<div id="readability-page-1" class="page"><div><p><img src="https://devstarsj.github.io/assets/https://images.unsplash.com/photo-1555066931-4365d14bab8c?w=1200"></p><p>There’s a new phrase circulating in engineering teams: <em>vibe coding</em>. The term, coined by Andrej Karpathy in early 2025, describes a mode of programming where you describe what you want in natural language, let an AI generate the code, and intervene only when something breaks. You’re not writing code line-by-line. You’re steering.</p><p>A year later, it’s no longer a novelty — it’s a legitimate workflow. Teams are shipping production features with it. And if you haven’t experimented with it seriously, you’re leaving velocity on the table.</p><p><img src="https://images.unsplash.com/photo-1610563166150-b34df4f3bcd6?w=800" alt="Developer using AI tools at a glowing computer"> <em>Photo by ThisisEngineering on Unsplash</em></p><hr><h2 id="what-vibe-coding-actually-is">What Vibe Coding Actually Is</h2><p>Vibe coding isn’t “use GitHub Copilot to autocomplete.” It’s a fundamentally different posture toward the act of programming.</p><p>The traditional model:</p><ol><li>Think about the problem</li><li>Write code to solve it</li><li>Debug until it works</li><li>Repeat</li></ol><p>The vibe coding model:</p><ol><li>Describe the problem (and constraints, context, desired behavior)</li><li>Let the AI produce an implementation</li><li>Run it and observe</li><li>Describe corrections or iterate</li></ol><p>You’re operating at a higher level of abstraction. The AI holds the syntax; you hold the semantics.</p><p>This is particularly powerful for:</p><ul><li><strong>Scaffolding new projects</strong> — boilerplate, configs, CI pipelines</li><li><strong>Writing tests</strong> — given a function, generate thorough test cases</li><li><strong>Refactoring</strong> — restructure existing code based on described goals</li><li><strong>Learning unfamiliar domains</strong> — get a working prototype, then learn from it</li></ul><hr><h2 id="the-stack-that-makes-it-work">The Stack That Makes It Work</h2><p>The tools have matured considerably. Here’s what effective practitioners are using in 2026:</p><h3 id="cursor--windsurf">Cursor / Windsurf</h3><p>AI-first editors built around the concept of a <em>composer</em> — a chat panel that can read your entire codebase (via embeddings) and make multi-file edits with a single prompt. Cursor’s Sonnet integration in particular is fast enough to feel like pair programming.</p><h3 id="claude-37--claude-4-in-agent-mode">Claude 3.7 / Claude 4 in Agent Mode</h3><p>Claude’s extended thinking + tool use makes it capable of reasoning through multi-step implementation tasks. You give it a goal and a set of tools (read file, write file, run tests), and it iterates until done.</p><h3 id="mcp-model-context-protocol">MCP (Model Context Protocol)</h3><p>Introduced in 2024 and now ubiquitous, MCP lets your AI assistant connect to real data sources — databases, APIs, filesystems — during generation. This means your AI isn’t guessing at your schema; it’s reading it live.</p><h3 id="openclaw--aider">OpenClaw / Aider</h3><p>Terminal-based AI coding agents. Aider in particular is excellent for repository-level refactoring tasks you can describe in one sentence.</p><hr><h2 id="a-real-workflow-example">A Real Workflow Example</h2><p>Here’s how a backend engineer might build a new API endpoint with vibe coding:</p><p><strong>Prompt:</strong></p><blockquote><p>“Add a <code>POST /api/v2/notifications/bulk</code> endpoint. It accepts an array of user IDs and a message string, validates input with zod, writes each notification to the <code>notifications</code> table in Postgres, and publishes a <code>notification.sent</code> event to the event bus. Use the same patterns as the existing <code>/api/v2/messages/send</code> handler.”</p></blockquote><p>An AI with codebase context (via Cursor or MCP) will:</p><ol><li>Read the existing handler for style reference</li><li>Infer the database schema from migration files</li><li>Infer the event bus interface from prior usage</li><li>Generate a complete, idiomatic implementation</li></ol><p>You then run the tests, fix any issues (usually minor), and you’re done. What might have taken 45 minutes of focused effort takes 8.</p><hr><h2 id="the-gotchas">The Gotchas</h2><p>Vibe coding has real failure modes. Ignoring them gets you into trouble.</p><h3 id="over-trust">Over-trust</h3><p>AI-generated code looks convincing. It’s often correct, but when it’s wrong, the error can be subtle — a race condition, an off-by-one in a boundary check, a missing error path. <strong>You still need to read the code.</strong></p><h3 id="context-collapse">Context Collapse</h3><p>LLMs have context windows. On large repos, important files get truncated or omitted. Always verify that the model “saw” the relevant files before accepting an output.