Vibe coding
Vibe Coding
What is vibe coding?
- Describe it in plain language, and AI writes the code
- No need to write code yourself
- An era when anyone can build
What vibe coding actually changes is not the tool but the distance between a request and its result. Before, a business user wrote a requirements document, put it in the IT team’s development queue, and weeks later reviewed the screen and filed another change request. Now the person who needs it puts the screen up themselves and fixes it on the spot, saying “move this button to the right, sort the list by deadline.” The builder and the user become the same person, and that is the core shift in this approach.
In practice, AI coding assistants and vibe coding are often used interchangeably, but the dividing line is whether a person reads the generated code. A developer who reviews the AI’s suggestions line by line is doing AI pair programming; someone who never opens the code and judges only by whether the result works is vibe coding. That also sets the scope where it applies. It is strong for work that can be undone if it goes wrong, such as a sign-up page used by 20 to 30 people in-house, a throwaway data-cleaning script, or a prototype for an investment review. For areas where failure is costly, such as payments, payroll, or public-facing APIs, you still need someone who reads the code.
What decides the outcome is not prompt-writing skill but how well you know what you want to build. Ask for everything at once, such as “build me a training management system,” and you get a screen that looks plausible but is unusable, and the more you fix it, the more earlier features break. Instead, start with “a screen that shows an Excel upload as a table,” confirm it works, then add sorting, duplicate removal, and email sending one at a time, so you can pinpoint exactly where a problem appeared. The people who build fastest are those who can write their work process down in five steps on paper.
The most common misunderstanding is reading “you don’t need to know code” as “you don’t need to take responsibility.” Real incidents aren’t dramatic: access keys left hard-coded in AI-written code, applicants’ names and phone numbers exposed at a URL that opens without a login, and a page nobody can touch once its creator moves to another department. Working does not mean safe. Before you open a tool to the company, simply checking that someone can explain in one sentence where the data is stored will catch most incidents.
John KimInstructor John Kim, author of B2B Marketing Design and CEO of Nanum Consulting
Seolyeong ParkGenerative AI and AX training specialist with 17 years of hands-on experience and 200+ lectures, helping non-technical staff put AI to work right awayThe term was coined in February 2025 by AI researcher Andrej Karpathy (OpenAI co-founder, former head of AI at Tesla), who described it as “fully giving in to the vibes and forgetting that the code even exists.” It was added to Merriam-Webster in 2025 and named Collins Dictionary’s Word of the Year.
Consider a hypothetical case: Company A, an auto parts manufacturer with 480 employees. Its two training managers handle sign-ups for about 60 courses a year, which arrive scattered across email and the internal messenger, and compiling them in Excel alone took 6 hours a week. They asked the IT team to build a sign-up page and were told “it’s on the backlog until the second half of this year.”
- Ms. Kim on the talent development team gave an AI coding tool a one-sentence request, “a web page that takes an employee ID, name, and course title and adds them to a list,” and had a first screen in 20 minutes.
- Rather than demanding a finished product in one go, they confirmed the sign-up screen worked, then added the next request: “block the same employee ID from signing up for the same course twice.”
- When they got stuck on handling over-capacity registrations, they asked Mr. Park on the IT team for a 30-minute review, and he caught on the spot that applicants’ contact details were exposed at a URL that opens without a login.
- They restricted access to the company network and ran a two-week pilot with 150 people across 3 courses. During the pilot they asked for two additions: “sort by deadline” and a “download to Excel” button.
- The 6-hour weekly consolidation task dropped to 40 minutes. Instead of waiting 8 months in the development queue, they rolled it out across all of the company’s programs in 11 days, and missed second-half registrations fell to zero.