Commerce Today Episode 130
===

Joshua: [00:00:00] So you remember when an e-commerce re-platforming project, whether it's Magento, Shopify, Adobe Commerce, it might be like a six month slog or a year long fight. Last week I actually had Claude code ship, a proof of concept re-platforming in a single sprint, actually half a sprint, basically in a single week with zero hand typed PHP.

Now this was a test running in parallel with a real project. Was not quite ready to put this into production yet, but the surprising result actually was we could put it into production. So I realized that the world of AI has changed even more than I think a lot of people realize. And it's gonna start showing up in your e-commerce projects really soon.

So I dove right in. But I am Joshua Warren of the e-commerce Agency, Creatuity and Host Commerce today. And today I wanna look at really what's been a blur so far this summer in the [00:01:00] past just really past handful of weeks, there's been some big developments in. AI powered coding. And these coding agents that the different AI companies have been releasing anthropic has released Claude Code and most recently some really amazing real time MCP hooks.

If you have no idea what that is, that's okay. I just learned it really in the past few weeks. It's basically a way that. Your coding agent can interact with just about anything. So there's one for GitHub, there's one for Jira but there's even one for Figma for pulling mockups, a lot. Open ais rolled out Codex, which is their new coding agent, and it's built on their latest and greatest model, O three.

And then these, iDE focused coding agents like Windsurf have released production ready versions and, came out of the beta phase. So AI has gone from, helping you with code completion, code assist to being a pair programmer that is sitting there working with you to now, possibly [00:02:00] making up almost an entire full stack e-commerce development team. And so I think we're gonna see some big shifts. We're gonna see differences in how agencies work and how projects are specified, maybe even in what the budgets are, and how they're priced out. So we're gonna dive into all that today.

So I've already mentioned some of the recent changes. Also just a couple weeks ago, OpenAI released O three Pro, a more powerful version of their O three model, and it's only available on certain plans. So most of y'all probably haven't experimented with O three PRO yet. I've been using it since it was released, and it is incredible.

It is a massive leap forward, and I know that. Open AI is actually teasing that their next new model could drop any day. Now they said basically by the end of the summer we'll see chat GT five, which is gonna be another leap forward. So it's just crazy what these models can do in surveying developers.

I have gone from hearing people say, oh yeah, it makes me a little bit faster, a little bit better. [00:03:00] To really with clawed code in the past month, developers are saying that literally in the span of a month, they've seen a 20 to 25% increase in their productivity or more. So really amazing how fast it's moving.

I will say that if you looked at AI powered coding for your e-commerce projects early on, you probably heard lots of people including me, saying, oh, these things. They can't really replace a developer. They're not there yet. You need a human in the loop and you do still need a human in the loop.

But you probably also heard some complaints of oh, it really doesn't know what it's doing. Like it's okay on very small tasks, but once you give it bigger tasks, it loses the plot. Doesn't quite know what it's doing well between the new tools that have come out and the best practices that are emerging.

Like I mentioned earlier, actually, it can build an entire Magento site now. With this there are, I mentioned some evolving best practices. I wanna share what I have found that's working best. So if you are a developer out there trying to use AI to help you in your [00:04:00] development work here's a few things that you absolutely should try.

First of all, don't just type in a prompt to the LLMI know we all think. That's how you use LLMs and for a lot of different things. That is, but especially with a tool like Claude Code. Claude Code can actually use any command you have installed on your computer. That means it can use the GitHub command line interface and it can pull information from GitHub.

Maybe you see where I'm going with this? Define GitHub issues. Don't try to cram everything into a prompt. Define a GitHub issue that is just like you would do for if you were just delegating a task to any other developer. Write up some good specifications, what you expect outta that, and put it all on that issue.

Now, the real trick here is. You'd actually use Claude Code in planning mode to even write that issue for you. So you can talk to Claude and basically say, help me plan this issue. Help me plan this approach. Analyze the code base based on that. Plus these requirements define a really good specification for [00:05:00] what we want to implement in this GitHub issue.

Next thing you need to do is prep work, is have tests. I've been talking about test driven development. For over 10 years, and most of you haven't been listening to me. Test driven development was something that, especially on the agency side, so many agencies, they love to talk about tests and test driven development, but they don't actually do it.

