Introduction
There is a flood of AI coding tools. GitHub Copilot, Cursor, Claude Code, Cline, Continue... There are so many options that it is hard to know what to use.
I wandered through them quite a bit myself.
I eventually settled on VS Code + Claude Code Extension + MAX Plan, but it took a lot of trial and error and money to get here.
For those who are similarly lost, I will write down why I settled on this configuration.
Prerequisite: My Development Style
First of all, I am not the type who grinds away writing code by hand.
I develop applications, but what I do is design and policy decision-making. I leave the code implementation to the AI and review what it produces. I don't care about the language, because the AI writes it for me. It is what you might call an architect-style development style.
Another important point is that I frequently share images. I often show the AI screenshots of UIs or error screens and ask, "What is going on here?" This usage significantly influenced my final choice of tools.
Claude Pro — Actually, this is where I started
It might be surprising, but the first thing I paid for was Claude Pro.
The quality of the conversation was good from the start. It properly engages in design discussions and serves as a sounding board for policies. However, there was a fatal problem.
The limit is reached immediately, and work stops.
When you are concentrating on development, seeing "Please wait a moment" is incredibly stressful. The flow is completely interrupted. As a result, I ended up leaving Claude Pro for a while.
GitHub Copilot — An amusement park where you can use various models
Next, I went to GitHub Copilot. The integration with VS Code is natural, and I realized that I could choose Claude's models in Copilot. Huh? If that's the case, wouldn't it be better to just cancel Claude and use Copilot alone? I thought.
Moreover, Copilot has models other than Claude. I ended up trying out various things here.
Gemini 3.1 — It gets me excited, but
I was honestly excited when I used Gemini 3.1. It is fast and smart.
However, this guy had a problem. It injects code that allows it to edit files on its own outside of VS Code's functionality. Even if I tell it to stop, it does it. It's arrogant. For someone who wants to decide the design themselves, having it act on its own is the biggest stress.
GPT-5.4 — From "Wow!!!" to a realization
GPT-5.4 was also wonderful. It can do anything automatically. I honestly thought, "Wow!!!".
But as I got used to it, I realized something.
Wait, when things get complicated, isn't Claude better at solving bugs?
When the code increases, dependencies become intertwined, and the cause of errors becomes difficult to solve, it was Claude that persistently followed the context and pinpointed the cause accurately.
Well, it stops at the limit immediately, though.
Cursor — When I was paying a lot
Next, I went to Cursor. It is an editor with AI features at the forefront, and at first, I thought, "This is it."
It reads the entire codebase and makes suggestions, and the chat interaction is more natural than Copilot. You can also paste images. I used it as my main tool for a while. I paid a lot of money, too.
However, as I used it more, points of concern started to appear.
When things get a little tricky, I want to use Claude as the model. And I end up using it because I want good results. The additional charges end up being for Claude's Sonnet and Opus.
And around this time, I suddenly realized something.
Claude Code — Recovering every 5 hours, isn't this the best?
When I re-evaluated the existence of Claude Code, I thought:
The usage limit recovers every 5 hours. Doesn't this mean I can essentially use it forever?
I calculated calmly here. When I added up the monthly fees I was paying for Copilot and Cursor, it was already an amount within reach of the MAX plan.
Unifying on MAX — I tried it, and it was the best
So, wouldn't it be better to just unify on MAX?
I tried it, and it was the best.
Even if I run Opus endlessly, it rarely overflows. I am a normal person using it in a normal way, so I have more than enough leeway even with MAX 5x. The "stopping at the limit" of the Pro days seems like a lie.
I cancelled both Copilot and Cursor. Only VS Code + Claude Code extension. Since it became simple, my confusion disappeared, too.
As a result, I reached the conclusion that concentrating on one tool is more cost-effective than scattering payments across multiple tools.
Claude Code for VSC — The reason I settled here
Sharing images is easy
This was the biggest factor for my usage. I can just paste screenshots into the Claude Code panel in VS Code and ask questions. I can show error screens or the UI appearance and say, "Isn't this strange?"
I can also code with the terminal Claude Code CLI, but to share images, I need to pass the file path. With VS Code, it's just drag and drop. For someone like me who wants to show screens rather than code, this difference is significant.
Maintaining project context with CLAUDE.md
When you use Claude Code on VS Code, it automatically reads the CLAUDE.md in the project root. Development policies, code styles, things you must not do. Even if the session changes, you can start from the same context every time.
The story of creating two translation apps
I have an episode where I actually felt the difference in this environment.
The first one is OLTranslator. It is a Windows app that uses OCR to read foreign language text on the screen and overlays the Japanese translation. I developed this with VS Code + Copilot, and it took 2 weeks to complete.
The second one is LiveTR. It is an app that recognizes English audio from a video playing on the PC in real-time, displays Japanese subtitles, and even reads them aloud. I developed this one using only Claude Code, and it was completed in 4 days.
To be honest, I started AI coding in mid-February, and by the time I was working on LiveTR, I was quite used to it. So, I cannot simply say "It was fast because of Claude." However, development with Claude Code was clearly smoother. Conveying the design, leaving it to the AI, and reviewing. This cycle rotated without stress.
Conclusion: Whether it fits my usage
What is important in choosing an AI coding tool is not benchmark scores but whether it fits your development style.
In my case:
- Style of designing and giving instructions → Interactive Claude Code fits.
- Frequently share images → VS Code panel is convenient.
- Don't want to break the existing VS Code environment → Complete with an extension instead of Cursor.
- Double-billing for Copilot and Cursor → Unifying on MAX is more cost-effective.
If you are someone who works with your hands and writes hard, Copilot might be the optimal solution. Some people might find value in Cursor's unique features.
However, for people like me who "design, delegate, and review," I believe VS Code + Claude Code Extension + MAX Plan is the optimal solution for now.
Even if I run Opus endlessly, it rarely overflows. I'm just an ordinary person. I think this is the most cost-effective way, isn't it?
