# Managing Your Vibe Coding Junk Drawer  ## Episode metadata - Episode title: Managing Your Vibe Coding Junk Drawer - Show: The WP Minute+ - Owner / Host: Matt Medeiros - Guests: [Kim Coleman](https://share.snipd.com/person/2aa64458-c1b7-499d-9a72-8d3a3f1fbf70) - Episode publish date: 2026-04-20 - Episode AI description: Kim Coleman, co-founder of Stranger Studios and lead on Paid Memberships Pro, shares her experience with WordPress plugins and integrating AI agents. She discusses vibe coding, using agents like Flint for workflow automation, balancing speed with responsibility, organizing AI projects, and shipping low-risk prototypes to learn quickly. - Duration: 37:08 - Episode URL: [Open in Snipd](https://share.snipd.com/episode/225c7442-39d4-47f5-83fd-8ac9273b1def) - Show URL: [Open in Snipd](https://share.snipd.com/show/6da84b69-b825-4cef-9a81-24e1aeeb5c5d) - Export date: 2026-07-08T01:10:39 ## Snips ### [Matt's Junk Drawer of Rapid Prototypes](https://share.snipd.com/snip/caad3298-a21e-47cb-b27a-577692882ae8) 🎧 04:58 - 07:54 (02:55) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/93e7d4fb-ae93-4f22-9626-075628671f24" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Matt describes building many quick AI projects and forgetting where they were hosted. - He launched podcast tools, prototype sites, and API experiments across Netlify and Cloudflare, creating a literal junk drawer of projects. #### 💬 Quote > I have launched a lot of junk that I have not done anything with. > — Matt Medeiros Matt Medeiros on launching many AI prototypes he later lost track of #### 📚 Transcript **Matt Medeiros:** I think it stems back. I think one of the real interesting things about AI is the same thing that really excited me about WordPress, you know, when I got into it 16, 17 years ago, was the ability to do something like cool that I couldn't do with my own abilities of writing lines of code, right? So when front page came out back in the day, and I was able to build websites graphically, that was really interesting to me, right? When I first started learning websites, like a lot of folks, it was just writing HTML, making a bunch of tables, and then putting spinning animated skulls in horizontal rule bars, you know, on web pages. And that was fun. But then as technology accelerated and matured, like all, I think, well, I think like for all software, for folks like us and those of us watching to create things, it iteratively gets better over time. And AI is just a massive accelerant to that, right? When I first got into building sites, it was with Drupal, and Drupal has CCK and Fuse, what we kind of know as ACF in the WordPress world, and then an ability to display this content like a loop block. It was amazing. I was like, look what I can do without having to write PHP and MySQL statements. And this AI stuff has really accelerated that. So I started playing around with it maybe like two and a half-ish years ago, specifically like ChatGPT. And, you know, know the overarching thing here today is like the junk drawer. I have launched a lot of junk that I have not done anything with. And as we were leading up to this conversation, I was like, God, I, the first things I was doing is with like making podcast downloaders and podcast players. And I'm like, I don't even remember the website I launched that on. Like, what did I do with that API key? I don't even remember when I was like launching this stuff, you know, in the beginning. But really, for me anyway, it comes down to the excitement of being able to do things I'm not able to do, learning along the way. So it's not just a blind eye to just, like, throw this stuff out there and, like, see what happens. It's accelerated my learning of other technology stacks, spending a lot of time in Cloudflare, understanding JavaScript more, understanding how other developers might look at WordPress and go, what is this thing? Look what I can do with these libraries here with whatever, like React app, so on and so forth. So it's both exciting. It's a great education process for me. But I have launched a bunch of stuff that I don't even know what to do with anymore. I'm just thinking about having to go and take it all down. I'm like, oh, God, do I even want to do that? --- ### [Vibe Coding Forces Better Architecture Thinking](https://share.snipd.com/snip/d7e9db87-29c4-45ca-a307-8465520855a4) 🎧 10:20 - 11:06 (00:45) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/55b4412a-bc6e-4db5-bea7-d76b6de3dc28" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - AI shifted Eric's focus from copy-paste fixes to architecture and code structure. - He notices AI produces standard folder structures and repeatable patterns, making vibe coding more like templated frameworks. #### 💬 Quote > With AI I find