Claude Code Sub Agents for Beginners: Building My First One

By Beau Johnson··7 min read

Quick answer

I asked Claude Code to look at my codebase and suggest three sub agents for easyflip.ai, then built the first one, an AI accuracy optimization agent. Getting it to read brand and size from photos took a lot of pasting browser console logs back into Claude Code.

In this video I ask Claude Code to look at my actual codebase and tell me which sub agents would help my project most. Then I build the first one and test it on real photos. I recorded this in August 2025, on day 13 of building easyflip.ai, an app for resellers that turns uploaded photos into eBay listings. It's for beginners who want to see what building a sub agent really looks like, errors included. Tools change fast, so check the current Claude Code docs before you copy my steps.

What I cover in this video

  • 1:31 The prompt I pasted into Claude Code to get sub agent ideas based on my real files
  • 3:50 How I understand sub agents and why I wanted them
  • 4:35 The three sub agents Claude Code suggested
  • 9:09 Building the AI accuracy optimization agent
  • 18:20 The first test fails and I start pasting console logs into Claude Code
  • 22:54 The Google Vision API key that was missing from my local env file
  • 32:46 The real reason the size was never detected
  • 34:20 It finally reads the brand, title, and size correctly

Asking Claude Code which sub agents to build

In the previous video I built an agent, but it lived inside my app and was not a sub agent inside Claude Code. So over the weekend I asked Claude whether I should let Claude Code pick the sub agents, since it has access to my local dev files. It said yes and wrote a prompt for me. The prompt opens and closes like this:

Looking at my actual project files and architecture, what would be the top three development sub agents?

Please analyze my actual code structure and suggest specific sub agents that would hook into my existing services, functions, and data flows.

In between, I listed the three areas I cared about: improving the AI photo analysis accuracy, managing infrastructure performance and costs, and tracking business metrics. I also told it the end goal, which is 10K MRR.

The way I understand sub agents is that each one gets its own prompt in a markdown file, and you can call on it for a specific kind of task. It could be a front-end developer, a backend developer, or someone who handles your API research. I don't think it saves much on tokens, but I think it gets work done faster and keeps things organized for me.

Claude Code came back with three:

  1. AI accuracy optimization agent. Tracks prediction accuracy, compares AI output against the user's final edits, and A/B tests prompts.
  2. Infrastructure performance agent. Watches for slow queries, timeouts, and API failure rates across Netlify and Supabase.
  3. Business intelligence agent. Tracks user behavior and what drives revenue.

Claude Code suggested starting with the infrastructure agent. I picked the accuracy agent first because photo analysis is the main feature of the product. I want it to read keywords off tags, put them in the title, and fill in the item specifics eBay requires.

Building the AI accuracy optimization agent

I told Claude Code to build it right away. It posted updates as it worked: designing the architecture, creating a database schema for AI performance tracking, then building an accuracy monitoring service. The build took around 7 minutes and used about 22,000 tokens on the $20 a month plan.

When it finished, it listed what it had created: tracking for every AI analysis, an eBay specifics validator, an OCR keyword optimizer, a prompt optimization engine, and an AI Insights dashboard in the account menu. Before I could test anything it had to run a database migration to set up the tracking tables.

Debugging with console logs until it worked

My first test was a leather jacket with a clear photo of the tag. The result was AI analysis failed. From there I repeated one loop for most of the video: run the test, copy the browser console output, paste it into Claude Code, and describe what I saw. This is the vibe coding part nobody shows you. Here is what turned up along the way:

  • Schema validation error. A Zod validation error was breaking the final output. It took two rounds to fix.
  • Google Vision API not configured. I had the key in my Netlify environment variables for production, but I was working locally and my .env.local file only had my Supabase and OpenAI keys. I had also mixed up which key it was calling.
  • Brand in the description but not the title. A Wall Street Bull hoodie came back as "hoodie black size unknown" until I told Claude Code exactly what I wanted in the title.
  • Size never detected. Claude Code tried more aggressive size detection and multi image OCR, and it still missed a tag that plainly said M.

The size problem had a simple cause. The generate listings page was only sending the primary photo for analysis, and the tag was the fourth photo in the SKU group. Once Claude Code changed it to read every photo in the group, the listing came back with the right brand, title, and size. My last request was to spell out single letter sizes, so M shows up as medium, and it handled that too.

