Vibe coding: from prototype to product
with Meghan Villard
I joined Meghan Villard on Data Fuel by Full Send to talk about vibe coding: getting a prototype working, deciding when it needs engineers, and taking responsibility for the software you keep. We also walked through research, planning, and implementation using a trip to Japan and patient-notes search.
Chapters
- 0:00 — Welcome to Data Fuel
- 2:30 — How far can a non-technical person go?
- 4:32 — Who owns the engineering decisions?
- 9:12 — Choosing tools and building a workflow
- 11:31 — Planning in a separate context
- 14:43 — Research, plan, implement: a trip to Japan
- 20:48 — Patient-notes search and domain expertise
- 26:04 — Where to start and what drives me
Transcript
Meghan Villard / Adam Hevenor
Proper nouns corrected from the automated transcript. Speech is otherwise preserved. Mixed or uncertain speaker turns are unlabelled.
0:00 Meghan: Welcome to Data Fuel. I’m Meghan with Full Send. Today we’re going to dive into Vibe Coding. So I have with me Adam Hevenor, who’s actually perfect for this discussion because his title is literally Chief Vibes Officer, helping companies build software, AI-enabled tools. So yeah, I’m excited to dive into just this world of like what’s going on with coding and software development. So yeah, I’m very glad to have you.
0:29 Adam: Awesome. Excited to be here.
0:30 Meghan: And you are actually the first person that I’ve met that has Vibes in the title. You know, we’d met at AI’s Builders event and just love that and led into a lot of discussion after that. So tell me, how do you define like Vibe Coding?
0:48 Adam: Yeah. So Vibe Coding really is using, Andrej Karpathy defined it and he defined it as I take a prompt, I put it into an LLM and then whatever it gives me, I try to use it. So the origin of Vibe Coding is really just using a language model. At the time, it was just taking it out of ChatGPT, putting it into your system, seeing if it worked. Vibe Coding has come a long long way since then. So, you know, today Vibe Coding really means still using that model to create software.
Adam: And it’s really grown into a whole spectrum of people doing really pretty challenging engineering feats with it to just applying it to, you know, small things you want to do or working with your kids and build a game or whatever you want to do. So it’s become a spectrum. But it’s also, I think, a lot about having fun while coding.
1:50 Meghan: The creativity.
1:51 Adam: The creativity.
1:52 Yeah.
1:52 Meghan: So it’s really just using the natural language in a way to go build things.
1:57 Adam: Exactly right. And, you know, a lot of people are doing exactly that. I use a lot of voice.
2:03 Meghan: Okay.
2:05 Adam: My kid loves to use voice. He’s not as good at like typing yet. So yeah, even like going that level, higher abstraction, thinking about how do I define the outcomes I want to achieve and use the latest models, use the latest tools to achieve that. And tons of exciting stuff out there to build with.
2:26 Meghan: And you have a technical background.
2:29 Adam: I do. Yes.
2:30 Meghan: How far can a non-technical person, because I feel like that’s one of the benefits is that like an operator or a business founder or a kid, someone who’s not technical can do this. So how far can someone who’s not technical get using this concept of like Vibe Coding?
2:46 Adam: Dangerously far. Uh, you can get as far as you want to get. And, uh, you know, I think that understanding what the problem is you’re trying to solve and how much engineering decision making needs to go into that should sort of help determine how far you should go. Uh, you know, if you’re building a game and you’re a kid and you want to experiment, you want to try as much interesting stuff as you can, there’s really no risk. You should do as, you can go as far as you can and use the model to learn, go do research, uh, you know, figure out new techniques.
Adam: Um, obviously you’re, if you’re trying to build a business critical piece of software, you can get very far. Um, and you should probably take a moment to think about, should you be doing what you’re doing? Um, uh, you know, there might be a good reason for it. And so the answer isn’t always no with these tools. Um, but, uh, yeah, being conscious about sort of how you’re applying the tools and what risks it introduce, uh, really determines how far you can go as a non-technical person.
