How to Debug a Bolt App with GPT-5 Using Console Logs

By Beau Johnson··9 min read

Quick answer

I paste my console logs and Netlify logs into GPT-5, ask it for a handoff I can give my developer, and paste that into Bolt's discuss mode. It traced my generic eBay listings to a JSON parsing problem, and after two rounds the app started generating real titles.

In this video I take a bug that had been stuck in my app for 8 to 12 hours and hand it to GPT-5 to see if it can figure out what's wrong. I recorded this in August 2025, on day 10 of building EasyFlip.ai in Bolt. Tools like these change fast, so check the current docs before you copy my setup. If you're a beginner vibe coding an app and you keep hitting errors you can't read, this is the workflow I'd try first.

What I cover in this video

  • 0:45 A quick tour of EasyFlip.ai and the generate listings tab that keeps failing
  • 2:16 Copying the console logs and describing the problem to GPT-5
  • 3:01 Hitting the free plan limit and switching to the paid model
  • 3:47 GPT-5's diagnosis: the problem is how my app parses the LLM response
  • 5:20 Asking for a developer handoff and pasting it into Bolt
  • 7:35 The first fix fails, so I add the Netlify logs for round two
  • 9:07 Bolt's project size error and why I'm planning to move to Claude Code or GPT-5
  • 12:11 The listing finally generates with a real title

The bug: generic listings instead of real titles

EasyFlip.ai is the app I've been building in Bolt. The whole premise is that you upload photos of an item and it automatically generates an eBay listing. I have the eBay API, the Google Vision API, and the GPT-4 Vision API connected.

The workflow had been getting better, but the generate listings section was stuck. I'd upload photos, assign a SKU, hit generate, and get back this:

Quality clothing item ready for review

What I wanted was for it to read the tag and put the brand name, the size, and keywords into the title. Bolt says it has integrated GPT-5, but I hadn't seen it, and when I asked Bolt whether it was using GPT-5 it didn't give me an answer. So I went to ChatGPT directly.

Round one: paste the console logs into GPT-5

My first try was on the free plan with a photo attached. I hit the free plan limit right away because the chat had images in it. If you're working on a big project, the free credits run out quickly. I moved to the paid model and asked the same thing:

Having a hard time with my app, EasyFlip.ai. I have Google Vision API and GPT-4 Vision API connected, but when I hit generate listings, it returns back a very generic response. It's supposed to analyze the photo and the tag and populate and create titles and keywords. It's just failing to do so. Here's the console log.

Within about 30 seconds it came back with a breakdown that was easy to read. It said the core issue wasn't the auth or the uploads. It was the LLM response parsing. The model was returning a JSON looking payload wrapped in a code block string, and somewhere my code was treating that string like an object. That's why the available keys in my logs were just character indexes and the title was missing, even though the JSON had a title in it.

I don't write the code myself, so my next message was:

Can you give me something I can give my developer to make these fixes?

My developer is Bolt. I copied the whole response and pasted it into Bolt using the discuss feature. Bolt outlined what the changes would do:

  • OCR first: Google Vision extracts the text from tags and labels
  • Enhanced vision analysis: OpenAI uses both the image and the OCR text for better brand and size detection
  • Structured output: JSON schema validation for a consistent data format
  • Better error handling: fallbacks when the APIs are unavailable

Then I tested it and got the same generic result.

Round two: add the Netlify logs

I went back to GPT-5 and told it the same thing was still happening. This time I gave it the console logs plus the logs from Netlify. Again it took less than 30 seconds.

The answer was more specific. The Netlify function needed to force JSON and return parsed JSON, not a fenced string. The Netlify log preview showed the response starting with JSON fences, and I was likely returning the raw string from OpenAI straight to the client.

I pasted that into Bolt's discuss feature too. I like using discuss before implementing because Bolt comes up with an actual plan first, and that keeps it from wasting tokens. Then I told it to implement the plan.

The next test was messy. My item didn't show up on the generate listings page and I had to sign back in. But then it worked. The app returned a title for a Wall Street Bull graphic hoodie in black, with a good description and keywords. It still said size unknown when it should have picked up medium, and it missed the condition. That's still real progress after hours of generic output.

