What Separates People Who Are Good at Vibe Coding
When I run vibe coding training, people who sit through the exact same session end up with wildly different results. The difference usually isn't coding skill. It's one attitude: how much they're willing to hand off to the AI.
The AI has a curious habit. Even when it's perfectly capable of doing something itself, it keeps handing it back to the human. "Go to this site and download this," "turn this on in the settings menu." The more obediently someone follows that guidance, the more likely they are to burn out along the way. People who are good at this push back instead: "You do it."
Case 1 — Have the AI Install Git
It starts at the very first gate of vibe coding. Ask the AI to run some code, and the first thing it says is "install git." You have to go to a site you've never seen, download a program you don't recognize, and get through an unfamiliar install process. Most people get stuck right here.
Instead of going through all that yourself, one line does it: "you install it." The agent checks your current computer setup, picks the right command, and installs it for you. All those "Do I click next? Install? What do I pick here?" moments where people get stuck — you can hand the whole thing off.
Case 2 — Have the AI Set Up Vercel and Supabase
Go a little further and you meet Vercel and Supabase. Vercel is a service that puts the website you built onto the internet so other people can reach it, and Supabase is a service that easily gives you a database for storing things like user accounts and posts. Once you build anything real with vibe coding, you'll almost certainly pass through both.
The problem is the setup. You have to sign up, go into the settings menu, and switch on this, enter that — and there are a lot of menus, most of them in English. Here the answer is the same: "you set it up." You do only the parts that need a human identity, like signing up, and hand the rest of the configuration to the agent.
One tip. When the AI works on Vercel or Supabase directly, the CLI resolves things faster than MCP. Ask it to "do it through the CLI." I'll unpack what MCP and CLI each are another time. For now it's enough to remember: when the AI is working on settings, have it use the CLI.
Told to install git → "you install it"
Told to set up Vercel/Supabase → "you set it up" (I only sign up)
Told to check localhost/deploy → "you go check it yourself"
CLI beats MCP on settings → "do it through the CLI"Case 3 — Have the AI Test and Verify It Too
The step where you check whether what you built actually works is no different. The AI commonly says "it's running on localhost, go check it yourself," or "it's deployed, verify it in your browser." Here, localhost refers to a temporary test address that runs only inside your own computer.
Here too you can hand it off with "you go check it yourself." Today's agents open the browser on their own, look at the screen, click buttons, and — if something's wrong — even use that as the basis to fix it. It cuts out the round trip of a human eyeballing the result and explaining it back in words.
So What Does the Human Do?
By this point you might wonder, "then what am I even doing?" There is clearly work left for the human. First, anything that needs a human identity or payment. Signing up, registering a card, verifying your identity — the AI can't stand in for you on these.
Second, deciding what to build and judging whether you like the result. Setting the direction and deciding whether this is what you actually wanted is the human's part. Hand the "hands and feet" — installing, configuring, verifying — to the AI, and the human is freed to spend energy on the "what and why." Delegation doesn't make people lazy; it lets them focus where it matters more.
What Delegation Really Means
To sum up: when the AI tells you to go do something, ask back — "you can do that, right? You do it." Unless it's genuinely something only a human can do, the AI will do it if you tell it to. It was capable all along; it just handed it back to the human out of habit.
Vibe coding is often described as "delegating to the AI." More precisely, it's about not taking on yourself the work the AI could have done. The difference between people who are good at it and people who struggle comes down to exactly where they draw that line.