Vibe Coding 101: Improvisation & Learning
Vibe coding is a bit like jazz improvisation. You're always learning, always trying new things, always experimenting. It pays off to push your boundaries and think outside the box.
In this blog I want to share some really cool workflows I've developed using Claude Code and ChatGPT's Codex and image generation tools. Working back and forth between various AI tools can have powerful results, and I want to fill you in on what I've been learning.
Different AIs Have Different Strengths
I'm by no means a coding or development expert, but I can tell you what I've experienced firsthand using Claude Code and ChatGPT's Codex tool to build various apps.
If you've read my previous blog about vibe coding, you'll know I originally started with an app called Cursor. While I liked it, I eventually moved on to Claude Code as my main tool for agentic coding. Over the last month, I've used a Claude Max subscription to build a large number of experimental codebases — some of which have gone on (or will go on) to become fully-fledged apps.
I find Claude Code really great at building out an initial prototype of an app. I've built some cool synths using Claude Code alone, and I think it's a great AI tool. However, there's one area where I feel it's lacking a bit, and I hope Anthropic can improve in this regard.
AI tools move quickly, and this blog will probably become dated quickly. That said, I want to explain why I feel ChatGPT's new 5.5 model has a bit of an edge over Claude Code right now, as of April 2026.
When Claude Code Hits a Wall
Sometimes with Claude Code, I run into a problem I just can't seem to solve. For example, I was recently working on a project called EP-1, an electric piano synth. For some reason, the sound had a fuzzy, distorted quality — kind of like an 8-bit Nintendo electric piano. Nothing wrong with that sound necessarily, but it wasn't what I was going for with this instrument.
I tried over and over to explain the issue to Claude Code, and it kept telling me it had fixed the problem. Then I'd test, and it would not be fixed. I've had this issue with Claude before. Sometimes it's amazing and fixes the problem perfectly; other times it's a struggle back and forth, where you don't see changes until suddenly Claude Code magically fixes the bug.
Enter ChatGPT's Codex and ChatGPT 5.5
Out of frustration with Claude Code, I decided to try ChatGPT's new 5.5 model via the Codex agentic coding tool. Immediately I noticed a change.
The model seemed to move faster and attack the bug at a deeper level. After 2-3 passes, Codex resolved the electric piano bug that Claude Code couldn't. As soon as I heard the fix flood my headphones with the clear, bell-like electric piano sound I'd asked for, I knew Codex was the right tool for my needs.
Don't get me wrong — Claude Code is still a great AI, and I'm sure it will improve in the future and may outpace ChatGPT again. Until recently, I felt Claude's Opus model was the best tool available, but with this new GPT 5.5 model, I'm not so sure.
Codex seems to quickly recognize problems it doesn't completely solve, and then go back and solve them without needing human intervention. Part of the tedium of vibe coding is the constant back-and-forth with the AI, trying to get it to achieve your vision or fix the bug in front of you. What I'm finding with Codex 5.5 is that it reviews its own work and then autonomously goes back and fixes what it didn't get right.
Of course, the model isn't always perfect. It still needs human intervention to complete projects, as it should. But I really enjoy how it iterates on its own to get things just right. An example: the website you're on right now. I completely redesigned it today using ChatGPT 5.5 and Codex, and I find it sharper now and more consistent with my vision.
Images: Another Reason GPT 5.5 Shines
Claude Code is great at creating code, but designing apps often requires imagery — whether it's a logo, marketing assets, or art for a game. Unfortunately, Claude Code can't generate images the way ChatGPT can.
Claude can generate "programmatic" images, as it calls them — images generated with code. But these user interfaces tend to appear basic or bland on screen. Some of the software synths Claude Code created for me looked hideous when I first opened them. You would not want to make music in those apps.
Part of the joy of using a nice music app or plug-in is a beautiful interface. Think about Fruity Loops, Ableton, or old Winamp skins.
To get that cool, photo-realistic look for your apps, it's very difficult to have Claude Code deliver. With ChatGPT's image generation, you can work back and forth with Codex to add images to your game or project. This is something Claude would struggle with, since it can't create images the way GPT's image generator can.
So if you had to pick only one model to pay for, I'd go with ChatGPT right now — you get agentic AI coding plus image generation. If you only used Claude Code, I feel you'd still have to pay for an OpenAI subscription, because image generation is so important to many workflows. If image generation or front-end design isn't needed in your workflow, Claude Code may be a better fit for you.
In Summary
The AI model race is neck and neck right now. In my view, the two top companies are OpenAI with ChatGPT and Anthropic with Opus. Times are changing quickly, and I'm sure new, even stronger models will be out from both companies soon.
Thanks for tuning into the blog today.