I screenshotted the result, gave it back to GPT-5, and asked if we could integrate eBay item specifics into each listing and have GPT-4 Vision fill them out. It wrote another handoff to fill in item specifics automatically using eBay's metadata and to improve size extraction so size unknown happens less. The end of the recording cut out, so you don't see that one get tested.

Why I'm planning to move off Bolt

While Bolt was implementing the fix, it showed me this message:

Project size seems quite large

I get that error at least once a day now. Bolt got me a long way. Ten days before this video the app didn't exist. But the day before recording was nothing except fixing bugs and errors, and I think the project has gotten too big for it.

The other reason is the manual work. Bolt has me going into Supabase to make changes in the SQL editor, add policies, and change buckets. I do want to learn how it all works, but I can build sub agents in Claude Code and have them handle the environment variable files and the Netlify changes for me. That cuts down the time it takes to build the app.

As for GPT-5 as a debugging helper, I liked it. The responses were snappy and easy to understand. I'd been using Claude as an assistant inside a project (not Claude Code), and the GPT-5 answers felt a bit clearer to me, though that might just be me.

Key takeaways

  • If you're a beginner and stuck on a bug, copy your console logs and your Netlify logs, paste them into GPT-5, and tell it what's going on. It will break the problem down for you.
  • Ask for something you can give your developer, then paste that into your builder. In Bolt, use discuss first so it makes a plan before spending tokens.
  • My generic listings came from a parsing problem. The model's JSON was wrapped in code fences and my app was treating the string like an object.
  • The first fix didn't work. Adding more logs for the second round is what got me a real title.
  • If your Bolt project throws size errors daily and you spend whole days on bugs, it may be time to move to a different tool.

FAQ

Can GPT-5 help debug an app built in Bolt?

Yes, in my case it did. I pasted my console logs and Netlify logs into GPT-5, it explained the problem in under 30 seconds, and I gave its fix to Bolt to implement. It took two rounds before the listing generated with a real title.

Why was my GPT-4 Vision app returning generic responses?

In my app, the model was returning JSON wrapped in a code block string and my code was treating that string like an object, so the title field looked like it was missing. GPT-5 told me to have the Netlify function force JSON and return parsed JSON instead of the raw fenced string.

What do I do when Bolt says the project size is too large?

I was getting that message at least once a day by day 10 of my project. I kept working through it for the video, but it's one of the main reasons I decided to move the project to Claude Code or GPT-5.

Can I debug with GPT-5 on the free ChatGPT plan?

You can start, but I hit the free plan limit almost right away because my chat had images attached. For a bigger project, the free credits run out quickly, so I switched to the paid model.

Transcript

Read the full transcript

0:00Build me an expense tracker with charts and categories. Wow, in just a few minutes, we have a fully working expense tracker. In this video, we're going to be going over some of the bugs and errors we've been facing in our project with using chat GPT5, the new release version. We're going to kind of work through a lot of these problems we've been facing and see how it stacks up against cloud code. Let's get into it. Let's get some broken code. Shouldn't be hard to find. All right, here's our project that we've been working on so far in Bolt and it's been great. It's been great at setting the foundation, but we do have so much going on here that we are really starting to push it to its limits. According to Bolt, they have integrated GPT5, but I haven't seen it. I've even asked Bolt if it's using GBD5. It doesn't give me an answer. So,

0:45um, let's see if the website is up and running and going well. Let's see if something will break. So, let's go to easyflip.ai. Sign in. Let's continue with Google. All right. So, the thing we've been having a hard time with is the generate listings tab. So, let's see if I have any photos I can upload. And the whole premise of the app is to be able to upload photos and have it automatic generate listings on eBay. So, let's see how well it does. So far, the workflow's been getting much better. But where we've been having a hard time is the generate listings