Adam: I think, uh, you know, the tools can basically let you go as far as your curiosity wants to take you.
4:11 Meghan: Which is very far for a lot of business folks. Um, but you help people implement this.
4:18 Yeah. So if people can get far on their own, where like, would you, or someone else who does have that technical background, like an engineer architect, where would they come in and actually help you get it to that next level? Like, can you give a few examples of where you or a technical person would fit into that?
4:34 Adam: Yeah. So like I said, when you’re building something that has lower business risks, you can let the model make basically all the decisions. So you’re making a game, you don’t care what framework it’s using. You’re not concerned about security vulnerabilities, really any of these things. Um, once you decide you’re going to take a piece of software, maybe turn it into a product or have it support some part of your business, uh, you got to think about the lifespan of that software at the very minimum in like quarters.
Adam: Hey, this is going to help us achieve a goal for next quarter. Um, but more than likely, um, but more than likely you’re going to take some time to build this out and run it for multiple years. And there’s going to be an accountability of what decisions get made over the next, let’s say two years for your vibe coded piece of software now. Uh, so it’s okay to build those with tools that vibe code the process, maybe generate all of the code using a model. But I think you’re going to want someone that has takes the responsibility of those decisions for the next two years, knows how to operate that software, knows how to observe it.
Adam: All the things that really go into software engineering. So, you know, that’s kind of like the point where it’s like, okay, I want to turn this into a product or a business critical system. Uh, you know, thinking about it like a product is something that I help teams kind of transition from, uh, you know, their vibe coded app to, to something like that. Um, and then, you know, scale, all those other things, of course, often come with that. Uh, you know, and, uh, again, if you’re curious and want to try experiments, you even can use the models to like figure out how to scale things.
Adam: Um, but again, even more increases, the odds are going to live with those decisions for a long time, the long-term maintenance of it.
6:37 Meghan: And one example. So we, we started a while back, we built, um, a very simple app just to scope projects to say like how roughly how big is this a small, medium, large, and what’s the timeline. And so to me, that was something very, it didn’t have risk to it. It’s just a lot of different inputs. How big is the client? Um, what are the tasks that we’re doing for them, um, to spit out, you know, the team size is very basic. That felt very doable and just like our own vibe code.
Meghan: But now we’re thinking about, okay, how do we build like a client portal to log into their own information? And that’s where like the security risk, the login information connecting to a database with sensitive data. Um, so I feel like all those aspects of like, yeah, like you said, the growing. So now other people are actually accessing this security, connecting into data. That’s where I feel like it gets a little more, um, where maybe you need that technical expertise come in and help get it across the finish line.
Meghan: But at least you can be handed something of like, this is my vision versus starting like from scratch.
7:36 Adam: Exactly. And yeah, having that curiosity to take that first step, I think will help a lot of businesses, uh, just be more effective if they’re gauging with someone like me or they’re, you know, trying to build a software product. Uh, the prototype, uh, the prototype, the code, all of that is no longer expensive to produce. Um, like I said, the liability in it is that running it for multiple years is exactly that. It’s a liability. You’re going to have to have people on staff and operations infrastructure and things like that to, to do that.
Adam: Um, but also. So the, like ability to hockey stick something, see like huge growth off of what, you know, if you have a really novel idea, being able to prototype it and then build something, um, uh, I definitely think is, uh, you know, it’s unique for people like me who have a technical background, but I actually think ideas like you’re talking about where I have this business problem. Um, I’m in a domain, um, I’m in a domain, domain specific application and building out that idea really has the potential for rapid growth and then.
8:46 Yeah. Can maintain it. How do you, you know, how do you scale it?
8:48 Meghan: Yeah. So then hire the engineers later, maybe the team just exactly the size. Um, and I know there’s like right and wrong ways to do this. Every tool is different. So we can eventually talk about tools like the Lovable Replit versus Claude Code. Uh, but, um, what is the best practice for actually coding this is I know people joke about getting it right in the one shot, but, um, what is your typical process of like a development cycle that you would recommend?
