It has been over twenty years(including as a student) of writing code, but recently, sometimes I started feeling awkward when I sit down to write it without any help.

You'll know why this happens, and I'm sure most of software developer will have same feeling.

Performance of LLM's coding

Yes, speed won.

Using LLMs on work are no longer optional, especially in the workplace. The shift in velocity expectations following LLM adoption is simply incompatible with manual coding(or even copy & paste from code on stackoverflow). As already proven, AI-generated code is now production-ready, and its quality + quantity is undeniable.

Still, something in my mind resists letting LLM go entirely. It’s not just nostalgia, but it’s about control. I'm feeling that I'll lose the ability to judge what's being produced if I stop reading/writing code. Ironically, the more capable and powerful these LLM models become, the more critical that human judgment will be.

Going Slow on Purpose

So I decided to run my personal projects in old way. Reduce LLM on implementation, just me and the editor. It does not meaning that I'll not use at all. I'm still using for research and debugging. It is more closer way of how I work on googling-era. But at least, I'm trying to typing myself.

Reverting to manual coding is instructive, yet frustrating experience. Reduction of productivity is significant. About a half-day work equals what my Claude could produce in 20 minute. I’m not ignoring this cost; the gap is staggering. However, I view this friction as a required investment.

What truly shocked me was that after 20 years of coding, I’m struggling. The habit of searching for answers myself and keeping entire context in my head, rather than offloading it to an AI, has faded faster than I ever imagined.

It is sobering to realize that 1.5 years of working with LLM could so easily dull the expertise I spent twenty years developing.

No doubt on working with LLM. But...

This is not a new pattern. The shift from terminal to IDE softened many hard edges of writing code. Autocomplete began filling in what we used to type from memory. Each layer made the act easier and less effortful. LLMs are the latest step in that progression, and a much larger one.

I am not resisting the direction. Productivity matters. The tools are good and getting better, and I have no doubt on using them is the right way for most work.

What I resist is losing one specific capability, becoming reading unfamiliar code and quickly approve whether it is correct or not.
I still believe that keeping judgment skill as an engineer requires keeping the habit of writing code directly.

The personal project crawls along. It makes me frustrating, but that's exactly the point I need now.