That's been the dirty little secret in the industry. But those of you out there that actually have been implementing tests on your projects and test driven development. You are really gonna appreciate that now and it's gonna really pay dividends in the AI world because if you have tests then the AI can do so much more for you and you can have so much more confidence in its output.

And let me explain. So I have basically a six step loop. So I mentioned I don't type the prompt in, directly to the LLM and say, implement this and then define the whole issue. I actually refer to a GitHub issue and what I do is I tell it. Read and understand this GitHub issue, and I give it the issue [00:06:00] number and then follow the six step loop.

Step one is comment on the issue with your plan of how you're gonna implement this feature. Step two is write the code. Step three is write or extend any tests to make sure there's test coverage for your new code. Step four is run your pre-commit hooks, basically your things to check your code style and formatting and run the test locally and make sure they all pass.

Step five is push the code into a branch on GitHub, submit a pull request. Step six, watch GitHub. And this assumes, another prerequisite here is you need to have GitHub action set up running your tests. So step six is watch GitHub and make sure that the test pass that your continuous integration system shows all green.

If it does, you're done. If it doesn't go back to step one, repeat in a loop until you're done, and I've actually done this with some pretty big features that Claude would work for 30 or 45 minutes completely unaided, unassisted on its [00:07:00] own. Going through this loop of figuring out how it's gonna solve it, writing the code, running the test, seeing a test failed, fixing that, and just going again and again.

Until finally it tells me, all right, the tests are all green, the pull request is ready to merge. One other thing that has really helped in implementing complex e-commerce projects with things like Cloud code is these tools can read. There's two different specifications. There's an agents MD file or a Claude MD file that can have information about your project, information about your coding standards, your expectations write those, use those. There's a lot of examples online of some best practices. But I would also encourage you to experiment with this. Windsurf actually has a file called Windsurf Rules.

And they have a public repo that you can view that has some good examples as well that you can pull from. But basically the most important things that I include in these files are things like testing, test coverage requirements, this pull request [00:08:00] process. I also will ask log code whenever it has a problem implementing something.

I'll just say, what are the top three things that if you had done at the start of this task would've prevented these problems from occurring? I'll then actually read those and if they make sense and look good, I will add those back into my agent's file and my claw file as new rules for the LLM to follow.

So we're really, as the models are getting smarter, we're also getting smarter about how to use these models, and I think that's why we're seeing this exponential increase in developer productivity using AI tools. So talk some about this Magento site that I built. In a week. And I wanna say also, like I haven't written Magento code in a while.

I get to talk to a lot of developers and I get to talk to a lot of merchants. I get to see a lot of Magento code, but typically my day-to-day does not involve writing Magento code. And that's one of the reasons why I wanted to try this. I actually, about six months ago ish used [00:09:00] windsurf. To develop and deploy a test shop wear site because I wanted to try something on a platform I'd never worked on before.

And so bringing it back to Magento was super interesting. And I decided this wasn't just gonna be, develop locally a Magento store. I decided to take a client's requirements from one of our real projects and go the whole way and say, Hey. I want you to set up a local Docker environment for me to be able to test on.

I want you to set up my AWS environment to deploy to. I want you to set up the whole infrastructure and manage it all, and so I. I tried Codex and I like open ice codex and I'm using it on some other projects, but Codex really struggled here. I tried Windsurf and that was okay, and it was really Claude code -once I deployed this approach that I mentioned earlier, the six steps that loop. Who was starting with those GitHub issues? Claude code is where I had the most success. And I honestly think that if I was gonna do this again, it would be a one or two day [00:10:00] process, not a week long process. Using that, and this wasn't just, again, a basic site, this was all of the requirements that this client had.

This actually involved developing some custom modules, developing a custom theme. I actually used Hova on it. So it was even switching out Luma to Hova. And yeah, it is going to change our industry. And I know that there's been a lot of people saying that AI is overblown. And I know there's a lot of people that have been saying AI is coming for all of our jobs, and I tend to fall somewhere in the middle.

But I do think that we are gonna see more disruption with e-commerce agencies and e-commerce development over the next three, six months. Definitely the next year. Than we've seen in the past 10 years. Like it's gonna be big. For instance, one of the things is I just don't see us buying really simple extensions or services anymore.

