Where vibe coding plateaus: the ceiling moves three times

The plateau is not where you think it is

The common advice online describes the vibe coding plateau as a single thing. Technical complexity gets too high. The generated code is hard to debug. You hit the doom loop where the agent keeps saying it fixed the bug and it did not. All of that is real. All of it describes the same plateau: the one where the model is the bottleneck.

That framing has a hidden assumption baked in, that the plateau sits in one place and stays there. It does not. The moment you clear the model bottleneck, the curve does not stop climbing. It climbs to a new ceiling, and that ceiling has nothing to do with the model. Then you clear that one too, and find a third.

So the useful question is not "does vibe coding plateau" (it does, three times). The useful question is which plateau am I on right now, because the fix for each one is completely different, and applying the wrong fix wastes weeks.

Plateau 1: the model

This is the plateau every other guide describes, and it is the least interesting one, because it is the easiest to leave. You are at the model plateau when the output is wrong in ways a smarter model simply would not be: it ignores half your prompt, loses a file you set up earlier, or writes code that throws the moment it loads.

The fix is not cleverer prompting. It is a different model. mk0r makes this a one-click move, and it does something quietly honest with how it presents the choice. The model picker in src/components/header.tsx does not show you raw vendor names. It maps each model ID to a word that describes how the model feels to build with:

The free tier is pinned to one model. In src/app/api/chat/model/route.ts the constant is literally FREE_MODEL = "haiku", and selecting anything else without a subscription returns a subscription_required 403. The point is not the paywall. The point is that climbing off the model plateau is a known, bounded, one-click action. If switching from Scary to Fast+ to Smart fixes your output, you were never really plateaued. You were just on the wrong rung.

Plateau 2: the friction

This is the plateau almost nobody writes about, and it is the one that kills the most hobby projects. Your app works. The model is fine. And you have quietly stopped opening the project.

It is not laziness. It is arithmetic. Every iteration has a cost you pay before you type a single word of prompt: log back in, find the session, re-explain context the tool forgot, wait on a build, click through a deploy. Each step is small. Stack them and the cost of one iteration rises above the value of one iteration. At that point the loop is mathematically not worth running, and you feel that as "I will get back to it later." You do not.

The friction plateau is removable, but only by a tool built to remove it. mk0r's entire shape is an answer to this one:

• No account. You open the site and start building. There is no signup step between an idea and the first prompt.
• Live streaming preview. The app renders as the model writes it. You are not waiting on a final export to see whether the change landed.
• One commit per turn. Inside the session VM, every agent turn is its own git commit (the commitTurn function in src/core/e2b.ts). Undo is a click that restores a real previous state, not a re-prompt that asks the model to remember.

None of that makes the model smarter. It makes one iteration cheap. When iteration is cheap, the loop survives, and a surviving loop is the whole game. That is also the bridge to the last plateau, because a cheap loop is the only thing that gives you enough reps to hit it.

Plateau 3: your taste

Clear the model plateau and the friction plateau and you arrive somewhere specific. The model does exactly what you asked. It does it fast. There is no account, no deploy, no waiting. And the app is still not good.

You go to write the next prompt and it does not come, because you cannot name what is wrong. The layout is a little off. The empty state is confusing. There is one screen too many in the flow. You can feel that it is not right and you cannot turn the feeling into an instruction. That is the taste plateau, and it is where vibe coding actually plateaus for anyone who has cleared the first two.

Here is the honest part, the part the voice of this product insists on. No tool removes this plateau. A stronger model does not. A stronger model executes your specification more faithfully; it does not supply the specification. Opus-grade output of a mediocre instruction is a more polished mediocre app. The model raises the floor on execution. It cannot raise the ceiling on judgment.

And yet this is the good plateau. It is the only one of the three that responds to practice rather than to a purchase. The model plateau you leave by paying for a better model. The friction plateau you leave by choosing a tool with less friction. The taste plateau you leave by doing reps: building, looking, noticing one concrete thing that is off, fixing it, looking again. Being stuck here is not failure. It is the only kind of stuck that is also training.

Which plateau are you on?

The fixes are completely different, so misdiagnosing costs weeks. The fastest test: ask what happens when the model does exactly what you literally typed.

What this framing does not solve

A few honest limits, so this reads as a map and not a sales pitch.

• The three plateaus are not perfectly sequential. You can sit on the friction plateau and the taste plateau at once. The order is the order you usually notice them in, not a strict ladder.
• Some ceilings really are walls, not plateaus. Real multi-user state, native mobile, a production payments backend: those do not yield to more prompts or a smarter model. That is a separate problem, covered in the iteration-wall guide below.
• Clearing the first two plateaus faster does not make your taste grow faster on its own. It just gives you more reps per hour. The reps still have to be deliberate. Cheap iteration plus no attention is just fast noise.