Since September 2024 I've been using AI to code every single day. That's 18 straight months on Cursor and, in the last week, I moved to Claude Code. During all this time, I learned what actually works, where AI still stumbles, and most importantly, what changed in my head as a developer.
This article is an honest account. No hype, no selling illusions.
From Cursor to Claude Code
Github Copilot was my gateway into coding with AI. I was impressed by how pressing Tab would already generate code that made sense for that file's context, I had no idea that was just the tip of the iceberg.
Some time later, about a month in, I was talking to a colleague at the company. He told me about a project he built inside Mashgin, a fully dockerized simulator of our kiosk. Blown away by the complexity, I asked more about it, and that's when he told me about Cursor and how it was a heavily improved VS Code with real AI integrated into the editor.
That was it for me. As a heavy VS Code user, I was completely hyped. Our conversation happened at the end of August 2024, and right after I started the 15 day free tier. By September I was already on the paid annual plan. For the next 18 months it was my main tool and it helped me a lot, I used it every day and I recommend you try it. The thing is, the Cursor team changed some things that left me unhappy, and when it comes to money and tools I'm pretty detached. I started testing other tools, including Claude Code, and that's when things shifted.
Vibe coding in the terminal... surreal
It's a fact that Claude Code is everywhere, I think it even surpassed Cursor in popularity and the number of people talking about it. I was pretty reluctant to subscribe, I started with the Zed integration using $5 in API credits first, then kept increasing, and now I'm on the Max plan.
The first thing that caught my attention was being able to run multiple tasks in parallel, each in its own terminal tab, all visible to me at the same time. In Cursor I didn't have that visibility (or at least I never managed to configure it that way). With the Claude Code integration in Zed, this created a whole new layer of productivity. I already liked Zed for its absurd speed, now with Claude Code alongside it, it became my main setup.
After the Cursor team launched Composer 1.5 (their new default model), which to me is very bad compared to Sonnet 4.5, it became clear that Claude Code is the best option right now. But not everything is perfect.
Where Claude Code still needs to improve
Cursor uses VS Code and its team has a strong focus on UX. There are things there that are simpler and more straightforward, for example: I select a block of text, press CMD + L and it opens a chat with that block already highlighted. Honestly, that saves time. In Claude Code, in every integration I've tested so far, I need to copy the block and paste it.
The @ for generating context inside the Claude Code terminal also needs a lot of improvement, it's quite immature compared to Cursor and even to Zed's integration.
These UX details are annoying, but they don't change the final result. What matters is the quality of what the AI delivers, and in that regard Claude Code wins.
Speaking of the features I explored in my last week.
Remote control: coding from anywhere
One of the things that surprised me the most was the remote control via iOS. I went out for coffee with my wife, and while telling her about how excited I was testing Claude Code I said: "look at this! I created a tunnel to my localhost and now I can send prompts from anywhere through the Claude app on my phone and it keeps updating directly on my machine so I can see the result."
She made a face like "interesting, I didn't quite get it, but you're excited, so great, let's go."
In my day to day, this is better than traditional "background jobs" because for whatever reason I had to leave my workstation, and I don't need to write a prompt while I was already immersed in it, wait for the AI to open a PR, follow that whole formal process and worst of all, not even have the feedback to see how the site is going.
The flow is simple: I write, it runs on my machine, I review, I repeat until it's ready. It seems like a small thing, but in practice it completely changes how I use my time, and since it takes a few minutes to execute I can have plenty of conversations with my wife while we have our usual coffee.
Or, I'm on the couch watching something? I send a prompt. I'm waiting somewhere? I send a prompt. The barrier between "I'm working" and "I'm not" got much smaller. The only thing to watch out for is not getting addicted, but that's a topic for another post.
This site was 90% built with AI
At my 9 to 5 job I have an extremely high standard, I don't just vibe code without understanding what's happening. Even if the AI writes something I don't understand, I insist on explanations until I get it, because inside Mashgin it's my responsibility to make sure the code works correctly, not the AI's.
