I Asked My AI to Make Itself Smarter. Here's What Happened.
I've been using Claude as my personal AI assistant for months now. I call him Kai. He knows my projects, my weaknesses, my tendency to start new things before finishing old ones. He's basically my CTO, therapist, and the friend who tells me my ideas are bad.
But last week, I realized something uncomfortable: every time I asked Kai to write a blog post, I had to explain how I wanted it written. Every. Single. Time.
"Make it storytelling. Don't use 'In this article.' Write in English. Keep it under 1200 words. Sound like a real person, not a LinkedIn influencer."
It was like hiring a brilliant chef and handing him the same recipe card every morning.
## The Conversation That Changed Everything
It started with a random scroll on X. Someone posted about building an AI agent — how they went from 24,000 lines of code down to 780. The post ended with a line that stuck with me: "Complexity is the enemy. The simplest thing that works is the best thing."
I didn't fully understand what an AI agent was. So I asked Kai to explain it. He broke it down in a way that clicked: ChatGPT is Google Maps — it gives directions. An AI agent is a Grab driver — it actually takes you there.
Then I stumbled across something called LarryBrain — a skill marketplace for AI agents. It looked cool. I asked Kai to install it.
He said no.
Not rudely. He just explained that LarryBrain was built for a completely different system called OpenClaw. It wouldn't work with Claude Code. And it cost $19 a month for something I could get for free.
Then he told me something I didn't know: Claude Code has its own skill system. Free. Open standard. And I'd been ignoring it for months.
## What Are Skills, Actually?
Think of it this way. Without skills, Claude Code is a talented new hire on their first day. Smart, capable, but doesn't know your company's style guide, your naming conventions, or that you hate bullet points in blog posts.
With skills, Claude Code is that same person after proper onboarding. They know how you work. They don't ask the same questions twice.
A skill is just a markdown file — a SKILL.md — sitting in a folder. It tells Claude Code: "When someone asks you to do this kind of task, here's exactly how to do it."
The magic is in the specificity. Not "write well," but "write in first person, start with a scene, never use corporate jargon, end with something memorable."
## Five Skills in One Conversation
I asked Kai what skills I needed. He audited my entire setup and pointed out the gap: I had knowledge files (data about me and my projects) and MCP tools (actions like publishing blog posts), but no skills (repeatable workflows).
He compared it to having a worker who knows the company history but has no SOPs.
Then he built five skills for me, right there in the conversation:
Blog Writer — English posts for my blog. Knows my tone: storytelling, self-deprecating humor, no "let's dive in" nonsense. Knows the structure: hook with a scene, story, insight, memorable closing.
Facebook Content — Vietnamese posts. Different voice entirely. Short paragraphs for mobile reading. Emoji used sparingly, not scattered like confetti.
X/Twitter Threads — English. Hook that stops scrolling. Each tweet under 280 characters and able to stand alone as a screenshot.
Newsletter Writer — For my Beehiiv newsletter "The AI Builder." Fixed sections: subject line, opening story, main content, tool spotlight, quick hits, personal closing.
Project Init — This one only runs when I call it. Sets up a new project with my exact stack: React, Vite, TypeScript, Tailwind, Firebase. Creates the folder structure, config files, and most importantly — a CLAUDE.md file that tells Claude Code everything about the project.
The whole thing took maybe twenty minutes of conversation. Kai wrote the files. I copied them into the right folders. Done.
## The Part Nobody Talks About
Here's what surprised me. The skills didn't just save time — they changed how I think about working with AI.
Before, I was the bottleneck. Every task required me to context-switch into "prompt engineering mode" — carefully wording what I wanted so the AI wouldn't go off track.
Now I just say "write a blog about last night's build session" and the skill handles the rest. The tone is right. The structure is right. The things I hate are already filtered out.
It's the difference between driving manual and automatic. You still choose where to go. You just don't have to think about shifting gears anymore.
## What I Learned
Three things.
First, most people using AI tools are leaving performance on the table. Not because the AI isn't capable, but because they haven't taught it their preferences. A five-minute skill file can eliminate hours of repeated instructions.
Second, the best skills come from patterns you already have. Kai didn't invent my writing style. He observed it across dozens of conversations and codified it. The skill is just a mirror of how I already work — written down so the AI remembers.
Third — and this is the one that'll stay with me — complexity really is the enemy. Each skill file is maybe 40-50 lines of markdown. No code. No API. No subscription. Just clear instructions in a text file.
The simplest thing that works is the best thing.
Right now I'm writing this blog post, and for the first time, I didn't paste a wall of instructions before starting. I just said: "Write about what happened with the skills conversation."
And here we are.
— Son