9:12 Adam: Yeah. So, uh, I mean, I think before you get into the process, like you mentioned, there’s a bunch of tools, I think finding something that someone’s recommended you’ve heard of and starting probably with something on the web. Like you mentioned Replit versus v0 Lovable, uh, Google, they all have good products to try it out, um, without really putting, installing anything on your computer. Um, uh, cause that does take the low, you know, you’re kind of up the risk once you install something on your computer and you start using tools on your computer.
Adam: Um, that said, um, that said, I definitely recommend, you know, getting to that next step of using something on your computer. And that’s really where you’re going to probably transition from, uh, maybe what is what, you know, the origin stories of vibe coding. You’re just like in a chat box and an app is sort of being built, um, to starting to think about, okay, how do I create a workflow so that I can work on my application such that every time I go to make an improvement, it’s easier to make another improvement and add another feature.
Adam: Um, and that takes a bit of discipline and, uh, uh, learning, um, but, uh, definitely recommend, you know, as you gain a little more experience using one of the either IDE products like Cursor, Antigravity, et cetera. Um, or better yet going to a terminal based product, like Claude Code or Codex. Um, I think that’s really the next level of being able to unlock lots of productivity hacks.
10:57 Yeah.
10:57 Meghan: And would you say like, whenever you go into the terminal for the first time, you’ve got a blank slate, would you just explain what you’re looking for? Do you need to actually go into like a different chat window and say, help me write a prompt to do X. What does that look like? Cause that’s what I’ve seen a lot of people actually go into ChatGPT and say, I’m going to build this in Lovable, right? Give me a prompt.
11:16 Sure.
11:16 Meghan: Is that right? Like, should they just start in the tool that they’re going to use? How much should they explain on like the first attempt? Um, is it best to almost explain everything in one or just start small and then build as you go?
11:30 Adam: So there’s some good things about doing it that way. And then there’s some, I think, you know, inefficient things about doing it that way. So the thing that is really efficient about doing it that way and is something that as you go to more advanced workflow, you start to, you, you don’t know you’re doing in this workflow, but you actually are. And you may bring to later workflows is by crafting your prompt in a separate chat window. You’re actually keeping this clean planning context window. Um, so the, the models, uh, use what’s called a context window.
Adam: And every time you interact and add to your chat history, the model is filling up its context window. And the likelihood that it’s going to hallucinate, make a bad decision goes up. It actually goes up quadratically, which means it gets a lot worse. Um, and, uh, what you’re doing by saying, Hey, I’m going to figure out what I want to prompt to make my app in a separate chat window. Is you’re basically coming up with a plan that then you’re handing off to a separate agent.
Adam: That’s going to do the coding for you. And that separation of concerns works really well with the models. Um, and it’s something, once you get into the IDE and code-based tools, it works so well, they build it into the product. So they have a plan step where you create a set of to-do lists. And then once you’ve created your to-do lists.
13:00 That looks good. That looks good.
13:02 Adam: And the more advanced IDEs and agent and, um, you know, agent frameworks, they will swarm agents on your to-dos. So they’ll like in parallel research the changes they have to make and then go make the changes.
13:16 Okay.
13:17 Meghan: So, so you’re saying there are some tools that you just start planning directly in there. And it’ll help you think of the to-do list and you come and say, yes, I like this. Go implement.
13:26 Adam: Exactly. And that’s actually all the coding tools I mentioned. So, you know, Cursor, Antigravity, Claude Code, you know, Codex, uh, they all have some, you know, toggle that says, hey, I’m in plan mode. And actually all that plan mode is, is it’s a little prompt in there that says, don’t write any code.
13:48 Okay.
13:48 Adam: And just think right now.
13:50 Just think right now.
13:50 Adam: Let’s, let’s come up with the plan.
13:52 Okay.
13:53 Adam: Um, and, you know, one of the ways I think about it is you’re creating leverage with those tokens that you spend on the plan.