1:30um section where it's just having a hard time identifying and putting in the actual brand name, the size, and keywords. So, you can see here's where we assign the skew. We're going to assign it as skew number one. Select all and then assign skew. Should go to the generate listings page. And then here we can generate and see what happens. Ideally it will generate. We have the eBay API. We have Google Vision API and then we also have the GPT vision API. Let's see what it does. So it just says quality clothing item ready for review. Okay. So, let's

2:16just take the uh console code and give it to uh GPT5 and just tell it what we want it to do. So, I'm having a hard time with my app. It's generating really generic responses. I'm using GPT4 Vision API. I wanted to be able to read the tag, put the brand name and the size in the title, and it's failing to do so. Here's the console logs along with a photo. There we go. And then we'll attach a

3:01screenshot of what's going on. Well, maybe not. All right, let's just get it and see what it does. See what it says and see how it can help. Uh maybe it's something I need to give to Bolt to fix this. So it says you've hit the free plan limit. You need 4 to continue the chat because it has images and my limit resets at 515. So obviously if you're doing anything crazy um as far as a big project, it's going to run out of uh free credits pretty quickly. So let's see what we can do here. Okay, let's just hop into the paid model and ask it the same thing. Having a hard time with my app, EasyFlip.AI.

3:47I have Google Vision API and GPT4 Vision API connected, but when I hit generate listings, it returns back a very generic response. It's supposed to analyze the photo and the tag and populate and create titles and keywords. It's just failing to do so. Here's the console log. Give that there and just see what happens. How long it takes to figure it out. And let's see. Okay, within about 30 seconds, it processed it and broke it down for easy to read. Um, it says the core issue isn't the O or the uploads. It's your LLM response parsing. you're getting a JSON looking payload wrapped in a code

4:34block string. Then somewhere you're treating that string like an object. That's why your available keys 945 are just character and missing title even though the JSON has title. What's going wrong? Model returns a string. You call object keys or similar on it. You got character and not fields. Um, okay. Quick fix today. So that looks great. Okay, cool, cool, cool, cool, cool. So, can you give me something I can give my developer to make these fixes? All right, it's still writing the code

5:20out to give to my developer, which I'm just going to give back to Bolt and see if this will fix the problem. Okay, just a quick summary. If your deb drops this in, the missing title, character index keys go away and you'll be getting a reliable title and keywords populated. So, let me just copy paste all of this into Bolt and just see what happens. Let's see. All right. So, here we go. Let's copy and let's go back to Bolt. Where is it? All right. Back into Bolt. discuss copy paste. See what happens. All right, I just gave it the code

6:05to bolt. All right, now it's just outlining what the changes do and what they indicate. Um, so let's see how this improves the AI analysis. OCR first. Google Vision extracts the text from the tags and labels. Enhanced vision analysis. Open AAI uses both the image and OCR text for better brand size detection. Structured output JSON schema validation ensures consistent data format. Better error handling. Graceful fallbacks when APRs are unavailable. So the combination of OCR vision analysis should provide much more accurate brand size and item specific details instead of generic responses. Let's see if this took effect and

6:50worked. Let's check. This has been a issue I've been working on for the last 8 to 12 hours. Um, so let's see. Select all. This can be skew one. Assign skew takes me to the generate listings page. Confirmation. Boom. Generate all items button is Oh, let's see. Hold on. Supposed to take me to the generate listings. Okay, let's just see if this worked. Generate all items. Quality clothing item. Edit. All right. Well, we obviously don't want

7:35that. We want it to populate with the title, the price, item specifics, and didn't do any of that. So, let's see. Console log. Still the same thing is happening. It's returning a generic response. Here is the console logs as well as the logs from Netlefi. All right, a great response. This took about less than 30 seconds to analyze all that. And if you're building something like myself and you're a beginner and you're coming across a lot of errors and bugs and maybe your project's getting too big, just go into GPT5 and just tell it what's going on. Copy the console logs. Copy the logs from Netlefi if you're using it and just copy paste. It's going