Near the end of the session I got a warning that I was approaching my usage limit. That gives you a rough idea of what the $20 plan covers: one sub agent built plus a long debugging session.

Key takeaways

  • Ask Claude Code which sub agents to build. It can see your file structure, APIs, and database schema, so its suggestions fit your project.
  • Start with the agent tied to your main feature and leave the rest for later.
  • When something breaks, paste the full console log into Claude Code. One time I pasted only part of it and had to go back and give it the rest.
  • If something works in production but not locally, check that your local env file has the same API keys.
  • This was not a single prompt build. I don't know what most of the code means, and feeding Claude Code context still got me to a working feature.

FAQ

What is a Claude Code sub agent?

The way I understand it, a sub agent is an agent with its own prompt stored in a markdown file that you can call on for specific tasks. Examples I mention are a front-end developer, a backend developer, or an agent that handles API research.

How do I decide which sub agents to build for my project?

I asked Claude Code to analyze my actual project files and suggest the top three sub agents that would hook into my existing services, functions, and data flows. Because it can see the codebase, the suggestions were specific to my app.

Can you build a Claude Code sub agent on the $20 a month plan?

Yes. I built my first sub agent on the $20 a month plan in this video. The build used about 22,000 tokens, and I got an approaching usage limit warning later while debugging.

Why was the AI missing the clothing size in my listings?

The generate listings page was only sending the primary photo for analysis, and the size tag was in the fourth photo. Once Claude Code changed it to read all photos in the SKU group, the size was detected.

Transcript

Read the full transcript

0:00Day 13 on this project and we made incredible progress and I think that we are about to make more progress in these next few days than in the first 13. We are now using cloud code. We're building out sub aents and we're starting to get a little bit more efficient in this project when it comes to building out easyflip.ai. In this video, we're going to be building out more sub agents and actually adding to the project, seeing how it can get better. And we're simply just going to be asking Claude, "How can we make this project better?" you're going to be following along in the process. So, in the last video, we actually didn't finish what we wanted to. We built out an agent, but it wasn't a sub agent that was not built within Cloud Code. So, that's what we're going to be working on today. I'm still happy with the agent that it built, but that

0:46agent is more implemented in my business when it comes to like when the user uploads the photo, it does the analysis, and recognizes the item, prices the item through the eBay API. That's great. And that's a great addition, but it wasn't what we intended to build. So, in this video, we're going to be going over that. And I had talked with Claude over the weekend. I said, should I ask Claude Code what would be the best sub agents for my project since it has access to the local dev files? Yes, absolutely. That would be the perfect approach. Claude Code has full access to our actual codebase, so it can be more specific on what it wants to build for us. So, um, Claude just came up with a prompt. Looking at my actual project files and architecture, what would be the top three development sub aents? So,

1:31these sub aents that we're going to be building in this video is um monitoring and improving the AI photo analysis accuracy. That's something that we're still trying to get better. Uh, it's failing to recognize like the size tag and also like putting those keywords in the eBay title is another thing. uh managing infrastructure performance and costs, tracking business metrics and development insights. Please analyze my actual code structure and suggest specific sub aents that would hook into my existing services, functions, and data flows. And of course, the end goal is to build this thing to 10K MRR, which we are already on track with almost 100 email signups, which is so cool. And I've barely even spoken about this project. So why this is much better

2:18implementing sub aents is it sees our actual file structure on our local. That's amazing. It understands our data flow how photos AI listings actually work. Knows our APIs, OpenAI, Google Vision and eBay integrations. It sees our database schema superbase tables and relationships. We're also going to implement the Superbase CLI today as well. and understands your business logic, revenue model, and user flows. So, this is going to be amazing. This is what we're going to be copying and pasting into Cloud Code right now. We are fresh out the kitchen and I just made my first commit from my local server to our production and it's amazing. Everything works. We're in all new territory. And why am I why I'm excited is because if you go back a few