14:01 Interesting.
14:02 Adam: Um, so, uh, you know, you can even back it up further. I often plan the plan. Um, so, uh, uh, this is called a research plan implement workflow. Uh, so before you even do the plan, you do research on the problem you’re trying to solve.
14:18 Okay.
14:19 Adam: Um, and that helps you frame the plan a little bit better. Uh, and you go back and forth. Um, and there’s sort of this compounding effect where the more time you spend on researching the problem, the better your plan will be. And then the more time you spend on doing the plan, the faster, more efficient your end implementation will be.
14:43 Meghan: So give us an example of the research. Where would you start with that?
14:47 Adam: Yeah. Um, we were talking about this ahead of time and, you know, it’s kind of a good, they’ll take us out of like software engineering world. Just make this, you know, easy to understand, um, a good example of research that we were doing recently was planning a trip to Japan and we, you know, want to see certain things. We want to go to certain places, hit the bullet trains, all the stuff you do in Japan. And, uh, my wife and I, we just started like chatting with Claude in the app.
Adam: Hey, like, what are we thinking this? Um, and, and, you know, the models, they really want to help you. They’re, they’re really good at planning. They like are just instantly doing the research. And so like very quickly, we’re like ending up with these chat histories of like, oh, go here, then go there. And, oh, when are your flights? And, and, uh, it got really unwieldy trying to do it in like the phone chat interface. Not surprisingly. Um, uh, so, uh, you know, we basically. Use that as some research.
Adam: We came up with like, Hey, here’s some research. We want, you know, here are some of the bullet trains we want to hit. We want to do this. Now let’s plan an itinerary.
16:02 Meghan: Got it.
16:02 Adam: Um, so, uh, you know, from there we did basically transitioned out of, uh, the, you know, Claude chat app. And it was right before Claude Cowork came out. I think Claude Cowork came out. Claude Cowork came out while I was working on our itinerary.
16:21 Meghan: You didn’t switch.
16:22 Adam: I didn’t switch. Cause I already have Claude Code. And I put all our plan into Claude Code, all our research into Claude Code. And I said, Hey, we need to plan an itinerary.
16:32 Meghan: Love that. That’s a good example. Cause it becomes real to people of like, you wouldn’t start with plan our itinerary when, you know, you know, the country you’re going to, it’s like, okay, we want to see the cities start big picture and then get down to like the hotels, the trains, the transportation, the details.
16:48 Adam: Exactly. And to extend the research plan implement, uh, metaphor further, the implement in this case was go book the hotels for the most part. We didn’t send our agent to do that. We went, I mean, just like, just like coding, you know, like you may go actually occasionally like just do the code yourself and I could see a world where I allowed my agent to potentially book things.
17:13 Meghan: Here’s my credit card. Here’s a limit. Go crazy.
17:15 Adam: Exactly. Here’s a credit card with a limit. Go crazy.
17:18 Meghan: Book maybe with cancellation options.
17:21 Adam: Uh, we have booked many travel things and made many mistakes. Like, so we ourselves chose cancellation opt-ins for like a little bit more booking fees and did in fact make mistakes as humans on a couple of things. So, uh, uh, you know, I think having the implement, separating the steps also gives you the ability to decide what pieces you want to hand over to the model and trust the model to do and which pieces you don’t trust the model to do.
17:54 Meghan: Yeah. I love that. And funny enough, we’re both planning a trip to Japan near the same cherry blossom season, but a great example with our research, uh, one thing we research is, can we drive there? Like, what is that like? And so that’s a good point of like, that’s not going to go in the itinerary. That’s just researching, but that impacts itinerary. Do we need to, can we just rent a car and drive to all these places kind of off, off the, you know, typical transportation routes, or if not, our, our itinerary changes quite drastically.
Meghan: And so that’s a good example too, of like the research part of, um, okay, what, you know, are there festivals we want to hit? Um, when should we be at certain times? And now we did the exact same thing. Now have more of an itinerary in something like Claude Code, um, that says, okay, here are, here’s our hotels. Um, how, and the dates of each help us plan. I want to do these five things. I want to go, you know, have the restaurant on the river.
