies. So first of all, it's gonna, yeah, list the agents.

    4. CVClaire Vo39:10

       I love this because you're using AI to create more AI.

    5. LHLuke Harries39:15

       Yes.

    6. CVClaire Vo39:15

       You're really just replicating agents on agents on agents, and then what this would do is create a specialized agent on the spot-

    7. LHLuke Harries39:25

       Yes

    8. CVClaire Vo39:25

       ... for a specific use case, um, that then you could use to give me a little call-

    9. LHLuke Harries39:32

       Yes

    10. CVClaire Vo39:32

        ... and get, get a case study done.

    11. LHLuke Harries39:35

        Great. Yes, exactly. And so you can actually see that here. It went through and created a prompt.

    12. CVClaire Vo39:41

        Yep.

    13. LHLuke Harries39:41

        So it's got the first message: "Hi, I'm your case study interview coach." But if you go to our Twitter, you can see Louis doing this workflow, where he then, um, phones up and orders a pizza using an AI. And you just say, "I would like to create a pizza ordering AI agent." But yeah, hopefully this gives the listeners a glimpse into, like, you can on the spot come up with these agents, which I think will be more and more abstract, which can then do these tasks for you using the tools as they go. So that's the promise of MCP. And transparently, it's still very early, and I think most of these are tools, uh, are toys, but I do think it's, it's going that way.

    14. CVClaire Vo40:20

        As a parent of kids who really like pepperoni pizza, I'm very worried about the ability to spin up a pizza order, uh, ordering agent in my house because we will end up with, with a lot of pizza. Okay, Luke, this has been incredible. Uh, you made a case study with me using Granola AI and your magical GPT. We eliminated t- forty thousand dollars of spend, um, by coding in Cursor as a, as a marketer. And then you built a way to use Claude or an AI to query your source of both personal information and industry news, which is WhatsApp, and do a bunch of really interesting stuff there. So, uh, this has been very eye-opening to me. I've learned so many things from what you showed me, um, including that everything is a launch. I'm just gonna keep that in my mind.

    15. LHLuke Harries41:07

        [chuckles]

    16. CVClaire Vo41:07

        So let's wrap with

11. 41:08 – 44:24

    Lightning round and final thoughts

    1. CVClaire Vo41:08

       a couple lightning round questions. The first one is, you know, we've, we've talked a lot about coding and text-based flows. Even in what you showed, a lot of it is coding and, and text-based flows. But I think what's so interesting about your point of view is you're starting to bring the idea of voice as input and output into, um, into the, the industry and into how people build products. You know, very quickly, what do you think kinda, like, voice modalities unlock, maybe for product managers to think about in terms of what they're building?

    2. LHLuke Harries41:40

       I think there's two broad types of things they unlock. So one of which is new experiences for customers, which just wouldn't have otherwise been possible. And so one great example is, like, if you're in education, suddenly you can make something which is way more engaging. So Chess.com shipped a app with, uh, Professor Wolf, which enables you to get, like, turn-by-turn guidance on your different chess. And so you can imagine a world where every- whether you're learning a language, you're playing chess, you're, you're, uh, learning maths, that everyone can have this, like, interactive tutor, and that's kind of, yeah, these new experiences. The other type, which I think is really exciting if you're a product manager, is you probably have a lot of back office functions. And but so particularly if you're an internal PM, if you look at all the places that you currently have, um, you know, people doing... manning phones, so often this would maybe be like, say, you're doing research collection, you know, you're a scaled marketplace, and you have large numbers of people collecting data, or you're doing customer support. Well, maybe actually currently you're not able to expand internationally because your team only speaks English, whereas now you could spin up an entire team of customer support agents who are fluent in French and in Spanish and in German.

    3. CVClaire Vo43:06

       Yeah.

    4. LHLuke Harries43:06

       So I think that's really exciting, too.

    5. CVClaire Vo43:08

       Yeah, I love the, I love the international, uh, angle to this. It's something that I haven't heard very many people speak to. Okay, Luke, this has been so great. Where can we find you, and what can we do for you?

    6. LHLuke Harries43:21

       Thanks so much for having me on the podcast. It's been a lot of fun. Uh, you can find me, my website is harries.co, um, and a couple of blog posts people may enjoy. So I've got the, uh, how to launch your product, where I literally talk through the checklist that we use-

    7. CVClaire Vo43:35

       Mm-hmm

    8. LHLuke Harries43:35

       ... uh, through a lot of launches to go from idea to value prop to core assets to distribution. I also talk about how to hire your first growth marketer, um, and you can follow me on Twitter. That's Luke Harries, and then underscore.

    9. CVClaire Vo43:50

       Great. Well, this has been so fun. Thank you so much.

    10. LHLuke Harries43:53

        Cool. Thanks so much, Claire. [upbeat music]

    11. CVClaire Vo43:56

        Thanks so much for watching. If you enjoyed the show, please like and subscribe here on YouTube, or even better, leave us a comment with your thoughts. You can also find this podcast on Apple Podcasts, Spotify, or your favorite podcast app. Please consider leaving us a rating and review, which will help others find the show. You can see all our episodes and learn more about the show at howiaipod.com. See you next time! [upbeat music]

Episode duration: 44:24