3:05days ago on the channel and you're watching this in sequential order, we were using Bolt to build this out and everything was great for the first like 10 days and then the project just got so big we started really messing things up and it just really got frustrating. So Claude code was a necessary step in the evolution process of building this out. You can start with bolt or lovable or replet or whatever you want to do to like build out the initial idea and then I think eventually depending on how complex your idea is and what you want to add into it, you would eventually probably want to pick something like cloud code to build this thing out. Um so yeah, so let's do this. Let's uh let's copy paste this into claude and um let's just see what it says. Let's go. And just to break it all down from my

3:50understanding why sub agents within cloud code why they're important is basically you can build these agents that are trained on specific data within a markdown file. That agent now can perform tasks by itself based on whatever you want. Front-end developer, backend developer, someone that manages your API research, whatever you want. So that's going to be really cool. And I feel like it's not really going to save as much like on the tokens, but I think it's going to get stuff done faster, which I think is cool. So that's why we're doing this. That's why we're implementing implementing it. And this is like our first time really doing it. So it came back really quick with an answer. Brilliant question looking at your actual codebase architecture. Here's the three most impactful

4:35developer sub aents to help you scale your business. So the first one was a great one, the AI accuracy optimization agent. That is kind of like our main product feature that we would be marketing this product through easy flip.ai is like the AI um analysis of your of whatever you're uploading. You know, I want it to be able to be really good at this feature and I want to focus on that as much as possible because that is our main feature. I I have so many ideas for this project on implementing. I just want to focus on the main thing and building out that main feature the best that we can before we start adding in other other things that I've already been like, you know, running around with. So that's why we're focusing so much on the AI analysis and making that

5:21the best that we can through Google Vision and the GPT vision API. So hope that makes sense, but that's something I really just want to focus on and not get too sidetracked. So that's a great sub agent to start building out. Um, it's going to look at the existing code and it's going to build track prediction accuracy, optimize prompt performance, cost efficiency analysis, and basically the track prediction accuracy, compare AI versus user final edits, AB test different prompts, accuracy prediction, skip expensive calls for low confidence items, business impact is going to reduce our cost by 40%. Infrastructure performance agent problem Netlify functions timing out superbase hitting limits user experience degrading solution real-time monitoring and auto

6:07optimization. I love it. So that's cool. So it's just going to like monitor slow queries and timeouts, track API failure rates. This is great. Number three, the third and final sub agent uh for this video anyways is going to be the business intelligence agent. Our problem is you're flying blind on what drives revenue. to user behavior and growth solution comprehensive business metrics tracking. So hooks into your existing codebase is off context TSX which is a user life cycle tracking. Wow, this is just insane. That's like definitely more than what I was thinking but it's going to be good. Uh it's going to help us make datadriven decisions. So um so yeah uh week one infrastructure agent, week

6:53two business intelligent agent, week three AI optimization agent. Each agent would have a developer dashboard accessessible. Cool. All right. All right. Um return on investment. These three agents could easily save you 2,000 a month in costs while accelerating growth by 3 to 6 months. Um I like the first agent because this is our halo feature for the product. We really want the AI analysis through Google Vision and GPT vision to be the best it can be. I want the best accuracy for picking up keywords that it reads, putting those keywords in the title, also analyzing the photos to get correct item specifics that are required and suggested by eBay. This is the most important aspect of our

7:38entire project. So, we're just going to enter that in and just go from there. All right. So, perfect choice. The AI accuracy optimization agent is absolutely critical for your success. Let me analyze your current AI pipeline and design a comprehensive comprehensive optimization system. Amazing. Came back with a great overview. Based on your codebase, here's how we'll build the most impactful developer tool for your Halo feature. Core problems will solve. OCR, keyword accuracy. Are we extracting the right text from the tags and labels? GPT vision specifics. Getting correct eBay item specifics, title optimization, keywords and titles that drive sales, cost optimization, reducing API calls while maintaining accuracy, confidence calibration, when to trust AI versus

8:24request human review. Amazing. So, this is so good. And how I believe this works, and correct me if I'm wrong, is that once it builds the sub agent out, it's going to put basically a prompt within a markdown file for that specific sub aent so it can cook essentially. It can do work, right? And that's just kind of how I see it in my mind. It's just um we can call upon these specific agents to complete specific tasks, which just helps things be a little bit more organized, I think, for me. Not saying you can't do it without sub agents, but I just think for me to be able to build specific agents out that do specific things within the business is going to be really helpful in the long run for this project. So, let's see what is it saying here. So, I'm liking all that