myself actually thinking more about architecture. > — Eric Karkovack Eric Karkovack on how AI changed his developer focus toward architecture #### 📚 Transcript **Eric Karkovack:** And it's funny, with AI, I find myself actually thinking more about architecture. And that's like the one thing I've really learned through all this is that I am thinking more like how this thing is going to work, not just that it works. I think that's been kind of valuable for me. **Matt Medeiros:** Yeah. Yeah, agree. Kim, **Kim Coleman:** I'm really- mean, when you say the word architecture, I think like not only how it's going to work, but also how the code is structured. So I think people who are using AI to VibeCode more with are realizing that AI is starting to use these kind of standard folder structures and building itself in a way that becomes repeatable. --- ### [Use A Human AI Human Review Loop](https://share.snipd.com/snip/739b14d1-46cb-4d09-834f-1a0646ce7f5b) 🎧 12:37 - 15:15 (02:38) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/0eb536c4-2c39-4dec-9368-a57e57cb9cbf" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Use a human-AI-human workflow for AI-generated code and assess risk before shipping. - Kim says design the brief, let AI build, then have humans test, review, or patch security issues before public release. #### 💬 Quote > At the end of the day, there has to be a human saying how important and how critical and how bad would it be if this code is insecure or fails or breakdowns. > — Kim Coleman Kim Coleman on ensuring human oversight for AI-produced code #### 📚 Transcript **Kim Coleman:** let you go first. matt okay yeah um so we actually published three plugins this week not this month not this year three in a week and um i think when you're vibe coding when you're launching things that are largely ai based you have to use a human ai human system. So Eric, you about the architecture thing, the human designed and created a brief, even with the help of AI, let's say. You went back and forth in chat with AI to design the framework structure, functionality, no-goes, rabbit holes about a project. You handed it to AI to develop it, and you got it back, and you either tested it, tested it thoroughly, had the code reviewed, handed it to another AI to code review. All those things are great. At the end of the day, there has to be a human saying how important and how critical and how bad would it be if this code is insecure or fails or breakdowns. You know, what am I opening my customer, my user up to? What risks am I opening them up to? And if it's relatively low and you're just like, you know what, this is a landing page about snow and has a map, like there's really no risk. Like people hit the website. I mean, the worst thing that could happen, I have some kind of injection point that is in the server and something goes haywire and I pay some crazy Google Maps fee. Like, I'm trying to think what the worst thing that could happen is. So I think having a human sense of the security and the risk of it being incorrect is kind of an important thing. But also, as quickly as that code was created, that's how quickly bugs can be fixed or faster because it's a day later and everything got way faster already. So we're kind of in this mindset of if it's low risk, push it. Let users be the testers. Let users be the stress testers. Let them try it and say, I wish it did this. Let them try it and say, it's broken. It doesn't work this way. It's way easier to get that kind of feedback when there's something for people to play with than it is to say, like, you know, what do you want? I don't know what I want. Give me something to try. The same is true with, like, localization for Paid Memberships Pro. We have a localization server. We are putting AI translations out to our customer sites. The people that were contributing translations on their own, zero. The people that are fixing translations that AI created, 100%. Because they're like, oh, it's easier to just fix this translation string than it is to start from scratch. They will take a 99% complete translation with fuzzy strings and get it to 100%. --- ### [Ship Low Risk Features Early](https://share.snipd.com/snip/a553a376-193d-47ac-be01-24b31ba04a03) 🎧 14:01 - 14:34 (00:33) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/2e75b493-1439-4c8e-ad57-99ddcc062ef5" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - For low-risk AI releases, push early and let users be testers. - Kim recommends shipping simple utilities like a snow map quickly, gather feedback, then iterate rather than holding back for perfection. #### 💬 Quote > If it's low risk, push it. Let users be the testers. Let them try it and say it's broken. > — Kim Coleman Kim Coleman on using users to validate low-risk AI