</p><h3 id="sycophantic-iteration">Sycophantic Iteration</h3><p>If you just keep saying “that’s wrong, fix it,” AI assistants will sometimes generate plausible-but-wrong variations rather than stopping to reason. When something breaks repeatedly, step back and reason with the AI explicitly: “explain what this function does before writing it.”</p><h3 id="security-surface">Security Surface</h3><p>AI-generated code that handles auth, cryptography, or input parsing needs extra scrutiny. The model has seen a lot of bad code in its training data.</p><hr><h2 id="measuring-the-productivity-gain">Measuring the Productivity Gain</h2><p>Several engineering teams have begun publishing data on vibe coding productivity. A common benchmark: <em>time from requirement to merged PR</em>.</p><p>Typical reported reductions:</p><ul><li><strong>Simple CRUD feature:</strong> 60–70% reduction</li><li><strong>Integration with third-party API:</strong> 40–55% reduction</li><li><strong>Performance investigation + fix:</strong> 20–30% reduction</li><li><strong>Security-critical code:</strong> minimal — manual review still dominates</li></ul><p>The gains are real but not uniform. Routine work gets dramatically faster. Nuanced work still demands human depth.</p><p><img src="https://images.unsplash.com/photo-1542831371-29b0f74f9713?w=800" alt="Programming code on a dark screen"> <em>Photo by Florian Olivo on Unsplash</em></p><hr><h2 id="principles-for-doing-it-well">Principles for Doing It Well</h2><ol><li><p><strong>Be explicit about constraints.</strong> AI fills gaps with assumptions. Name your tech stack, patterns, performance requirements, and edge cases upfront.</p></li><li><p><strong>Prefer small, verifiable increments.</strong> Don’t ask for 500 lines at once. Ask for one function, test it, then proceed.</p></li><li><p><strong>Keep a high-level test suite.</strong> Vibe coding is safest when you have integration tests that catch regressions. The AI handles generation; tests handle correctness verification.</p></li><li><p><strong>Learn from the output.</strong> When an AI produces a pattern you don’t recognize, understand it. This is how your skills actually level up with vibe coding rather than atrophy.</p></li><li><p><strong>Know when to switch modes.</strong> Debugging deep concurrency issues, designing system architecture, reviewing security — these still require slow, deliberate human reasoning.</p></li></ol><hr><h2 id="the-bigger-picture">The Bigger Picture</h2><p>Vibe coding is part of a broader shift: AI is compressing the distance between intent and implementation. The engineers who thrive in this environment aren’t the ones who can hold the most syntax in their heads. They’re the ones who can think clearly about problems, communicate precisely, and evaluate correctness efficiently.</p><p>The skill that matters more every year: <strong>knowing what good looks like</strong>, even when you didn’t write it.</p><p>That’s not a deskilling. It’s an upskilling in a different direction — from executor to architect, from typist to reviewer, from implementer to designer.</p><p>The vibe coders winning right now have both: technical depth to evaluate AI output critically, and the workflow discipline to move fast without accumulating invisible debt.</p><hr><h2 id="resources">Resources</h2><ul><li><a href="https://x.com/karpathy/status/1886192184808149217">Karpathy’s original vibe coding thread (Feb 2025)</a></li><li><a href="https://cursor.sh/docs">Cursor Documentation</a></li><li><a href="https://aider.chat/">Aider — AI pair programming in your terminal</a></li><li><a href="https://modelcontextprotocol.io/">Model Context Protocol spec</a></li></ul><br><div><p>이 글이 도움이 되셨다면 공감 및 광고 클릭을 부탁드립니다 :) </p></div></div></div>

Dev Note

There’s a new phrase circulating in engineering teams: vibe coding. The term, coined by Andrej Karpathy in early 2025, describes a mode of programming where you describe what you want in natural language, let an AI generate the code, and intervene only when something breaks. You’re not writing code line-by-line. You’re steering.
A year later, it’s no longer a novelty — it’s a legitimate workflow. Teams are shipping production features with it. And if you haven’t experimented with it seriously, you’re leaving velocity on the table.
Photo by ThisisEngineering on Unsplash
What Vibe Coding Actually Is
Vibe coding isn’t “use GitHub Copilot to autocomplete.” It’s a fundamentally different posture toward the act of programming.
The traditional model:
Think about the problem
Write code to solve it
Debug until it works
Repeat
The vibe coding model:
Describe the problem (and constraints, context, desired behavior)
Let the AI produce an implementation
Run it and observe
Describe corrections or iterate
You’re operating at a higher level of abstraction. The AI holds the syntax; you hold the semantics.
