Learning Vibe Coding for the First Time, 73 Days Until AI Writes My Code — CEOs Shim Jae-woo and Seon Ung-gyu Compare Claude Code and GitHub Copilot in Real-World Practice

In January 2024, a certain startup couldn't wait for developer recruitment. The deadline for Series A funding was approaching, and they needed to launch an MVP (Minimum Viable Product) within two months. On the day they first heard about Vibe Coding—where AI writes code based on just words or requirements—the team stood at a crossroads: Claude Code or GitHub Copilot. Regardless of which they chose, it was completely different from their previous development approach. A business team that didn't know how to code had to take AI's hand and start playing the developer role themselves.

Vibe Coding is not just an autocomplete tool. As outlined in Part 1's comprehensive guide, the core principles and key elements represent a paradigm shift where AI understands natural language requirements and designs and implements entire project structures. However, in real development practice, "which tool to choose" goes beyond simple feature comparison. The 73-day experience of Shim Jae-woo, CEO, and Seon Ung-gyu, CEO, of AX Education Group, who actually used AI coding tools in real projects, proves this.

---

Claude Code is a real-time development environment created by Anthropic. When users make requests in natural language, AI actually writes code in a browser-based development environment and shows results instantly.

The moment the AX Education Group team first experienced it was nearly shocking. They needed to design a backend database structure, and when they explained a Supabase PostgreSQL-based schema in Korean, Claude generated precise SQL queries. It included CREATE TABLE syntax, uuid defaults, and RLS (Row Level Security) policies. The most surprising part was completing everything from environment variable settings to API routes in a single breath in the Vercel deployment environment. The integration of Supabase and Vercel was the most practical use case of Vibe Coding.

* Real-time preview: Results appear in the browser immediately after writing code, shrinking feedback loops to 5-minute intervals
* Natural language iterative requests: Simple sentences like "make the button blue" or "make this form mobile responsive" enable code modification
* Supabase native support: When generating PostgreSQL queries, it automatically optimizes for relational DB features like JOINs, triggers, and functions

The most valuable discovery from 73 days of experience was that Claude Code doesn't just function as 'automatic code generation,' but operates like 'a real-time discussion partner in development'. When the team asked, "This logic seems slow, can you optimize it?", Claude suggested indexing strategies and caching mechanisms.

---

GitHub Copilot approaches with a different philosophy. It goes directly into developers' existing IDEs (Integrated Development Environments) like VS Code, JetBrains, and Vim, providing real-time autocomplete. It predicts the next line while you're typing code and completes implementations the moment you define a function.

The decisive reason AX Education Group chose Copilot was speed and context understanding. As the project became complex and dependencies between files increased, Copilot read not just the current file but the entire project structure (project name, directory structure, import statements, etc.) and suggested consistent code. When creating API routes in Next.js App Router, the moment they created "app/api/todos/route.js", Copilot automatically wrote the appropriate HTTP methods (GET, POST) and Supabase client connections for that path.

* IDE integration advantage: Developers can accelerate development speed by accepting or rejecting autocomplete suggestions without taking their hands off the keyboard
* Code review feature: You can ask about the code you've written: "Are there security vulnerabilities?" or "Is there a better performance pattern?"
* Multi-file context: Because it understands the entire project structure, it automatically references environment variables correctly when writing Vercel Functions

According to CEO Shim Jae-woo's journal entries, "Using Copilot feels like borrowing an experienced senior developer's elbow to write code." It's not thinking alone and pondering each line, but rather having the next line predicted and modifying it.

---

For 40 days of the 73-day journey, they stood at a crossroads. They needed both Claude Code's powerful automatic generation capability and Copilot's fast feedback loop, but neither was perfect.

Claude Code excelled at strategy planning. When first designing Supabase's database schema, when building Vercel's deployment pipeline, when facing 'big picture' questions like "How should we implement user authentication?", Claude provided deeper answers. Like an architect, it first drew the entire system and then filled in the details.

GitHub Copilot was superior in execution speed. Within an already-determined structure, it was more efficient when quickly creating repetitive components or writing CRUD logic in ultra-short timeframes. Especially for serverless code like Edge Functions, patterns were clear enough that autocomplete was nearly perfect.

In the end, the team's choice was this:

Early project stage (Weeks 1-2): Claude Code to complete overall architecture, DB schema, and API design

Development progress (Weeks 3-6): GitHub Copilot to mass-produce components, functions, and test code

Optimization stage (Weeks 7-10): Claude Code for performance bottleneck analysis, refactoring, and deployment automation

This hybrid approach was the core strategy that made it possible to deploy a high-quality MVP within 73 days. While maximizing Supabase's real-time DB functionality and Vercel's global CDN, code quality didn't suffer.

---

The biggest lesson the team learned over 73 days was this: AI excels at writing code, but what to build is still humanity's responsibility.

Is Claude's proposed database design always optimal? No. After receiving Claude's proposal, the team asked, "But we need real-time alerts in under 1 second—can this query structure handle it?" Claude then proposed a caching strategy utilizing pgvector.