features #### 📚 Transcript **Kim Coleman:** Like, I'm trying to think what the worst thing that could happen is. So I think having a human sense of the security and the risk of it being incorrect is kind of an important thing. But also, as quickly as that code was created, that's how quickly bugs can be fixed or faster because it's a day later and everything got way faster already. So we're kind of in this mindset of if it's low risk, push it. Let users be the testers. Let users be the stress testers. Let them try it and say, I wish it did this. Let them try it and say, it's broken. It doesn't work this way. --- ### [Domain Experience Trumps AI Clones In Emergencies](https://share.snipd.com/snip/dd9d59df-9dd1-4097-812e-919e669ca7d2) 🎧 18:20 - 19:38 (01:17) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/e6ad5d80-3b00-49f4-8aa1-5865e458141c" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Domain knowledge and user empathy beat generic AI clones for emergencies. - Matt's snowplow map prioritized zero logins and fast reporting, unlike other creators who built social-style apps that failed in crisis use. #### 💬 Quote > I built it knowing what the end user is going to want in an emergency situation. That is zero logins, zero stupid social media like chat rooms, and the fastest way possible to report an emergency. > — Matt Medeiros Matt Medeiros on designing the Snowplow app for emergency needs #### 📚 Transcript **Matt Medeiros:** There's things that you know, like there's things that you have decades of membership software experience. So if somebody comes along and goes, well, I can just replicate Paid Memberships Pro. I can just put the repo into Cloud Code and I can just clone it and just move on and they could uh but they don't know the business and even all of the chats that you might have with cloud code it's not going to give you the 20 years of like customer relationships and experience that you've built same thing when i built the snowplow app i mean i'm not a developer but i had people messaging me saying oh oh, I was building one too. And, you know, he was like, take a look at it, see what it is. And it's like user registration, user profiles, like a social media site built in. I'm like, this is somebody who's just having fun with vibe coding. I built it knowing what the end user is going to want in an emergency situation. That is zero logins, zero stupid social media like chat rooms, and the fastest way possible to report an emergency. Everyone else that threw their app at me during the storm saying that they did it better, it was like they were replicating Facebook for snow plowing. It's just like that ain't going to work because you don't know. Unfortunately, you don't know what customers need or what users need in a situation like this. --- ### [Store Projects In One Folder And Update Status](https://share.snipd.com/snip/8043b239-f8a2-4aff-a2dc-08ed45d3219e) 🎧 19:55 - 21:09 (01:13) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/913a2f81-d23f-4b25-baa0-6e6b450d7d5f" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Keep experiments organized in a single folder and update a status file after work. - Kim stores all AI projects in one folder and records a status summary to avoid losing track of repos and keys. #### 💬 Quote > Every time I work on AI, I use the same folder on my computer. All the projects exist in it. And before I close, I say update the status file with all the things we worked on today. > — Kim Coleman Kim Coleman on practical organization for AI projects #### 📚 Transcript **Kim Coleman:** And I think WordPress products that are open source and coded to be customizable used to get a really bad rap. So Paid Mortgage Pro is probably known as the free option. It's probably known as the one you need to use code to customize. And that used to be a point of friction. And for us, it was always the value, right? For years it was. But we don't ever tell you no. There is no no. There is just yes with this code. Yes with this hook, you can get it done. And it's interesting to be here now where people are no longer afraid of that and they're valuing it more. If you're on a hosted membership platform, for example, there's no way to drop custom code in, at least most of them, and at least currently, to tweak the way something works. So yeah, I think that positions WordPress and other well-coded open source pluggable plugins in a really good place for where we're at. Yeah, Matt, I a solution for that. I don't even know where they're all located thing. I don't know if you're doing this, but every time I work on AI, I the same folder on my computer. All the projects exist in it. Maybe you do this too. And before I close, I say update the status file with all the things we worked on today. --- ### [Bulletproof