Meghan: I want to see the cherry blossoms, uh, um, give me an itinerary. And it went down to what seat should I book on the train to Kyoto to see that? Like that’s wild, um, that it could go find that research and tell me down to like, based on what I want, you know, nature and some of the themes it told me down to like, what seat to sit on a train. So that’s pretty cool.
19:09 Adam: It is pretty cool. And I think it’s a bit like GPS where I don’t necessarily need the instructions, but when I booked my trains, I also screenshotted every single step and said, Hey, is you sure this is right? Can you make sure this is the right time from the hotel and everything? Um, so yeah, I definitely think as I mentioned, I could see a world where you could give the model access to go do those things. Obviously there’s still like a lot of like tooling and stuff like that.
Adam: Um, I think making it more convenient. Some of the ideas that I was just toying around is it would be great to set up an email as like where I could forward all my confirmations.
19:55 Meghan: And then that’s your agent just takes that.
19:56 Adam: And then the agent could take that. And that’s definitely some of the things you can do with Clawdbot or open, excuse me, OpenClaw.
20:04 Okay.
20:04 Adam: No, you know, don’t want to use the old name. Uh, OpenClaw, formerly Clawdbot. Um, I can send it Slack and stuff like that. So, you know, you can send it the receipt that way. Um, but I think, I think email would be extra nice for travel confirmation since you’re already getting the email.
20:24 Meghan: Yeah. I love that. Um, so love that example of the Japan. I feel like that’s a great to con to grasp of like, this is when you use your chat, you’re planning, and then you get into the, like, um, or you’re researching, then you get into the plan to implement. Can you give us one more example of like vibe coding, how you’ve seen it work? Well, kind of what the use case was and how it’s being implemented or used today?
20:48 Adam: Yeah. Um, I mean, one of the projects that I’m working on with our client is, uh, around using what’s called EMR data or patient notes. So when you go to the doctor, uh, everyone knows now you kind of, there’s a, there’s a recording, there’s a transcript of, uh, you know, your interaction with your doctor and, uh, you know, that data is anonymized and made available for research of new drug companies. So for drug discovery, uh, basically these companies are able to go in, you know, they search for Jane Doe and they find out, oh, the drug was too expensive or it had a side effect or this or that our, our new drug should be able to capture more of the market.
Adam: Um, so, uh, you know, this was something that, uh, I use some of my, you know, engineering background and really worked with their engineering team to do something that with a small team, they really couldn’t have done just a few years ago. Um, so, you know, that’s obviously on the spectrum of things like taking AI with an engineering team and sort of multiplying that engineering team. Um, lots of other examples of people building whole projects by themselves and things like that, of course. Um, but, uh, I think, uh, really impressed by small teams that are accomplishing a lot.
Adam: Um, so.
22:23 Meghan: Yeah. And who’s the stakeholder of this tool that you helped build?
22:27 Adam: Yeah. So for this company, their stakeholders and their subject matter experts are PhD epidemiologists. Um, so, uh, you know, we have basically built out with the engineering team, this infrastructure ground layer of APIs where they can go and search these patient nodes and a bunch of different techniques. Um, and it’s really cool to see, cause now the epidemiologists, they’re vibe coding an application layer on top of that to go research specific disease areas, pull out data on those disease areas, um, they’re, you know, as productive with the tools as the engineering team, um, and bringing a whole new set of expertise and skills to the table.
Adam: So that’s really cool to see.
23:19 Meghan: And sometimes I feel like that deep expertise and the domain knowledge is the most important skill set to have in building these products. Like if you know your space well, now you don’t need to know the coding as much. So that gets you like a level up.
23:30 Adam: Absolutely. The knowing how to use the tools is still one of the valuable roles at the table. But I think having the subject matter expert is as important. And I think a team where you have both of those and really understand your end user, um, is how you build successfully.
