이 기사는 아직 이 언어로 제공되지 않습니다. 영어 버전을 표시하고 있습니다.

Thinking in Vibe: A New Workflow with Claude Code

From Copilot to Cursor Composer to Claude Code—exploring the evolution of AI programming

· Yiming
AI ProgrammingWorkflowClaude CodeMCP

By the end of this, you’ll know how to set up Claude Code, connect it to Chinese LLM APIs, configure MCP servers, and use an Ebook MCP project to understand and modify code efficiently.

In the age of AI, coding isn’t just about writing code anymore.
It’s about humans and AI writing and thinking together.

You need to learn how to collaborate with AI:
How to divide work, how to guide, how to correct course, how to sync up.

You iterate your understanding through practice,
finding balance between efficiency and control.

I call this state:
Thinking in Vibe.

Part 1: My Vibe Coding Starting Point—The Copilot Era (2023)

I started using GitHub Copilot in 2023.

Back then, I used it in two main ways:

  • Smarter code completion
  • The built-in chat window in the IDE for explaining code and exploring ideas

Copilot at that stage felt like:

A better autocomplete + a technical Q&A assistant

It definitely boosted productivity, but fundamentally it was still:

  • Human in the driver’s seat
  • AI as the assistant
  • Changes usually limited to a single file

This was the starting point of Vibe Coding,
but it wasn’t really changing the programming paradigm yet.

Part 2: The Real Game Changer—Cursor Composer (August 2024)

What made me realize “AI programming has entered a new phase” was Cursor’s Composer mode.

The key changes with Composer:

  • Can modify across multiple files
  • Can do structural refactoring
  • Acts more like an execution agent

That’s when it hit me:

AI programming has entered the Agentic Era.

Part 3: The 2048 Demo—A Complete Vibe Coding Experiment

In September 2024, I built a 2048 game demo and posted it on Xiaohongshu.
It got over 1,000 likes.

The whole process was almost 100% Vibe Coding:

  • Used screenshots to help AI understand the UI
  • Implemented interaction logic
  • Added sound effects
  • Added animations
  • Tweaked visual details

A lot of people found it “mind-blowing,“
but some questioned it:

“2048 is simple. AI is just piecing together old code.”

But for me,
the value wasn’t in building the game itself.
It was in discovering:

The development process and paradigm are undergoing a revolutionary change.

It’s shifting from “you write code”
to you and AI continuously interacting to achieve a goal together.

Problems I Noticed Back Then

  1. AI output was unstable
  2. It would often stop before finishing
  3. I had to keep asking it to self-check

This made me realize:

AI is powerful, but it needs structured guidance.

Part 4: Chat with PDF—From Demo to “Serious Project”

A few days later, I built a full Chat with PDF app.

This time, my approach was completely different.

Right from the start, I wrote two documents:

1️⃣ Functional Spec (Business Requirements)

2️⃣ Technical Spec (Technical Approach)

I wrote these docs together with ChatGPT,
not in the AI coding tools of that time.

The reason was simple:

AI coding tools back then didn’t support systematic planning modes yet.

So I chose:

  • Use ChatGPT for planning
  • Use Cursor for coding

After the docs were ready, I had Cursor read the specs and start implementing.

I also created prompt templates for common instructions, constantly guiding it to:

  • Check
  • Fix
  • Iterate
  • Complete

In the end, the entire feature was fully implemented.

My Key Takeaway

Vibe Coding isn’t about winging it.
It’s about:
Structured input + clear goals + human course correction.

Another crucial reason:

I had very clear understanding of this project’s business requirements and implementation details,
including:

  • Which libraries to use
  • How to structure the architecture
  • How to break down features

The specs were relatively clear,
which was one of the key reasons AI could deliver good results.

Part 5: Kiro—Exploring the Dual Mode of Vibe + Spec

Later, my friend Jesse recommended Kiro to me.

Compared to Cursor, Kiro introduced Spec-driven development:

This created a dual mode of Vibe + Spec.

It emphasizes:

  • Spec-driven development
  • Planning mode
  • Task breakdown
  • Implementation

It not only supports brainstorming,
but can also break ideas down into executable tasks.

YOLO Mode: The Experimental Form of AI Driving

Kiro also offers YOLO mode (You Only Live Once):

  • Higher autonomy
  • Less human intervention
  • Longer automated execution

My take:

It’s more experimental right now.
For complex scenarios, it’s not production-ready yet.

But it represents an ultimate vision:

Like the ultimate goal of autonomous driving—
truly fully automated AI programming.

We’re still in the exploration phase,
but this will definitely become mainstream in the future.

Part 6: Why I’m Paying Attention to Claude Code Again

Claude Code actually launched mid-last year,
but I didn’t take it seriously back then.