9:09without reading through it all. Um, it's going to create a monitoring dashboard as well. Um, so this is great. Would you like me to start building this tomorrow? This agent will be the key. Um, let's build it today. I would like to build all three of these sub agents today, but let's start with this one first. One thing that I like as it's building this out and it's been working behind the scenes for maybe couple minutes. Um, it's just going through what exactly it's doing. So, so far it's designed the AI accuracy optimization agent architecture and then it just finished create database schema for AI performance tracking. I believe that's happening in Superbase. And now it's build accuracy monitoring service. So,

9:56don't know exactly how long it's going to take, but it's nice that it kind of just gives you an update as to what it's working on. All right, it's going through it. It's been about 5 minutes now, and I I love it. I I just sitting here thinking, you know, trying to keep busy, but also at the same time is like this is such a great time in history to build something, build your idea. And that's why this series is existing because I want to just help people like yourself that maybe you are from the outside looking in. You're seeing all these cool projects, apps, ideas come to life, but you just like don't know where to start. I was the same way just a couple weeks ago. Actually, months back, I had this very same idea. And then I just decided to go all in. Start messy. Start before I'm ready and just go. Hit the go button and record while doing it.

10:43So, I just want to like encourage you that maybe you have an idea right now that you're working on, but these tools may seem intimidating or you don't really know how to go about learning these skills to put your plan into action. Just I just encourage you is just start before you're ready. make mistakes, make errors, and it's okay if you mess up because you're going to learn from those valuable mistakes just like I have. There's been several days along in this project, like day nine, day 11, where I literally just spent all day fixing an error. Like, it was so frustrating. I was even telling my wife like, "Is this even worth doing?" Like, I just spent all day making no progress on the app, you know, more than eight hours. And it is worth it because the lessons that you learn along the way while building these skills up is so

11:28valuable. And I think if you're not taking the time to learn these skills is like you're going to be able to learn these skills and bring them into the marketplace, whether you want to work for a startup, build your own startup, or just create something cool that serves yourself, you know, that that just like a a walking app, a habit tracking app. I just think that we're living in such a good time to be a builder, to be a creator, and to take ideas into in putting them into action. There was a great tweet by Russell Brunson over the weekend that your idea is just a secret without marketing. So maybe you have a great idea that can actually help people. But if you're not talking about it, then you're the only one with that idea and that secret. So I just feel like, you know, everyone has a an idea to build something great and do something great. Now's the time to be

12:13that builder, be that creator. Come on now. Who am I talking to? This is a great day to be alive. Great time to be alive, to be building things. You know, could you imagine five years ago learning full-on computer languages in building this thing from scratch? Now we live in a time where I can just hit F1 on my keyboard, talk for nonsense to these to these uh vibe coding tools, and they will they will take out whatever out of whatever I'm taking. and you know, formulate plans, formulate business ideas, and that is just such a beautiful thing. And we're seeing it happen right now, y'all. So, it's so cool. So, this has been going on for about like maybe I would say 7 minutes now building out this first agent. Um,

12:58it's okay for me cuz I got things to do. It says it's used 22,000 tokens. I'm still on the $20 a month plan when it comes to Claude. I know I'm gonna probably eventually the next step will just be paying for that $200 a month plan which is supposedly supposed to give you 20 times the amount of tokens or context or whatever um to just build this out even more. So pretty cool. I'm loving it. I'm excited. And we're going to document the journey every single day y'all. I'm going to be working on this and I feel like by using cloud code it's going to help speed up the development process. So I'm thinking what used to think maybe 3 months to build this out. I think that within 30 days we can actually launch this product with the main feature and actually start getting paying customers. It's already in demand