Workflows With Agent Memory Layers](https://share.snipd.com/snip/8b9f8049-e3a2-4190-b975-c04cccea21aa) 🎧 24:24 - 25:21 (00:57) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/953d30af-5523-4f96-bfe6-ed43617415ca" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Use agent layers with shared memory to avoid vendor lock-in and preserve context across LLMs. - Kim highlights AutoMem and multi-LLM agents so tools can pick the best model per task and keep project memory. #### 💬 Quote > For people that are using agents with AutoMem, they are not linked to a specific API, a specific AI tool. > — Kim Coleman Kim Coleman on using AutoMem and agent memory to avoid single-vendor dependence #### 📚 Transcript **Kim Coleman:** So I don't know if you know the WP Fusion developer, Jack Arturo. His company is very good plugins, but he has something he's open sourced and is working on called AutoMem. So for people that are using agents with AutoMem, they are not linked to a specific API, a specific AI tool. They are choosing between the ones they are connected to what to use for what outcome. So that's more like a human saying how they want to work on something, where they want to go, making decisions based on what's going to work best for them and doing it. So as quickly as possible, move to something like that where you aren't using one tool specifically, that you have an agent that makes those decisions for you and has the memory of what you've worked on already stored. So I don't know. That's how I'm approaching it and why I'm not worried about it. --- ### [Keep Play Projects As Learning Artifacts](https://share.snipd.com/snip/fe575e56-e55d-4d61-ac43-c32bf43e2d78) 🎧 31:34 - 32:43 (01:09) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/59334917-86b2-4dae-8303-2b10a86f38a0" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Treat some AI experiments as learning projects and be comfortable not shipping them. - Kim advises saving prototypes on GitHub as resume items and evaluating attitude toward AI when hiring. #### 💬 Quote > Some things are just done for fun and just be comfortable with that. Ship it to GitHub and do nothing with it. It's a resume item. > — Kim Coleman Kim Coleman on the value of unsold AI experiments as learning and hiring signals #### 📚 Transcript **Kim Coleman:** would say some things are just done for fun and just be comfortable with that, be okay with that. Put in the reps, learn the systems, see what AI can do for you. And it's okay if you don't ship those things. That's okay. What did you learn from it? Go that route. Ship it to GitHub and do nothing with it. It's a resume item. I'm definitely, when I'm hiring people, I'm looking at their usage of AI, their sensitivity toward it, if they're skeptical of it, if they're trusting or untrusting, if they have concerns about it, that's a no-go for me, because I know we're going to all need to be using it. So just their attitude toward it, that's important. I would say like loop AI into the, not just like the shipping of it, but the marketing and the nurturing, like Matt was saying, maybe you're not there yet to let it run, but at least ask, you know, does this have legs? Run an analysis of if this is something I could charge for. I think we're not, we're looking at it as coding products for us, but we're not looking at the AI as like a mentor on top of that, or a marketer on top of that, or a salesperson on top of that. --- ### [Ask AI To Document And Mock UI Early](https://share.snipd.com/snip/1f56c743-9be8-45eb-b58b-9fdf68060474) 🎧 33:03 - 34:25 (01:21) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/c204d7eb-d693-4e9e-a6e4-4010ce046ade" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Ask AI to document architecture and generate quick HTML mockups during design. - Matt asks agents to create architecture docs and three HTML UI variations so he can iterate locally before heavy frontend work. #### 💬 Quote > I'm constantly asking it to document everything as much as possible. ... I'll often ask the agent to quickly mock this up in HTML. Give me three variations in an HTML file so I can just view this locally. > — Matt Medeiros Matt Medeiros on using AI for documentation and UI mockups #### 📚 Transcript **Matt Medeiros:** And this is like if you're just getting into this stuff and trying to figure things out, like I'm constantly asking it to, you know, document everything as much as possible. Create a systems like breakdown of I just especially if you're in the territory of building stuff that you've not done before, especially if it's like outside of WordPress. Making like an architecture document just so you completely understand like where the requests are going. Again, I do a lot with Cloudflare, so it's like understanding where are