Is GitHub Copilot's autocompleted API route truly secure? No. The team's information security officer pointed out, "This endpoint is missing RLS policy," and explicitly instructed Copilot, "Apply Row Level Security." Copilot then accurately reflected Supabase's RLS policy.

Characteristics of teams using AI coding tools correctly:

* Clear requirement definition: Rather than "Make a navigation bar," specify "The navigation bar must be responsive, with a hamburger menu on mobile and horizontal menu on tablet and above"
* Maintaining code review perspective: Don't unconditionally accept AI-generated code; ask "Isn't this a security problem?" or "Can this O(n²) performance improve to O(n)?"
* Tool selection flexibility: Don't fixate on one tool; alternate between Claude and Copilot depending on the situation

In their final retrospective, AX Education Group's CEOs Shim Jae-woo and Seon Ung-gyu recorded: "Vibe Coding doesn't replace coding ability. Rather, it makes 'judgment in deciding what to build' even more important. AI speeds up implementation, but architecture decisions, security policies, and performance goals remain humanity's responsibility."

---

Q1: Between Claude Code and GitHub Copilot, which should beginners learn first?

A: It depends on project stage. If starting from scratch with nothing, Claude Code is recommended to first understand the overall structure. Claude explains "why this structure" like a mentor. Conversely, if a project skeleton already exists, using Copilot to quickly handle repetitive work is better. In AX Education Group's experience, they transitioned from Claude initially to Copilot after becoming familiar.

Q2: When using Supabase and Vercel, does an AI coding tool really handle environment variable setup perfectly?

A: Environment variables are a very careful area. Claude Code accurately generates public keys like NEXT_PUBLIC_SUPABASE_URL but understands that service_role keys must never be exposed to clients. However, GitHub Copilot might be incomplete in this regard. Therefore, for sensitive environment variables, it's essential to explicitly instruct Claude, "This key should only be used on the server."

Q3: How much did it cost during the 73 days?

A: Claude Code costs $20/month subscription (Claude Pro) for unlimited use, and GitHub Copilot costs $10/month. Vercel was upgraded from hobby plan (free) to Pro ($20/person/month), and Supabase changed from Free ($0) to Pro ($25/month). Total cost was approximately $300 (around 400,000 Korean won) to launch a production service within 70 days. If they had hired a traditional development team, it would have cost roughly 100 times more.

---

The 73-day journey of AX Education Group from the first encounter with the term "Vibe Coding" to production MVP is not merely a case of technology adoption. It means democratization of development. An era has arrived where business teams without coding knowledge can take AI's hand and directly implement technology.

Facing choices between Claude Code and GitHub Copilot, the right decision isn't "picking one" but "judging when to use what." On top of modern stacks like Supabase's PostgreSQL-based databases and Vercel's global deployment infrastructure, these two AI tools work complementarily.

Vibe Coding is still a field requiring learning and judgment. However, what you now learn isn't 'programming languages' but 'the ability to convey clear requirements to AI'. This ability is much easier to learn and leads to results much faster.

AX Education Group has conducted Vibe Coding-based talent development and technology consulting for over 10 years in Seoul's Jung-gu, and under the guidance of CEOs Shim Jae-woo and Seon Ung-gyu, approximately 100 teams have already successfully built MVPs with Vibe Coding. Your next project can also be completed within 73 days. All it takes is the courage to start.

For Vibe Coding practical consultation, contact jaiwshim@gmail.com or 010-2397-5734.

---

| Item | Claude Code | GitHub Copilot | Considerations |
|------|-----------|----------------|--------|
| Learning Curve | Beginner-friendly, mentoring style | IDE-based, requires developer | Claude recommended for early projects |
| Architecture Design | Excellent (proposes overall structure) | Average (partial optimization) | Claude safer for large-scale systems |
| Development Speed | Slow (thorough) | Fast (repetitive tasks) | Copilot for implementation speed focus |
| Supabase Integration | Excellent (SQL queries, RLS policies) | Good (API route autocomplete) | Claude for complex DB design, Copilot for basic CRUD |
| Vercel Deployment | Good (environment variables, CI/CD) | Excellent (accurate environment variables) | Both strong but verify security with Claude |
| Code Quality | High (sophisticated from the start) | Medium (requires review) | Production code recommend Claude validation |
| Cost | $20/month (Claude Pro) | $10/month | Copilot cheaper but Claude recommended for time savings |
| Real-Time Feedback | Excellent (browser preview) | Good (IDE integration) | Claude for completion focus, Copilot for dev environment preference |
| By Project Situation | New projects, complex structure | Existing codebase, repetitive tasks | Hybrid utilization optimal (early: Claude, mid: Copilot, optimization: Claude) |

---

📍 Learn More About AX Education Group

•🌐 Website: https://www.yes24.com/product/goods/188879054

•📝 Blog: https://metabiz101.tistory.com/