13:44within this niche of resellers. People want this. People don't want to do listings. People just want to upload photos and the listings are done. They're priced. The keywords are optimized in the title. The item specifics are filled out and it's just plugandplay. Easy. It solves a simple problem. So, I'm hopeful, but we're going to track along the journey, y'all. And if you guys aren't subscribed and you like seeing AI startups come to life, please do me that one favor and just subscribe. Give me some feedback. I'm a lifelong learner, so I'm not proclaiming to know it all at all. So, please, I'm I'm here to to get any feedback as well. So, yes, I I appreciate all the new subscribers. We're almost at 100 subs in just 14 days. That's crazy. So, let's see what uh Old Cloud is doing and where we're at. Cool. So, it just finished. Let's

14:30see. I've successfully built your AI accuracy optimization agent. Your Halo feature optimization system. Here's what I've created. AI accuracy optimization agent. Complete system. Core components built. Um, database schema AI predictions. It tracks every AI analysis with detailed accuracy metrics. eBay specifics tracking validates eBay compliance and accuracy. That is the main one there too. Um, meaning that eBay won't let you create a listing until required item specifics are filled out about your product. So, say you're listing a men's button-up shirt. Well, it needs to know the size of that shirt before you can list that item. And if the size isn't read correctly from the AI, then it's not going to be able to up

15:16the user will have to interfere and do that manually, which we're really trying to avoid. Um, AI accuracy agent tracks prediction accuracy in real time. So this is like all the features I believe that are built into this one agent uh the core components. So eBay specifics validator OCR keyword optimizer prompt optimization engine just really cool. This is just it blows me away. It really does. This is just so cool. Um key features real-time performance tracking. Every AI analysis is monitored for accuracy. Automatic cost optimization. Track spending versus accuracy versus for efficiency. Smart recommendations. Personalized tips to improve your AI

16:02accuracy. Cool, cool, cool, cool. All right. So, how to use access dashboard. Go to your app account menu. AI insights. Monitor performance. See real-time accuracy. Costs and trends. Follow recommendations. This is insane. Track improvements. Watch your AI accuracy over time. Improve over time. Integration points. The system automatically hooks into your existing APIs. This comprehensive system will help you achieve the best possible AI accuracy for your clothing analysis. Your true competitive advantage. The AI accuracy optimization agent is now ready to help you scale to 10K a month by ensuring your AI delivers consistent excellent results for keyword extraction. Okay, so cool. All right. Amazing. Am I able to test this out? Is

16:48the AI analysis better? Should I upload some photos to see if it has become more accurate? Okay, so before we could test that feature out is it's going to run a database migration to set up the AI accuracy tracking tables. Then we can test it out. Okay, just after about 2 minutes, it says your development server is running smoothly and we can see the hot module reloading has picked up our changes. The AI accuracy optimization system is now integrated and ready to test. You can finally test this now. What's better? Enhanced AI analysis. Every photo you upload now goes through four layers of optimization. Um, cool, cool, cool. Better brand size detection. The system now cross references OCR text with AI vision for maximum accuracy. Smarter keywords. How to test? Upload

17:35new photos. Go to your app. Watch the console. All right. It says, "We'll we'll just notice better brand detection from the clothing tab tags. More accurate sizes extracted from the label. commercial keywords like vintage, rare, premium. Uh, okay, cool. Okay, so we got to go to our incognito window. All right, so let's sign in. We're going to upload these photos from a leather jacket that we have. We know that the titles are clear. Everything is good. And we know that there's a photo of the tag, so it should be able to read that. And the photo is clear. So, so once the photos are uploaded, we're able to select all the photos and then assign a skew number. So, let's just say this is skew number

18:20100. Assign should go to the generate listings page. All right, I see it here. Generate. And we also want to work on the time that it takes to um generate these listings. So, so far you can see AI analysis failed. I also want to see if we're even on Oh, we are AI insights. Okay, I see it there. Okay. Um that's cool. So let's go back to why it failed. So let's select it. And then what I do is I just copy paste the Google console um into cloud and just give it like here's what happened. So AI analysis failed Google console. Uh I re recommend anybody doing that if you are experiencing just problems just give it to Claude code. That's what we pay the

19:06big bucks for. Claude, we got problems. All right. Okay. So we got problems in the AI analysis. I have copy pasted the console code. That's just what I had been doing. So um copy pasted that in there. All right. Now let's give it the details. See how fast it comes back with an answer. Okay. I love it. So I guess after I gave it all that um context, it says I could see several issues happening in the AI analysis pipeline. The good news is that the AI optimization system is working. You could see it running through the enhanced steps. What is working is the keyword optimization, eBay validation and enhanced analysis. The issues that we need to fix are the Google vision API not configured in OCR not working schema