But here I decided to lean more into the prompts and less into manual intervention. This site you're reading right now was built almost entirely with AI. I'd say 90% of the code was generated by AI, specifically by Claude Opus split between Medium and Hard thinking. The other 10% were things it genuinely couldn't solve and I had to code by hand, or adjustments that wouldn't be worth spending a prompt and providing all the context again to fix. Having an AI doing things for you doesn't mean you can be lazy.
To give a concrete example of that 10%: things like fine tuning layouts, animations that needed a human eye to feel natural, or very specific business logic where the AI simply didn't have enough context to get it right the first time. They're small things, but they make a difference in the final result and only someone who understands the project can solve them.
The most important point here is about time. What used to take days of writing and implementation now comes out in hours.
MCP Tools: the workflow of the present
Something that got stuck in our heads is that AI is just a tool for chatting, and occasionally generating images. I need you to stop believing that right now.
AI isn't just something that answers your questions. Today, in 2026 (and even before that actually), it's something that executes actions for you. And MCP (Model Context Protocol) tools open a completely new horizon.
I've been testing this workflow in the last month in my daily work at Mashgin:
Linear → Zed → Claude Code → GitLab
I connected Linear via MCP, and built a smarter MCP for GitLab using the glab CLI. From start to finish, the AI understands my task, implements it and opens the pull request. When my colleagues leave comments on the code review, I ask the AI to check and fix them. All through Zed or the Claude Code terminal.
As I said before, it's MY responsibility to make sure the code works correctly, so of course there are cases where I need manual intervention and where I'll read and understand 100% of what's happening, but in general this is how it's been.
And you know what's the best part? The people at the company support it and want more of it. It's not something I do in secret, afraid of losing my job. It's a change the entire team is embracing and that will secure it.
The question isn't whether AI will replace devs. It's whether the dev who uses AI will replace the one who doesn't.
Not everything is sunshine
I want to be honest about the limitations because there's a lot of content out there that only shows the pretty and unrealistic side.
The code isn't 100% how I would write it. Even when I explain in detail what I want, with context, with examples, and even with Skills and a well defined CLAUDE.md. The AI has its own way of solving things and it's not always the way I'd choose. Maybe there's something inside it that tries to balance cost vs effort, I don't know.
Artificial Intelligence in reality isn't that intelligent. Business rules from scratch, UX flows tailored to your audience, the best tools for the job. All of that is something AI won't decide. It goes by the probability of what people use most and what's inside its model. It's not something intelligent, it's a big follower of the crowd.
The hard truth is that AI can't run an entire business or a large project on its own. Forget it. If someone says that and you get scared because you think you'll lose your job or be out on the street in a few months, it's because you clearly didn't understand anything.
The person behind the screen (or the mobile app hehe) needs to know what they're doing, needs to understand the rules, the tools, and then figure out how to make the best use of them.
AI is a reflection of the professional typing the prompts. If you don't understand what you're asking for, the result will be proportional.
At some point you'll need to intervene manually. You'll need to read the generated code, understand what was done and fix it. If you don't have that ability, AI becomes a dangerous black box.
And yes, you'll need to learn how to code.
The hardest lesson for me: letting go of perfect code
This was the biggest lesson of these 18 months and I confess I'm still working on it.
I need to let go of perfect code. Not to the point of accepting anything, but enough to escape over engineering. And look, over engineering has haunted me since my early days in programming. That urge to abstract everything, to create the perfect architecture, to cover every edge case.
With AI, that changes. Code is no longer the most valuable asset. It's hard to accept, because it's been (and still is) hours upon hours of studying. Hours that will continue, because technical knowledge remains essential. But the relationship with code has changed.
Were my problems solved?
Being very direct, a good chunk of them yes, but others were created. The game changed. And whoever wants to get into it the right way needs to understand the combination that companies will demand from a dev:
- AI: know how to use the tools, understand the limits, integrate into the workflow
- Product vision: understand what's being built and why, not just how
- Solid code: master enough to step in when AI gets it wrong
It's not about replacing the developer. It's about the developer who knows how to use AI delivering at a level that wasn't possible before.
These 18 months showed me that AI is the most powerful tool I've ever had in my hands and I'll keep diving deeper into it. However, it's still a tool, and as such, the result depends on who's using it.