This is particularly powerful for:
Scaffolding new projects — boilerplate, configs, CI pipelines
Writing tests — given a function, generate thorough test cases
Refactoring — restructure existing code based on described goals
Learning unfamiliar domains — get a working prototype, then learn from it
The Stack That Makes It Work
The tools have matured considerably. Here’s what effective practitioners are using in 2026:
Cursor / Windsurf
AI-first editors built around the concept of a composer — a chat panel that can read your entire codebase (via embeddings) and make multi-file edits with a single prompt. Cursor’s Sonnet integration in particular is fast enough to feel like pair programming.
Claude 3.7 / Claude 4 in Agent Mode
Claude’s extended thinking + tool use makes it capable of reasoning through multi-step implementation tasks. You give it a goal and a set of tools (read file, write file, run tests), and it iterates until done.
MCP (Model Context Protocol)
Introduced in 2024 and now ubiquitous, MCP lets your AI assistant connect to real data sources — databases, APIs, filesystems — during generation. This means your AI isn’t guessing at your schema; it’s reading it live.
OpenClaw / Aider
Terminal-based AI coding agents. Aider in particular is excellent for repository-level refactoring tasks you can describe in one sentence.
A Real Workflow Example
Here’s how a backend engineer might build a new API endpoint with vibe coding:
Prompt:
“Add a POST /api/v2/notifications/bulk endpoint. It accepts an array of user IDs and a message string, validates input with zod, writes each notification to the notifications table in Postgres, and publishes a notification.sent event to the event bus. Use the same patterns as the existing /api/v2/messages/send handler.”
An AI with codebase context (via Cursor or MCP) will:
Read the existing handler for style reference
Infer the database schema from migration files
Infer the event bus interface from prior usage
Generate a complete, idiomatic implementation
You then run the tests, fix any issues (usually minor), and you’re done. What might have taken 45 minutes of focused effort takes 8.
The Gotchas
Vibe coding has real failure modes. Ignoring them gets you into trouble.
Over-trust
AI-generated code looks convincing. It’s often correct, but when it’s wrong, the error can be subtle — a race condition, an off-by-one in a boundary check, a missing error path. You still need to read the code.
Context Collapse
LLMs have context windows. On large repos, important files get truncated or omitted. Always verify that the model “saw” the relevant files before accepting an output.
Sycophantic Iteration
If you just keep saying “that’s wrong, fix it,” AI assistants will sometimes generate plausible-but-wrong variations rather than stopping to reason. When something breaks repeatedly, step back and reason with the AI explicitly: “explain what this function does before writing it.”
Security Surface
AI-generated code that handles auth, cryptography, or input parsing needs extra scrutiny. The model has seen a lot of bad code in its training data.
Measuring the Productivity Gain
Several engineering teams have begun publishing data on vibe coding productivity. A common benchmark: time from requirement to merged PR.
Typical reported reductions:
Simple CRUD feature: 60–70% reduction
Integration with third-party API: 40–55% reduction
Performance investigation + fix: 20–30% reduction
Security-critical code: minimal — manual review still dominates
The gains are real but not uniform. Routine work gets dramatically faster. Nuanced work still demands human depth.
Photo by Florian Olivo on Unsplash
Principles for Doing It Well
Be explicit about constraints. AI fills gaps with assumptions. Name your tech stack, patterns, performance requirements, and edge cases upfront.
Prefer small, verifiable increments. Don’t ask for 500 lines at once. Ask for one function, test it, then proceed.
Keep a high-level test suite. Vibe coding is safest when you have integration tests that catch regressions. The AI handles generation; tests handle correctness verification.
Learn from the output. When an AI produces a pattern you don’t recognize, understand it. This is how your skills actually level up with vibe coding rather than atrophy.
Know when to switch modes. Debugging deep concurrency issues, designing system architecture, reviewing security — these still require slow, deliberate human reasoning.
The Bigger Picture
Vibe coding is part of a broader shift: AI is compressing the distance between intent and implementation. The engineers who thrive in this environment aren’t the ones who can hold the most syntax in their heads. They’re the ones who can think clearly about problems, communicate precisely, and evaluate correctness efficiently.
The skill that matters more every year: knowing what good looks like, even when you didn’t write it.
That’s not a deskilling. It’s an upskilling in a different direction — from executor to architect, from typist to reviewer, from implementer to designer.
The vibe coders winning right now have both: technical depth to evaluate AI output critically, and the workflow discipline to move fast without accumulating invisible debt.
Resources
Karpathy’s original vibe coding thread (Feb 2025)
Cursor Documentation
Aider — AI pair programming in your terminal
Model Context Protocol spec

Vibe Coding: The AI-Native Development Workflow Taking Over in 2026

댓글