19:52validation error breaking the final output and missing database tables AI insights dashboard failing. Let me fix the critical schema validation error first. So it worked through all this and it says perfect. Let's test and let's test the fixed AI analysis. Um it says the Google vision API is not configured. Interesting. Um but for now it says let's test what we have. The AI optimization is working without the Google vision API. It's just the Google vision alone which is still very effective. Try this now. Okay. All right. Back. I don't know. For some reason I just like to open a new. All right. So all right. So after making those changes, restarting the development server, I was able to log in. But now the analysis is still failing. So I'm going to tell that to claude code. All right. So I still see

20:37the analysis still failing. I have copy pasted the console code. And just like that it um it said the issue is still the zod schema validation error happening at line 40. So it made some changes and it says to test it again. The enhanced debugging will show us exactly what data is being passed and what's causing the validation to fail. Um okay. So we're just going to repeat the process. Let's come back over here. We lost it. Right back into generate listings. Let's see what happens. Jacket black size unknown. Okay. Well, at least it didn't have an error. Um, it did not read the brand, which is clearly in the photo, but that's cool. At least not too many errors. Let's go back. Okay, it

21:22finished without any errors, but also it did not include the brand name. It didn't include the size, which are major problems. We need it to be able to know the brand and the size by the photos that we upload. Is this because the Google Vision API is not connected? Do we need to integrate the Superbase CLI? I have copy pasted the console logs as well. Let's upload some different photos. Let's just see if maybe that was the difference, but let's try it. should clearly see it's a Wall Street bull New York and should just be able to pick that data up. Let's go. This is 55. Select all assign skew. And let's just see what happens. See if this makes any

22:07difference. If there's problems, we can copy paste hoodie black size unknown. Go to edit Wall Street Bull hoodie. I mean, it got that and it put it in the description, but it didn't put it in the title. Okay, I'm going to tell that to Claude as well. For example, when doing an analysis on a Wall Street bull New York hoodie, it puts those details in the item description, but I want those details to be in the title. Wall Street Bull hoodie should be in the title, and it's not. Okay. And then we're going to go back. Well, let's see what happens. Can I get the console log? It's not. Let's go back. Okay. All right. So, I just brain dumped all of the things that I was thinking into Cloud Code. and along with the console logs. Okay, cool. So, it optimized it as well. But I I do

22:54want to point out that this is key. I have my Google Vision API set up in my environmental variables on Netlefi on the production side. But since I'm working on a local level, I also need to go to my ENV.local file and put in the Google AI vision API into that document. So, that's what we're missing and that's what I need to do as well. And hopefully Claude code can pick that up since I'm doing it on a local level and access this because right now all I have is superbase uh a non-key URL and open AI API key in this uh file. So that's what I'm going to do right now. I'm going to go into Netifi and grab that point that I want to make. I'm on the $20 a month

23:39plan and I've been using it. You've been seeing it. Um, I haven't used it otherwise, you know, off air, but we've been building this agent out and it says approaching usage limit resets at 2 p.m., which is in 17 minutes. But, um, just to kind of give you a frame of reference on like what can you build with the $20 a month plan. So far, we're finishing up our first sub agent and kind of walking through some of the errors that we're going through. So, just keep that in mind. So, one of the main reasons why it wasn't working, the Google Vision API, is because I just mixed up the API key that it was calling to, it was wrong. It was um now that that's inputed into the local ENV file, I have the correct API key in there. So, it should now be able to use that to

24:25help read the photos. All right. So, we're going to re-upload the photos that were um in our database. And let's see if it can read these much better. So, let's try with this Worththington. Actually, let's try with the steel jacket since I think that that's more clear. Um, just to give it a little little softball to start it off. So, okay. So, that one we're doing. Let's start that off as skew number two. And let's go to the generate listings page and let's just see if this is better after all those changes. Cool. So, it got steel jacket blue. Still the size is unknown. Having a hard time with that, but let's see. Steel jacket blue. Size unknown. Yeah, it still doesn't get the size, but it does have at least it's reading some of it. Okay, so click X.

