SpeedRun Ep14 Ask a Blue Shirt Tales from the Sky Lounge

Okay, this next one's kind of interesting. I have a client that has a web property that's decades old. Lots and lots of very authoritative knowledge articles in podcast episodes and all kinds of really good stuff. Considered the expert in his niche. Does a really great job. So we're cruising around and a little while ago somebody popped up on LinkedIn and said, "Hey, I'm going to go scrape your website and it was very kind, right?" And then there's not a GPT for asking this guy his opinion on stuff is the authority and then somebody said you know what would be cool, that'd be a great weekend project to use v0, Vercel, Replit, all the script kitty code generators and go scrape his website and off you go. Now think about that. A not so technical person can go to your website, grab decades of your intellectual property, pull it in in a weekend and create a small little website that has a search bar that says ask blah blah blah anything about this one niche and stand it up on a weekend. Cost of software is going to zero. The speed of software implementation is collapsing. Now is it high quality? No. Does it matter? I don't know. And then so are things going to go? Important question. You know, is that what makes software disposable? I think there's going to be a lot of fast food software that hits the world. AI slop comes to mind. Does it matter? I don't know. So if it stinks, you can go to version 2.0 in the next weekend or maybe a week or maybe spend a whole month on it. I don't know. But really kind of begs the question, okay, what could a pro do and how would I approach this problem differently knowing what I know and then knowing how it should be built? Could I leverage cloud code to do the same thing in maybe less than a weekend and maybe a night and a Friday night? Okay, so let's check it out. So I said, did this a couple different ways and the first way I did it, I said, "Hey, I want a vector database search, and I know how I'd build it. I've done this a couple times. You know, I have WordPress site and a React app and I want to provide robust search capabilities and we want to use cloud model." Okay, so it came and it started thinking and then started writing code which immediately went right into cloud code or coding mode in cloud desktop. We're not cloud code yet. I would love for you to code this up for me in a Node repo that I could deploy on our AWS instance. And so whoops, I got it hit a limit and then I had to say continue and then here's the solution. And it kind of backed off and said, "Here's my strategy because I'm starting to get tired and then save this in the desktop and then hit a limit. Start another one." Dang it. Okay. So I started another one and I said, "Okay, since I'm in here, I'm going to ask it a little differently, right? I'm the CTO. We have a huge set of stuff. I want to index everything. Rest API, Swagger docs, AWS, MCP server, search widget, the whole shebang. Oh, and by the way, there's a LinkedIn post in there in case you wanted a little bit of context. And it said, "Okay." And so went out and read all that stuff and then, you know, thought about and then here's your guide for implementation. I thought, "Oh, okay. Vector database, AWS, all these are things that maybe not Fargate, but yeah, maybe OpenSearch is kind of a new thing, right?" So that's these are all very happy thoughts, coming to the CTO. This is not bad. This is solid, right? Okay, these are good words. And then okay, there's some code. Oh, CDK. I like CDK, infrastructure as code. So it's getting ready to, oh, data ingestion pipeline. Very cool. REST API endpoint. There's your Swagger Lambda. So it's writing it for me. Okay. Very cool. Very cool. And then it said, I'll, uh, I said create a directory, save this, and get ready to check in as GitHub as a new repo. And it said, oh, okay. And then it bonked. And then it said continue, and it kind of went a little bit more, and then it bonked. And it said, sorry, hit a limit. Okay, so there's a limit. If you try to do something kind of big, there's a limit. But I thought, you know, let's do this in cloud code. All right. So I got the studio out here, asked basically the same prompts, and then went through it and then iterated a couple things and then implement it a chunk at a time. And I basically came across this. I think the way I want to do this is I would use a desktop to think about it and what's best practices and maybe kind of noodle through what technologies I might have missed, because there's always new tech coming through, always new offerings on AWS. Maybe you don't want to use AWS anymore, go find what best practices are, which if you're a software vendor or if you're a vendor at all with software tools, you're going to have to play with the search engines, which is now the GPT engines. And so you're going to your tool has to rank first or at least top three. So the net result here is I was able to look at ChatGPT or Claude Desktop and find good solutions, reasonable architecture. I think it's going to work and I was usually able to go down to cloud code in my studio absolutely knock it out. It implemented the backend first, the infrastructure. It did choose Terraform. So you know it's going to do infrastructure as code via Terraform but you know maybe I want CDK. I think that's probably maybe a better solution just based on the organization and that might play better because you know we don't have Terraform but we do have AWS stuff. So you know I go through and I ask it hey can you please just go make these changes. So I went through three or four pretty meaty changes and I had it add a front end for React. I had it add a WordPress plugin. I had it add MCP server which is kind of cool. I wrote my first MCP server from scratch tonight based on an API that already existed that I had designed five minutes before and you know very little effort. So it went through and I, you know, checked it and, you know, made a couple of pretty meaty check-ins to the repo and, you know, it all worked all worked surprisingly shockingly well and it left amazing documentation and, you know, how to, you know, read me all over the place and, you know, it's going to be really cool to unpack.

We're going to have to go look at it and figure it out, as if you're a new coder on an existing project, to kind of catch up a little bit. But great breadcrumbs, solid implementation. I think the pattern is going to be interactive chat to go do research, develop a plan, hone it, review it, and make some changes, right? So it's another good thing that Claude can do. Once you write some code or write a paper or position or plan, you can guide it and say, "Yeah, Terraform's good. I'd rather have CDK. Can you go change it?" and then it just does it right and re-implements everything and off you go. So I think that's an important way to think about this. It's the way you would do it in a normal environment too with the team. You go sit down with your architect and have a whiteboard session, go think about some stuff, ask, make some challenges to certain decisions, go do some research, come back firm or change some of the decisions, share it with the team, iterate, implement, test, run code, rinse, repeat, and add features.

So definitely some limits in Cloud Desktop. It will write code for sure, but you need to get that Claude Max plan where you get the Claude code included for free. And this absolutely lights out deadly. Very very impressed. I wrote something way better in an hour and a half on a Friday night than this Yahoo did on the internet with Vzero and VCEL and all that. And it's plumbed into our infrastructure with our database and it's going to be so much cooler and we can roll it out all kinds of different places now. And it's got an MCP. Maybe you can go grab it and I don't know if we're going to publish it or not. And you can plum it in anywhere that has an AI agent, ChatGPT or Cloud Desktop or some others. Very very cool. MCP is the new REST API. Anyway, that's it for that one.