The main reason was practical:

It wasn’t friendly to Chinese users. Accounts and APIs kept having issues. My account got banned multiple times.

But I’ve always been interested because it represents a different philosophy:

CLI-First

I often use OpenAI Codex on my phone for code reviews now:

  • On the subway
  • In the park
  • During fragmented time

Then I execute the changes when I’m back at my computer.

This makes great use of my non-office hours.

Codex can do this
precisely because of its CLI form.

The core value of CLI is:

High flexibility + freedom of context.

Part 7: MCP—Letting AI Understand Your Context in a Standardized Way

MCP (Model Context Protocol) isn’t just:

“Let AI read your knowledge base”

It provides:

A standardized way for AI to get context, understand tools, and call capabilities.

It enables AI to:

  • Better understand your data
  • Better use your tools
  • More reliably integrate with your systems

My MCP Practice

Early 2025, I built my own MCP server for:

  • Reading ebooks
  • Parsing documents
  • Summarizing content
  • Supporting Q&A
  • Assisting with creation

I didn’t promote it at all,
but it naturally gained:

  • Hundreds of stars

I use it myself when reading,
and it genuinely helps with reading efficiency.

Now, I can say to Claude Code:

claude "Help me find if there are any Musk-related ebooks in my downloads folder. If yes, list the directories and give me three questions about Tesla's autonomous driving to test my understanding."

Claude isn’t just “answering questions” anymore.
It’s building a learning scenario.

Part 8: Why I Choose to “Mix and Match”

I stopped obsessing over:

Which tool is best?

Instead:

Different tools fit different scenarios.

  • Cursor: IDE environment + rapid iteration (Cursor for Chinese users only supports Auto models)
  • Kiro: Spec + Planning (stronger Claude models)
  • Claude Code: CLI + automation

The core isn’t the tools—it’s workflow design.

Part 9: This Is What I Understand About Vibe Coding

Vibe Coding isn’t:

❌ Just talking
❌ Not writing requirements
❌ Handing everything to AI

It’s:

✅ Expressing intent with language
✅ Guiding execution with specs
✅ Amplifying efficiency with AI
✅ Humans responsible for direction

You shift from “the person who writes code”
to the person who designs collaboration methods.

The recently popular Claude Code Skills
further strengthens AI’s execution capabilities in this direction,
which is one reason I highly recommend diving deep into Claude Code.

Part 10: You Only Really Master Tools by Getting Your Hands Dirty

Next, I’ll walk you through:

  • Installing Claude Code
  • Configuring Chinese LLM APIs
  • Getting into my Ebook MCP project
  • Understanding the code structure
  • Modifying features
  • Starting the MCP server

Because I firmly believe:

You only truly understand a tool’s characteristics, boundaries, and strengths
by getting your hands dirty.

The deeper you understand,
the better you know which tool to use in which scenario
to get maximum benefit.

Part 11: Claude Code Installation and Configuration

# Detailed installation docs: https://code.claude.com/docs/en/quickstart
curl -fsSL https://claude.ai/install.sh | bash

export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_AUTH_TOKEN={$YOUR_DEEPSEEK_API_KEY}
export API_TIMEOUT_MS=600000
export ANTHROPIC_MODEL=deepseek-chat
export ANTHROPIC_SMALL_FAST_MODEL=deepseek-chat
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1

claude

Part 12: Getting into the Ebook MCP Project and Understanding the Code

cd ~/Users/onebird/github (go to your code installation directory)
git clone https://github.com/onebirdrocks/ebook-mcp.git
cd ebook-mcp
claude "Explain the overall structure of this project"
claude "Tell me how to run the tests"

Part 13: Fork the Code and Make Changes

fork the code and create a branch

claude "Please help me check the latest dependencies and update package.json. After refactoring, run unit tests to make sure they all pass"

Part 14: Starting the MCP Server

cd {$installration_path}
uv pip install -e .
npm install
uv run {$installration_path}/src/ebook_mcp/main.py

Part 15: Starting Claude Code and Using Ebook MCP Capabilities

claude mcp add ebook uv run /Users/onebird/github/ebook-mcp/src/ebook_mcp/main.py 

claude "Help me find Musk's books in my downloads folder and list the relevant directories"
claude "For the xAI chapter, give me a few questions to test my understanding of that chapter's content."

I kept the installation and usage process brief, mainly because I want you to explore hands-on. If you’d prefer a complete walkthrough, I can record a video taking you through it step by step.

Closing

In the age of AI,
you’re not just writing code.
You’re designing how humans and intelligence collaborate.

This is Thinking in Vibe.

If you’re also looking for:

  • More efficient ways to learn
  • More natural creative rhythms
  • Easier programming experiences

Please like, save, and share. Your support gives me more motivation to keep going.