25:10We'll still copy the console code. Um, let's see if there's any errors. Looks like there's some errors. So, let's copy that just in case. Copy. Paste. Go back to cloud code. Awesome. Read the brand name and put that in the title, but still unable to detect the size. Here is the console code. Okay. The thing we're working on now is just focusing on size detection. So, it's got the brand red and it said to put this size pattern recognition uh I think as a prompt within this sub agent um to just be a little bit more clear on the sizing. So, let's see how that works. It says your system now has much more aggressive size detection. That should catch sizes from the clothing tags. It will examine the

25:57images more thoroughly and you could test it right now. All right. So, we are giving it another try after all those changes. Let's see what happens. We are trying to analyze the same item to see if it makes any changes. Wall Street Bull hoodie black. We got it. So, versus just hoodie black size unknown. Still didn't get the size. So, let's see here. Yeah, still missing the size, but Wall Street Bull. Yeah, just didn't get it. And even though the photos are right here for the size, you can see medium should be detecting that. Um, so that still needs to be worked on. Let's take the console code and just see. Looks like there is some errors in there. So, we'll just drop that in as well. It read

26:42the Wall Street Bull hoodie fabulously and put that in the title. It still was missing the size, which I uploaded a photo within that listing to include a photo of the tag size. It is a size medium, but it didn't pick it up or read it. Here's the console logs. Watching Claude Code cook is so good. Oh my goodness. Just reading the updates and the issues it finds during this development process is actually really helping me understand kind of how it goes about fixing these bugs or errors or maybe it just wasn't put in the markdown file correctly. I have no idea. Um, but it's pretty cool just to see like how it's debugging and figuring it out because now we officially have the

27:29open AI GPT vision API and the Google API for the vision uh to work now. So, let's see. Okay, it says I've significantly enhanced the size detection specifically to address the medium size detection issue, individual text processing, aggressive partial matching, enhanced debugging, Google vision detects. Okay, cool. And it's actually on the specific item we're uploading is actually just it just says M, but it should say it should know that M is medium by reading it on the tag. So, let's give it another go. All right, so at the start of the video, this is kind of what we were getting. hoodie black size unknown. And even though it completed the process, it still didn't do a very good job. So, here's where we're at now. Let's see if it got any

28:14better at developing a title, reading the photos, reading the tag in the photos, and let's see. So, it's still Let's see. Had a hard time getting medium. Um, everything else looks pretty good. Um, but it's just failing to get the size. So, we will get the console logs again. Um, maybe I'll just copy paste cuz no errors are going on there until about here. So that way it's not so much. Okay. So then we're going to go back. Claude. Awesome. Still not reading the size the tag in the photo. All it says is M. M is for medium. So it should be able to read this and analyze it and put medium in the title and in the item

28:59specifics, but it's failing to do so. Here's the Google console logs. All right. Well, it implemented some new changes. Multi-image OCR processing now processes up to five uploaded images instead of just the first one. We want it to ideally up uh process all of them. We want it to read all of them. So, um I will refer back to that if this doesn't work, but we'll go ahead and and try it. Um better debugging, comprehensive text collection, and enhanced single letter detection. So, um let's see. Sh. So, we should be good to go. All right. So, you can see the evolution of the process where we started, where we were, and hopefully on this third one, we can get it to work. We'll see. All right. Let's give it another try. We made a few

29:44changes and we'll see what happened. Okay, so let's go back and copy paste all this. Errors will not analyze now. And I made a mistake. I didn't copy paste all the console logs in our last request. So here is all the console log from the AI analysis. Okay. Something about being something between the server side on netlefi environmental variables and OCR text transmission processing. That's all I know. But here it gives you a breakdown what's going on. So what I fixed fixed function crashes fixed OCR text transmission enhance multi-image processing. And this should at least um also condition is a huge one. Detects

