I've been holding to not write this post because I don't want to be part of this AI hype. But lately, there are a lot of things on my head that I want to share.

If you are already familiar with AI coding tools, you can skip this post anytime. If you are skeptical & don't believe in this stuff, or you don't have time to explore it, this post might be for you. I've been at that point. So I will give you a quick summary.

This past few months, AI coding agents have been improving a lot and very quickly. I'll tell you which part is different.

If you ever heard about MCP, it's a tool which you can use to make AI models access some of the applications you have. Previously you probably just copy & paste, then prompt to build or fix some part of the code.

With MCP you can connect any resource you have with LLM models. They can now read your deployment logs, your documentation, your PRD. Not only read, they can also create all those stuff.

Imagine you have a project, start discussion with AI and make detail of product requirement, then ask them to generate a complete breakdown, then let them solve it one by one. It's just like having a complete PM assistant plus a junior engineer.

Nowadays, I heard people move away from MCP to CLI, so instead of connecting GitHub MCP, just log in to GitHub CLI on your local, run AI agents in terminal, and let them use that to interact with GitHub. It's amazing, switching from just a coding assistant into a real entity that can solve their own job (which used to be our job).

With tools like Claude Code & Codex, it makes them even smarter to manage their knowledge. So instead of just reading & writing one file or just part of your code, they can think and move like an actual engineer.

Even, you can make them ask you questions if your instruction is not clear. Or if you're too lazy to explain something, just write it in markdown files, and they will read it.

What I'm doing now, I ask them to build a feature then make documentation about that feature so if we need context for those feature, we still have it, no need to re-explain it.

Their performance depends on the knowledge you give them. I believe writing documentation is more important now, rather than writing the actual code.

Not only sharing the context about your feature, now you can also teach them your engineering skills. Instead of telling how to build the code every time you write a prompt, you can just write one markdown file about your preference, and let them know when they need to read this markdown skills.

For example, now you can teach them how to write better unit tests, write better styling, or even read a design context from Figma.

Even if you're too lazy to do this, there are a lot of people who already built a set of skills to do coding work (e.g., Superpowers). It's like upgrading from an intern engineer that just started coding into a junior that has basic coding skills. The more skills we can write, the smarter they are.

In the beginning, I thought that AI is not as smart as us. They make mistakes, they screw up our codebase. But humans also make mistakes. That's why we still have a PR review mechanism in our workflow.

I believe, it's not a choice to not use this. This stuff has become a complete part of our normal workflow.

And here is the reality, this thing will rewrite how we should build an application.