Thoughts on Claude Code

Background

As a full stack developer who loves to code complex logical things by hand, I have always been opposed to AI Agents writing code for me. Partially because I love writing code, but mostly because I never really trusted them.

I remember the day when ChatGPT 3 came out. I was staying late at the office with just another Senior, and I randomly saw ChatGPT on his screen. He was chatting with it, saying how interesting it was and joking when it made stupid mistakes. "Another chat bot?" I said. I mean we've had those since I was a little kid. Just some weirdly programmed chats that could only respond in a very basic manner with almost pre-baked ideas.

When I came closer to observe, I noticed... the thing on the other side of the screen was actually intelligent. At least seemed to be intelligent? Immediately I went to my computer and typed ChatGPT into Google. It was during that time when Google didn't suck and the thing I was actually looking for appeared as the first hit.

The era of ChatGPT

As ChatGPT became smarter over the years, ChatGPT 4 got released. I still remember asking it things and how it helped me to type things in Typescript that I could never type before. Hidden APIs, secret knowledge. I was blown away.

But as time passed this became a routine thing, it didn't spark amusement anymore, and every mistake and hallucination actually made the experience more and more frustrating. "Well.." - I said - "I mean this is still better than nothing".

People pasted their code in ChatGPT, some tried to actually use it in a way that would generate code for the project, but lack of codebase context and bad prompts just led to more and more chaos. Carbonara everywhere, pain and suffering. I didn't want to write any code with ChatGPT at all. I said no every single time the "Copy" button appeared next to the code that was generated with ChatGPT. It was great to bounce ideas off, but never to actually use code in production.

During that time I reached an all time high in my typing speed - 195 words per minute.

The era of Cursor

Then Cursor came out. People were impressed and I was impressed too. I saw how more and more of my co-workers vibe coded, but I tried it and I still wasn't impressed.

I was actually hating the experience. Both because I saw that it wasn't smart enough, but also because I saw how some of my co-workers became too dependent on AI-Assisted coding tools. Not a great place to be in.

The era of Agents

Then I tried Codex. At the time it felt like magic. It was mid 2025 and I was just looking for some AI tool to go through 200 files and perform a very simple operation. I tried Cursor, but it failed. I wrote a script to perform the work manually, but it wasn't enough in most cases - since some files were a bit different and needed slightly different approaches.

Then I tried Codex - I was blown away how good and smart it was. It was pretty slow, but it did the job. But now my problem was that I wasn't good enough to explicitly tell it what to do. Like I could explain in detail what I needed, but it always found some edge case where I didn't explain enough and did a completely weird thing.

After a while I stopped using Codex and went back to typing code by hand. Not even using Github Copilot. I thought to myself - "We are really close, if this could just be smarter and think through the same way I do? Could it behave not like a junior developer?".

Claude Code

Claude Code. CLAUDE. CODE. ANTHROPIC CLAUDE CODE. This was Codex, but better - much better. It could now solve my setup problems, go through a bunch of files, follow the way I worked and quickly review my code. It still produced sub-optimal code on a large scale, but now when I needed to go and refactor some files it was the first tool I went to. When I had a setup issue, it was the first tool I went to, when I was wondering how multithreading in Python worked, I went to - you guessed it - Claude Code.

Today it's the tool that I don't use to vibe code, but rather to speed up those parts of my workflow that I hated the most - routine, boring, repetitive tasks, useless search for info on a tool that I will forget in a week, etc.

In the past two days I have written over 2k lines of high-quality advanced and structured code. Claude Code gave me the power to only think about how I want things to be structured. I could write multiple files by hand and ask it to continue as I'm going. It just worked. The perfect combo of my fast typing speed and a friendly assistant that will never lie in code reviews, will always find things to nitpick when asked and resolve any tooling problem that I had.

Instead of spending countless hours searching for some stupid typo I made somewhere, or some tool breaking the transpilation pipeline, I could now offload it to Claude Code. I could focus on the thing I love the most - writing scalable and performant code that powers beautiful things.

How do you know if you use Claude Code right? When I asked myself that question I said - I don't feel Claude Code. It's such a hidden part of my workflow that I don't think about it. Even though I use it all the time, Claude Code performs the routine work while I can spend time on actually writing the code we need.

Need to setup fixtures in tests? No worries. Need to find out why that one transpilation tool breaks? Claude Code got you! You forgot the structure of .circleci config? The thing you have learned like 200 times, but because there are big periods when you don't need that knowledge - Claude Code will remind you and help if something breaks.

Claude Code is just a tool, but it marks a revolution - a thing that actually helps you - not trying to replace you.

With that said, today I had probably one of the best days at work - today I truly understood that coding (even in such a messy, infrastructure-wise, language like JavaScript) doesn't have to be painful. Focus on the good parts, focus on the important parts and leave the parts that are important but routine to Claude Code.

Peace.