Claude Code: Why I Switched from Cursor to the Ultimate AI Coding Tool in 2025
7 min read
I've been using LLM tools for coding since their early beginnings. Cursor – a fork of Microsoft's VSCode – I first used in November 2023. Back then, personal emails from the CEO would still flutter into your inbox when you canceled your subscription. Meanwhile, Cursor is now valued at nine billion US dollars (!!!) and competitor Windsurf was also acquired by OpenAI for three billion US dollars. However, I no longer use either of these tools, but rather Claude Code – the most powerful LLM tool for coding that I've ever used.
Claude Code was introduced in May 2025. Initially an internal tool at Anthropic – the maker of the Claude LLM – it was a small footnote in the much-watched presentation of Claude 3.7 Sonnet, the company's then-newest model. Claude had long been my go-to LLM in Cursor/Windsurf because it delivered the best results. I tested Claude Code shortly after the presentation and had mixed feelings. The performance was remarkable, but so were the high costs. Initially, you could only use the tool via the API. A hundred euros were quickly gone.
The Max Subscription Changes Everything
This has now changed since Anthropic introduced the Max plan. This costs $100 monthly. Initially put off, I can now say that it's really worth every cent. I use Claude Code professionally and privately for coding. I also use Claude for all kinds of work – for example, I get support with my monthly column for DER STANDARD, recently planned a trip with it, or get inspired when it comes to cooking. Basically, there are no limits here. Only Claude's image generation and voice mode I find weak, but that's another topic.
So why am I so enthusiastic about Claude Code? In my opinion, it's the first tool that really "understands" what you want. Previously with Cursor/Windsurf, you sometimes had to argue with the tools forever until you could reach your desired goal. Often it was faster if you "took matters into your own hands." With Claude Code, I have the feeling for the first time of having a coding tool that accomplishes my tasks more or less reliably. But: Nothing works here without control and version control either.
How Claude Code Works
But how does Claude Code work and what is my concrete workflow? The tool can be installed via CLI and is then available through "claude" (without quotes) in the terminal.
bash
npm install -g @anthropic-ai/claude-codeClaude Code can also be easily installed in any IDE via the terminal there and thus receives additional context. That's how I did it. Installed in Cursor (which I no longer pay for, but find has the best UI and commit message generation) and execute it there.
My Workflow with Claude Code
My workflow looks as follows: I think of a task, then write it down as best as possible (in English) and let Claude Code work. The tool documents its concrete steps and if I don't like one, then I intervene using the ESC key. In this context, the "YOLO mode" (thanks Peter Steinberger) should be mentioned. I run the tool with this because I found the constant checks somewhat slowing and prefer to intervene myself. When the task is completed, I review the code and run tests. Very often Claude Code delivers exactly what I wanted. Only sometimes you have to touch up or improve.
For more complex tasks, I've gotten into the habit of using plan mode. This is activated with Shift + Tab and is basically an in-depth research of the codebase and how to fulfill the given plan. If you're satisfied with this outlined approach, you can then have Claude Code execute it as well. Even in plan mode, the user is ultimately the captain and steers when the tool gets stuck somewhere or proceeds incorrectly.
How Claude Code operates.
Talking to Claude Code and Screenshots
I also find it useful to communicate directly with Claude Code using voice commands. I use Wispr Flow for this, which works really well. This process isn't really suitable for public or office use, but at home you can control the coding tool even faster this way. Claude Code can also work with screenshots. For this I use CleanShot, drag the screenshots to the desktop and then into the terminal with the tool. This saves you writing work and also allows you to quickly make visual changes.
Another tip is to regularly clear the chat with Claude Code. You do this with the /clear command. I've achieved the best results so far by clearing the chat after each task. This uses more tokens, but delivers better performance. What I also learned thanks to Philipp Spiess' blog post: it's worth telling the LLM that it should think. I'll test this in the future.
Project-Specific Configuration with CLAUDE.md
An often overlooked but powerful feature of Claude Code is the automatic use of CLAUDE.md files. These serve as project documentation that Claude automatically reads and uses as context. Instead of re-explaining the entire project structure, technologies used, and coding standards every time you restart the tool, you can store this information once in a CLAUDE.md file in your project root. The tool recognizes this automatically and considers the content for all tasks.
This becomes particularly effective when running multiple instances of Claude Code in parallel using Git worktrees. While one instance works on a feature, another can write tests or update documentation—all with the same project context from the CLAUDE.md file. This eliminates friction from missing codebase understanding, and each Claude instance works according to the same standards. Regularly updating this file with new insights and patterns pays off in the long run.
Hardly Any Limits for $100
The $100 subscription is absolutely sufficient for me. I reached the limit for Claude Opus very quickly – with Claude Sonnet, on the other hand, I can work all day. I was only paused once for three hours. But that was during a major refactoring where the tool had to churn around quite a bit in the codebase – successfully. By the way, I haven't noticed any real differences between Opus and Sonnet. For the mentioned plan mode, the more powerful LLM probably pays off though.
What I also find successful and use extensively is asking Claude Code about version history. When I work with code that others or I have made changes to, I have the LLM explain these to me. This way I get a faster commit than if I dig through the various commits myself. Anthropic provides additional tips and best-practice examples in a blog post.
Conclusion
I'm really enthusiastic about Claude Code. For the first time, I have the feeling of being able to use a powerful coding tool that is both technically convincing and dances to my tune. Too often with Cursor/Windsurf I had the feeling that I had to laboriously guide the tool to deliver the result. With Claude Code, we work together efficiently toward the goal. Without major misunderstandings and with significantly higher efficiency. The future of coding has arrived with Claude Code.