23:53 Meghan: Yeah, I love that. Um, and one example that I’ve, we’ve built down kind of seen it progress. I feel like the progression is always nice to like, see, um, one client we had, they needed to, um, they had a whole data life cycle for this like marketing project that they were going to do. Um, and so, you know, the, the product, it was like a test, um, an assessment, if you will, of what is your current situation like? So they would offer this for free to clients, um, do the tests, give the results.
Meghan: And then hopefully that would be, um, in a way in the door to go sell a different service to them. Um, but they had a lot of different touch points on this data. And so we’re like, what if we just built a very lightweight app internally where different teams all the way from sales and top of funnel to the person running the test, the results, the marketing, the salesperson on the other side of it, they can come in and input that data and see kind of the results over time.
Meghan: And so it started, um, or the team built it themselves as internal, very kind of basic.
24:54 Adam: Yep.
24:54 Meghan: Um, and then that’s where we came in to actually help implement this. And now it’s a customer facing portal. So there’s the security. It’s actually pulling in data from HubSpot. It’s pulling data from the test results. It’s pulling data from a few different places and the customer now sees it. And so that’s where it’s cool of like the concept of the vibe coding is they started with something that was amazing and working for them and then they wanted to grow it even further. And so that’s where kind of the, the architecture mindset came in, um, to make this like a true product.
Meghan: So it got them off the ground very quickly.
25:26 Adam: Very cool. Yeah, definitely just a different paradigm now to build these types of products. And a lot of the conventions we’ve used in the past are changing. Um, I wouldn’t recommend just, just try to go build it as the first step, even just like 18 months ago. Um, but that is the best way for business leaders, entrepreneurs, if they have a problem that they’re close to, they should really try to solve it with the tools first. Um, and then, like you said, as you want to scale it to a product, start to bring in the experts and validate that demand super early before you invest in that.
26:03 Meghan: I love that. Test it out. So I think you kind of got into this, but the number one thing people can do to learn to get in that space. It’s not really like videos or learning anymore. It’s just doing, would you say that’s correct?
26:15 Just get in there and definitely get in there.
26:17 Adam: Uh, you know, try out the tools, uh, like I said, I would say level one, go try the v0 Replit Lovable to everyone you’ve heard of first, maybe level two, try Cursor Antigravity. Both of those have pretty decent little free limits. And then level three, Claude Code.
26:41 Meghan: Love it. Co slash Cowork now slash Cowork.
26:45 Adam: Uh, I, I’m not, I don’t have a ton of experience with Cowork. I definitely think though Cowork, uh, it sits kind of somewhere close to like that Cursor use case, but isn’t just for coding. Um, so if you’re gonna like plan a trip to Japan, Cowork probably is the thing to use. Uh, you know, I think I just like Claude Code and just use, we’ll use it anyways, but, uh, yeah, uh, Cowork, um, I think could solve a lot of workflow problems for a lot of, uh, you know, uh, entrepreneurs and business owners.
27:24 Meghan: It’s making this accessible.
27:26 Adam: Yep.
27:26 Meghan: I love it. Well, okay. Last question, Adam, we at Full Send are passionate. Like our team is out hiking right now. We’d love to just that mentality of like, what drives you? What makes you like passionate about just the work that you do or outside of work? Uh, so what is your Full Send?
27:41 Adam: My Full Send is traveling, also being outdoors, you know, as already mentioned, going to Japan and, uh, doing a family trip here for two weeks. So that’s definitely my, my Full Send.
27:55 Meghan: I love it. Go see the world.
27:56 Adam: Yep.
27:57 Meghan: Well, uh, any final, final words?
28:00 Adam: I think that was great.
28:01 Meghan: Okay. Well, I hope this has been helpful. Um, I know this is a hot topic, so hopefully people listening go build one cool thing this week.
28:08 Adam: Yes.
28:08 Meghan: Um, yeah. Thanks for tuning in.
28:10 All right.