I first learned programming in the summer of 2018, so I spent about seven years writing code by hand.
During those seven years, what I loved most was that feeling of being completely immersed in coding. It was the moment when I was building something from scratch and bringing an idea to life.
But around the time I left my job last year, I started using Claude Code for the first time, and honestly, it was pretty shocking. Once you give an LLM a terminal environment and access to a bunch of tools, it becomes a completely different beast from the usual chat-based LLMs.
It genuinely felt like this thing had been trapped inside a chat UI the whole time, unable to show even a tenth of what it could really do. Then, the moment it got full access to a computer, it just took off.
So from that point on, I was basically glued to Claude Code / Open Code. And as of March 29, 2026, the day I'm writing this, I probably haven't written more than 50 lines of code with my own hands.
I used to think I loved coding, but it turns out what I really love is building things. Coding was just a tool for getting there.
At first, it felt strange to hand off the entire coding process to an agent and build things purely through prompting and management. I kept wondering, can I really say I made this? And compared to projects I had coded myself, I definitely felt less attached to them 🥲
But now, all those interests I once had in writing “Clean Code” and building “Clean Architecture,” and more generally all those things related to writing code, have gradually shifted into a different question:
How can I use agents better?
Skills, subagents, agent teams, harnesses, sandboxes.
That's what matters to me now.
I don't really care how “crappy” the code written by agents looks. What even is crappy code in the first place? If code that looks bad to humans turns out to be what agents perform best with, should we still call it crappy?
I wouldn't say this is a 100% accurate representation of everything I've been thinking lately, but it's clearly the direction my thoughts are heading.
A lot of people don't seem to care much about skills because they're “just markdown files.” Although they're not really that simple anymore.
That's what I thought at first too. Aren't they just a written list of what to do?
But then I realized that code is also basically a written list of what to do. The only difference is that code is written in programming languages, while skills are written in natural language.
They're basically the same thing.
A well-written skill creates a completely different experience.
Let me bring up Garry Tan's GStack for a moment:
https://github.com/garrytan/gstack
Garry Tan is the CEO of Y Combinator. He has advised thousands of top-tier startups in the Bay Area, and he knows what matters and what doesn't in a startup way more than I do, and honestly, more than almost anyone on Earth.
And his GStack is basically a near-clone of how he operates at YC.
Can I get a similar level of feedback just by talking to vanilla Claude Code or Codex? Probably not.
Can I write a skill that operates like GStack entirely on my own? Probably not.
The key is the experience he has built throughout his career and during his time at YC. LLMs and I might get a rough sense of how he thinks by reading his work and writing, but we still can't know exactly how he really thinks and operates.
And that's why I believe skills are basically the next form of computer programs.
A computer can be just a fancier calculator, or it can run something like ChatGPT, depending on what programs it's running.
Similarly, an agent can be just another Simsimi, a Korean chatbot originally built in 2002 mostly for entertainment, or it can become something like a clone of Garry Tan, depending on what skills it's running.