the functions, calling, you know, the database and at what layer and where is it sending this data. If we're doing edge caching, like help me understand that. That way, one, it's great to bring it to another, you know, agent in the future for it to understand what's happening, but also for your own understanding of every time you launch a new feature, constantly asking it to document. The thing that – and another thing that I'll do is because what I've found anyway, at least from my workflow, is design and UI is still not solved 100%. There's a bunch of apps I've been testing this last week, which are doing an okay job, but still not great. I'll often ask the agent to, like, if I'm launching a new feature, it's something like, hey, quickly mock this up in HTML. Give like, three variations in an HTML file or files so I can just view this locally on the browser. I'm going to tell you which one I'm going to pick, and we're going like, build off of that. --- ### [AI Boosts Output But Humans Still Sell](https://share.snipd.com/snip/c4fb2ebe-27fd-4d5a-b01c-c7ef380cda35) 🎧 33:38 - 36:36 (02:57) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/624cf4d1-f7ed-41de-9524-8f3c1436944f" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - AI accelerates productivity but doesn't replace human sales and relationship work. - Matt stresses agents can't yet handle nuanced sales conversations with governments or long-term trust built via podcasts and direct outreach. #### 💬 Quote > To still land the deal, it's still happening face-to-face. You still have to be a podcast, doing a YouTube video, doing webinars, emailing people and responding to them in a human fashion. > — Matt Medeiros Matt Medeiros on the ongoing need for human-led sales and relationships #### 📚 Transcript **Matt Medeiros:** That way, one, it's great to bring it to another, you know, agent in the future for it to understand what's happening, but also for your own understanding of every time you launch a new feature, constantly asking it to document. The thing that – and another thing that I'll do is because what I've found anyway, at least from my workflow, is design and UI is still not solved 100%. There's a bunch of apps I've been testing this last week, which are doing an okay job, but still not great. I'll often ask the agent to, like, if I'm launching a new feature, it's something like, hey, quickly mock this up in HTML. Give like, three variations in an HTML file or files so I can just view this locally on the browser. I'm going to tell you which one I'm going to pick, and we're to, like, build off of that. Basically, that's just, like, into again this is just my workflow this is just insight into i'm not just burning tokens constantly being like can you move the button to the left can you move the button to the right you know like change the label of the button and it's like rewriting all the javascript code or all the front end code i'm trying to to take it in steps especially with design because that's usually the biggest burn for me. The last thing they'll say from a conceptual point of view, I guess, or a business point of view, because I know everyone is trying to figure out how to, you know, make money with this stuff or improve their margins, is you still have to be a great human communicator right now. I mean, unless Jason has found an agent system that is like doing cold outreach emails and landing customers, I still don't know how you're going to make the sale of your thing if you can't be out front talking about it. You know, I saw a lot of people talking about like I actually made a video just to test things out. Like made a video with Cloud Code of the Snowplow app when I was launching it. And it like created the whole MP4, like showing interactions in the browser and doing all this stuff. And it was like really cool. But that would never sell anybody to buy it. I'm having deep conversations with local governments, and I'm a human, and I am experiencing the sales process. There's no agent in the world who's going to do that for you right now because the person on the other side is not even going to what to say to some kind of automated system. So you still have to be, and this is where you can spend your time on your marketing, on your messaging, on your connection with the customer, and hopefully the right customer. And it's, you know, all of this stuff that I've been doing with AI is great on the productivity side and the coding side and the brainstorming side, but to still land the deal, it's still happening, what I'll call face-to but you're still doing a podcast, doing a YouTube video, doing webinars, emailing people and responding to them in a human fashion. --- Created with [Snipd](https://www.snipd.com) | Highlight & Take Notes from Podcasts