30:30clothing condition um whether it's new with tags, gently used or heavy wear, etc. So, let's give this a try. All right, let's see if it works. We're just keep using the same example and needs attention. Let's try this other one down here. See what's going on. By no means is it a oneshot code prompt to get your app going. You are seeing the process right here and comment down below. What am I doing wrong? You let me know. But here we go. I'm copy pasting all of this and still not working. Here's the complete console log. And I think the main point of view here for beginners is like I don't know exactly what's going on. Like I don't know what's breaking it. I don't know what's causing it not to read it. I don't know why we make two steps forward, two steps back. That's

31:16the whole point of just asking Claude code to fix this. like I don't know what the code means specifically. So just giving it the context of the console logs and saying hey continuing to get errors just helps us kind of like at least make a bit of progress even though it doesn't seem like we are. Here's the issue. The system is designed to process multiple skew group multiple photos per skew group but it's only receiving one image. For size detection to work you need to upload all photos. Yes, we did that. Assign them. Generate the listing current workflow. Boom boom boom boom boom boom. Let's give it a try. So, we're basically selecting the photos that are going to be in this listing. And this is going to be working work. We're going to work on organizing this a bit more because it's not completely finished. But at least I'm picking all

32:01the photos that are relevant to the listing. I'm assigning a skew to assign the skew. Move on to the generate listings should be there. We have three of them. Those are the failed ones. Um, so let's generate and see what happens. Okay, so it failed on that one. See if it fails on all. AI analysis failed. Tag something. Okay, so back to console logs. Boom. Copy. Still errors. Giving it the console logs. Let's see what it says. It's supposed to work. All right. Still didn't get the size. All right. Looks okay. It just missed the size. Again, nothing's perfect, nor do I expect it to be. Just on an hour on working on it. So, it's going to take a little bit more. So, let's go console

32:46code again. you've seen it back and forth. The problem, it was just reading only one photo and that's why it just wasn't getting all the information. Even though just off of that one photo, it did read good amount of information. It just didn't have wasn't processing the photo of the tag, which was the fourth photo that we uploaded within the skew. And it just says it wasn't doing that. The key breakthrough was in source pages generate listings page. We where we fixed it. The old one was uploading just only reading the primary photo. Now it reads all photos in the skew group and it should now work. So let's give it another try here. Right. So we have it inputed and we're just continuing to test the same item. So it didn't work there. Didn't work there. Let's give it another try. Doing something different.

33:32I'm just going to try it on a different item here. And doesn't work. So console logs back. Copy paste. The cool thing though, even though it seems like we're not making any progress, is that just the difference coming from bolt new to using cloud code, I can like give it this information and we can attack a specific error and I can see it go through kind of how it analyzes that error and kind of just read along with it. And you know, even though we're still struggling to make much progress, it just seems like it's understanding what's going on and just from the pre previous context as well. So, we're like slowly making progress even though it's not visible yet, right? On the local side. All right. So, let's go back to this one. Generate console logs. Copy.

34:20Paste. Give it to Claude. 900th attempt. Let's see what happens. I feel like we're getting closer. I really do. I really do, guys. I feel like we're getting close. Redo the process again. All right. So, we are back at it. Let's see the changes. See if it works. I feel like we're getting closer. It worked. It worked. Oh my god. Oh, I'm going to cry. It worked. It worked. It worked. It worked. It worked. Oh my goodness, y'all. It's not like we had been working on this for 12 hours straight, but we have been going back and forth with Claude and it's been a little bit tedious, but it's working. So, that's cool. So, let me go back to cursor bolt cursor uh claude code within

35:05cursor. Tell it the good news. It worked, Claude. It worked. It analyzed the image. It got the brand right. It got the title right. It got the size right. Thank you. The only thing that I would that when someone uploads a photo and the tag just shows S, M, L, or extra large, I'd like it to be able to process that size character into the actual size. So, say it's a size M. Well, I want the size M to be spelled out as medium. Is there any way to fix this? All right, so let's see if it implemented those changes. All right, it did it. It did it. It did it. So cool. Amazing. Amazing. Wow, that's so awesome. So, if you like seeing AI startups come to life, make sure you

35:50guys subscribe. Comment down below on any thoughts. Tomorrow's video is going to be continuing to build out these sub agents and seeing what Claude Code could can come up with to help optimize our business plan. We'll see you in the next one. Read to us.

Ready to start building with AI?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.

Join Shipping Skool

Keep reading