So what I mean by that is, especially on, e-commerce sites that maybe had a lower budget for customization or that were developed by [00:11:00] developers that didn't know Magento as well, you would often see 20, 30 different extensions or third party services used. And a lot of those would just do one or two basic things like a classic from way back in the early days of Magento is these little image labels that you could put on your products that would be like, Hey, this product's on sale, or This is a best seller.

Like those little graphical tags used to, there's an extension, probably just a $50 extension that you would buy that would do that. If for some reason that requirement came up on a project, now I'd have Claude Code write that. Like it would be so easy for Claude to write that. There's no sense in the merchant paying for an extension to do that.

So I think a lot of these single purpose extensions, they're just gonna go away or they're just not gonna be used anymore because you can generate a module or a customization. Cheaper and easier and simpler than deploying that extension. So that's gonna be interesting. I also think that billing is gonna have to change.

And we've already been seeing a shift. It's been interesting talking to [00:12:00] other agency owners, talking to others in the Magento Association. And really just even seeing the competitive nature of recent bids there's been a shift in the market already for pricing and just the scoping and size of e-commerce projects.

But there's gonna be an even bigger shift now because if agency A is gonna deploy a team of developers and it's gonna be a six month long project, we're gonna put a ton of hours into it. Obviously it's gonna be more expensive and they're gonna most likely charge by the hour. That's how most agencies do it.

But let's say you have agency B and they have a couple of engineers that are augmented by things like Cloud Code, these actual coding agents, the billable hour doesn't really make sense anymore because like I mentioned earlier, I had a whole task, whole features, whole branches that Claude code would implement and Claude code would work on it for 45 minutes.

But I wasn't actually. Doing anything on the project during those 45 minutes. So do I bill for [00:13:00] my, one or two minutes of working with Claude Code and starting the task? Do I bill for the 45 minutes that took Claude code? Do I bill for the, 40 hours that amount of code and that level of functionality would've taken a developer to implement?

I don't know. I think value based pricing now suddenly starts making a lot more sense. A fancy way of saying fixed price, but I think that's starting to make more sense. And overall though it may be hard for merchants to make decisions of who to hire in the next six to 12 months because comparing apples to apples between an agency that is using and leveraging these coding agents versus one that isn't, it's gonna get really tricky.

And then of course like anything else. There are different levels of competency with these tools. So you may have some agencies that have super low bids, but it's because they're not developing that six step loop that I outlined. They're not putting the time into getting the task to find really well upfront.

They're literally just handing an [00:14:00] LLM a copy of their meeting transcript with you and saying, go build this site and yeah, you might get an okay site, but it might be pretty brittle. It might not actually run very well, especially if there was no testing involved or test driven development just might not actually function that great versus an agency that is using ai, but has that testing scaffolding in place has, a little bit more overhead and a little bit more experience with ai.

They're gonna charge you probably more than that other agency. But you're gonna get a better result out of it. But when everybody's saying we use AI and they're also probably saying it's a proprietary process, we don't wanna show you our prompts and our processes it's gonna get real hard to make a good decision.

Definitely don't envy the merchants out there looking to make a decision for a major e-commerce platform in the near future. So that's a little bit about some of the concerns I guess from that case study. I will say merchants, there are a [00:15:00] lot of advantages for you here. Definitely we're gonna see the timeframes collapse.

So instead of new ideas, new features, taking months, they're gonna take weeks or days. And that means you can prototype more, you can actually prototype out an idea and even launch a small test on your site. In a matter of days, instead of, like I said, weeks or months. I also think you'll be able to reallocate more of your budget from the basics of building a site to actually growing that site.

I'm most excited though about that prototyping idea. I actually finished the draft for my new e-commerce Growth Playbook, the book that I'm publishing in about a month. And in looking at it, one of the things that came up again and again is the companies that are shipping more improvements faster to their website are doing better.

They're the ones that are the most successful with their e-commerce initiatives. And this just puts that into the hands of so many more merchants now. Like before, you might have to have a [00:16:00] bigger team or a bigger budget to be shipping and deploying weekly or every two weeks. Now anybody's gonna be able to do that.

So I do think there's some big upside for merchants on this on agencies. Y'all. It's not gonna be simple. I think all the agencies out there need to sit down, look at how they're utilizing ai, how they could utilize AI better, and also just start having an honest, frank conversation about what your billing ought to look like and how you can stay in business and you can be compensated for the value that you're delivering to your clients, but also recognize the reality of.

