2. Vibe Coding (AI-Assisted Development Approach)

# 🧠 Vibe Coding

How I Approach AI-Assisted System Design

🧠 How I Think About AI

I don’t treat AI as something that “writes code for me.”

I treat it like an engineering component inside my system.

That means:

If AI behaves randomly, that’s not an AI problem —

that’s a system design problem.

🎯 My Approach

Instead of prompting AI repeatedly and hoping for better results,

I design the system so AI has no choice but to behave correctly.

💡 Core Principle

Define the system first
Then let AI operate inside it

🧩 How I Structure AI Systems

I break the system into clear layers so AI can understand:

🔗 Context (MCP — Model Context Protocol)

I strictly control what context goes into AI.

👉 If context is messy, AI output will be messy

📄 Rules (AGENTS.md)

I define how AI should behave:

👉 This acts as the guardrail

🧠 Reasoning (SKILL.md)

I define how AI should think:

👉 This prevents “AI doing everything badly”

🧭 Planning (Critical Layer)

I force AI to plan before execution.

Instead of:

input → output

I enforce:

input → analyze → plan → execute → validate

👉 This alone dramatically improves reliability

🤖 Agents (Separation of Concerns)

I don’t use one general AI.

I split responsibilities into focused agents:

👉 Similar to microservices, but for AI reasoning

🔧 Tools (Execution Boundary)

AI is not allowed to “decide actions”.

It can only execute predefined tools:

👉 This eliminates hallucinated operations

🎨 Design (UI Consistency)

For frontend generation, I define:

👉 Prevents inconsistent UI across AI outputs

🔄 How It Works in Practice

User Input
→ Context Builder
→ Agent Selection
→ Planning
→ Tool Execution
→ Validation
→ Structured Output

⚖️ Trade-offs I Accept

This approach is not “quick and dirty”.

It requires:

But in return, I get:

🧠 What This Is Inspired By

This approach aligns with modern AI agent systems (like Claude Code architecture), where:

📌 How I Think as an Engineer

When working with AI, I follow the same mindset I use for backend systems:

define boundaries
define responsibilities
define constraints

Then — and only then — I use AI to accelerate development.

💬 Final Thought

AI is powerful, but without structure it becomes unpredictable.

> I don’t rely on better prompts

I rely on better system design