In February 2025, Andrej Karpathy, former Director of AI at Tesla and co-founder of OpenAI, introduced a term that quickly caught on. He called it “vibe coding.” Here’s how it works: instead of sitting for hours writing code line by line, you converse with an AI agent that has access to your editor. You state your goals, the agent writes, you read the diff, and accept or reject it. Done.
A year later, the ecosystem changed rapidly. Cursor has become the default choice for many engineers in Silicon Valley. Claude Code and GitHub Copilot Workspace followed suit. Then a question that used to feel awkward began to frequently arise in product team conversations: do we still need to hire new developers, or is it enough to have people who understand the product plus an AI agent?
The answer is not black and white. But what’s interesting: the role of non-developers in the midst of this question has changed. From mere users, they can now become creators.
What Is Vibe Coding, Really?
Vibe coding is not magic. What has changed is: the portion of your time that was previously spent memorizing syntax is now redirected to tasks that are closer to your actual work. Defining problems. Deciding what goes into v1 and what gets held back. Reading whether the agent's output makes sense or is a blunder. Deciding when to rollback.
Karpathy himself laughed when he tweeted that he “barely touched the keyboard” while working on his last side project. But note: he knows exactly what he is asking for. And he knows when the agent's output needs correction.
Why Non-Developers Can Now
Indonesia has a persistent developer gap. Data from Kominfo 2025 states that the demand for digital human resources will reach around 9 million by 2030, while available talent is far below that number. Interestingly, this gap will not be resolved simply by increasing the number of computer science graduates. Many small business problems do not require an enterprise-level engineering team. They need a simple app for tracking inventory, a landing page to launch a product, or a small dashboard to view daily numbers.
This is where non-developers plus AI coding agents become a sensible combination.
“The hottest new programming language is English.”
Andrej Karpathy, January 24, 2023
That tweet became a meme in its time. Now it feels almost literal.
Three Principles for Collaborating with a Coding Agent
First, treat the agent like a smart but easily confused junior developer. Juniors need context. Juniors need clear PRDs, specific constraints, and concrete examples. Not a one-liner instruction like “make a login.” What makes sense is: “create a login page using email plus OTP, save the session in JWT, validate errors and display them inline below the field, do not modify files in the utils folder.”
Second, work in small loops. Plan, build, test, commit. Then repeat. A single jumbo prompt like “create a complete e-commerce application” is almost guaranteed to explode. But ten small prompts, each generating a diff that you can read in three minutes? That’s a consistent way to achieve stable results.
Third, drop checkpoints often. Git is no longer just a tool for engineers. Git is your save point. Every time a small feature works, commit. If the agent starts to blunder and messes up the working code, you can rollback without crying.
Signs You Are Ready to Move from “Using Apps” to “Building Apps”
You have a specific product idea. You know who the first users are, even if it’s just yourself or five friends. You are ready to go through the stage of “why is this error happening again?” without running away. And you understand that the AI agent is a co-pilot, not an autopilot. If you fully autopilot it, you will end up with an application that runs on your laptop but fails in production.
If those four things apply, you are actually just a few weeks of practice away from your first live URL.
References:
- Andrej Karpathy – Vibe coding tweet thread → x.com/karpathy
- Kominfo RI – Roadmap for Digital Human Resource Needs in Indonesia 2030 → kominfo.go.id
- Cursor – AI code editor docs → cursor.com/docs
- Anthropic – Claude Code documentation → docs.anthropic.com