More and more of this work is gonna shift from people to these tools and these LLMs. I also think smaller agencies have a bigger opportunity to compete now. You don't have to have a team of 20 people dedicated to a project to deliver an enters enterprise sized project now. So that's gonna be real interesting.

For developers, I guess first of all, for extension [00:17:00] authors you're really gonna have to think about the value that your code is delivering and if it can be easily replaced by an LLM and you need to really go deep on some specific features, functionality and like vertical specific knowledge so that you are adding more than just the code in your extension, that you're almost a consultant and someone that can enable a merchant to.

Better solve a specific problem. Now, developers I've had some interesting conversations over the past few weeks with different people, different developers that are embracing this technology. And I will say one surprisingly common thread is. It's exhausting. We all are feeling a little burnt out by it.

Like it's very exciting. You can do more, you can get more done. But we all kind of notice, hey, we're feeling tired, like more so than usual. And in digging into that, I think the issue is as a developer, you get into a flow state when you're coding, like when you are in [00:18:00] the middle of implementing a feature things just flow.

You know what you're doing. You can almost do it automatically. There's a joy to the process. With using something like Claude Code, suddenly all that flow state work, the AI gets to do it. And now you're having to just do the really hard work at the beginning and the end. So it changes up what the tasks look like and it changes up really how your brain is interacting with your computer as you are developing new features.

So I do think that, it's gonna be an adjustment. I think there's some people that aren't gonna make that adjustment and are gonna want to keep doing it the old way. And those are the people that I think are gonna find the job market get especially challenging. Those that are willing to embrace becoming prompt engineers becoming really solution architects and understanding how to better define tasks for the LLM.

They're gonna be just fine. They're gonna have plenty of job opportunities. And I think that if anything, there's a new type of developer [00:19:00] emerging that is part e-commerce developer, part AI engineer. And that's gonna be an interesting development. And I think even that's gonna change pretty rapidly.

I would not have trusted the LLMs out there a year ago. To successfully define a task for another LLM or for a human developer to implement the models that are out there right now though, O three Pro absolutely can write great specifications and user stories. So the models that are out a year from now, who knows what they'll be able to do, but I think the difference is.

The people that can adapt to that change and learn the new models the fastest are gonna be the most productive and the most in demand. And so that's why I think all developers, e-commerce developers that want to be successful you're gonna have to also become an AI engineer. Yeah, that is the state of the art of where things are at with AI right now and e-commerce development.

If a lot of this sounded crazy or far [00:20:00] futuristic to you I totally get it. I was reading a stat recently that, I think it's 60% of people still aren't using AI still aren't using tools like chat GBT at work or they're using them. I believe that stat included people that were using them, but only using them like once a week.

Only about 5% of people have ingrained AI into basically every task they're doing at work. And that's me. And those of us like me that are doing that, we're getting a glimpse of this future and. It is wildly different than where we have been in the past months with e-commerce development. So buckle up and there's a lot of opportunity out there for people that are going to embrace these changes.

There's gonna be a lot of disruption and it is gonna be interesting to see how our industry adapts to all this. I would encourage you, if you are a developer and you haven't yet tried Claude code, go try it out. Really focus on that Claude MD file and on. Working with Claude in planning mode to [00:21:00] write up a GitHub issue for a task before you actually implement it.

Also if you're a developer or even, a someone, tech adjacent, developer adjacent, and you've always had this personal project, either you're halfway finished with or that you've wanted to start try that out in Claude Code. I've been reading so many stories of people that took, their.

Five or six personal projects that they've had up on GitHub for years, never really did anything with. And in the span of a weekend, they got all of them to a place where they are making a meaningful contribution to their life. So a lot of cool stuff you can do. Definitely check it out. For the non-developers out there there is a lot that AI can do to help you in your work.

I also think that, helping developers better define tasks is gonna be super important. And really thinking through what does billing, what does the commercial arrangement between agencies and merchants look like? How does that need to change? That's some big important work that needs to be figured out really.

Sooner than I think a lot of us [00:22:00] realized. So with that, I hope you enjoyed this episode of Commerce Today. If you have any questions or if you wanna see like my agents MD file or anything like that, find me on LinkedIn. My name is Joshua Warren. I have a Creatuity gold background behind my headshot. Send me a DM there and I would love to help you out and I will see you on the next episode of Commerce Today.