8:21to break it down for you. It says Netlefi function force JSON and return parse JSON, not a fenced string. The problem right now is Netlefi log previews shows the response starts with JSON fences. Also, you likely return the raw string from OpenAI straight to the client. So, let's see there. Okay. Okay. Uh, if you paste your current Netlify function minus keys in the client extractor where you log title field variations, I'll mark the exact lines to delete and change. Okay. All right. I grabbed all that response from GPD5 to put into Bolt. And let's

9:07see what it says. Bolt came up with the plan. I like using the discuss feature before just getting into the implementation. That way it can come up with an actual plan that won't waste tokens. So let's implement the plan. Project size seems quite large. I get this error at least once a day. And this is one of the reasons why we're going to be probably moving to GPT5 or cloud code just to be able to build out this app. There's many things I want to add to it. Bolt.new new got us to a point, but I feel like there is a lot more errors and bugs that I need to fix constantly. So, regardless for the presentation, let's just have it go

9:53through it and see what it does. And one of the main reasons why I'm making that switch is just because with using Bolt, it has me go into Superbase, make a bunch of changes in the SQL editor, and add policies and change buckets. And yeah, I'm learning how to use Superbase. It's just going to make it so much easier when I'm making changes to environment variable files and making changes in Netlefi, when I can build sub agents out in Claude and have it take care of it for me. Not saying that I don't want to learn how it all works. It's just going to cut down on the time it takes to build this project out. We've built this project that you've seen, we've been testing on easy flip.ai. Today's day 10. Literally 10 days ago,

10:39this did not even exist. So, we have made crazy progress thanks to Bolt. I just think that in the project now, we're getting to a point that we need to make the change now more than later. Um, because I am starting yesterday was just problem solving bugs and errors because I just think that the file or the the project has gotten too big. So, if you've experienced that, maybe it's time to make the change. Okay, just a couple minutes later and it says that it applied all the changes and let's just give it a second to deploy and then we will test it out. All right, so back in the app. Let's see if it recognizes this Steo jacket. It's obviously plain and clear that it's the brand STEO. It's a size 2XL

11:24and the color is pretty obvious. So, let's assign the skew on that one. One. See if it takes us to the generate listings page. It's supposed to. Um, let's see. All right. Generate listings. It's not in there. Let's delete this. And this is what I'm talking about. Things just happening just like this. Uh, let's just sign back in. All right. Let's see what it does. Ideally, it works, but that never happens, right?

12:11Oh my, it worked. Amazing. Revolutionary. So good. Wall Street Bull graphic hoodie, black size unknown. Even though it should pick up medium. Let's see what it says in here. So, condition it didn't get pretty cool on the keywords description. Just missed the size and the brand. I would like it to just be Wall Street Bull, just a basic hoodie. But very cool. That's That's progress, right? That's progress. So, awesome. Um, let me screenshot this and paste it into GPT 5 and just said we made progress.

12:58Um but it failed to recognize the size. Okay. So, next project of mine and um just kind of continuing what I'm working on is um I just attached a photo to uh GPT and it just says it failed to read the size of the item, but for the most part it worked well. Is there any way we can integrate eBay item specifics into each uh listing and have it filled out by GPT4 Vision? I'm already connected to the eBay API, so let's let it work and do its thing. I just really think I'm getting to the point in the project that I need to um switch from Bolt um either to Chat GBT or Claude. Um it's just getting to that point. So,

13:44um been liking using it so far. I think it's pretty snappy in its responses and very easy to understand. And I've been using Claude, not Claude Code, but Claude is just an assistant um within a project that I have set up. And it seems like the responses on GBT are a bit more clear, but um I don't know if that's just me, but um here is the fix. Like I asked it and it says fill in eBay item specifics automatically using eBay's meta data and boost size extraction. So size unknown happens way less. Here's a clean handoff for your dev that can be implemented. So I'm just

14:29going to copy paste all of this once it's done. And I'm just going to give it to Bolt. And the very end of this video had cut out, but we are still working with GPT5 to kind of problem solve some of these errors. So, if you like troubleshooting and you like seeing AI startups come to life, make sure you guys subscribe because we're releasing videos every day on this channel documenting our journey from zero to 10K MRR. We'd love to have you. We'll see you in the next one.

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