Gustav In the last two years, "vibe coding" went from a joke on tech Twitter to a real way that real products get built. You describe what you want, an AI writes it, you nudge it until something runs. Tools like Lovable, Bolt, v0, Cursor and Replit turned that loop into a business model. Suddenly a marketer with an idea could ship a working app in a weekend. No computer science degree, no team, no permission.
That is genuinely amazing. It is also the setup for the most predictable market in software right now.
The 80% problem
Here is the pattern, and once you see it you cannot unsee it. AI builders get you to about 80% astonishingly fast. The screens look right. The demo works. You show your friends and they are impressed. You start telling people it is almost done.
Then you try to do something real with it. Add a second user. Take a payment. Put it on the internet where strangers can reach it. And the whole thing starts to wobble.
The last 20% is where software actually lives: authentication that does not leak, a backend that survives more than one person, security that does not expose your database to anyone with a browser, and code organised well enough that the next change does not break the last one. That 20% is not a small final step. It is most of the real work. It is just invisible until you hit it.
This is not a failure of the people using these tools. It is a structural feature of the tools themselves. They are optimised to produce something that looks finished, because that is what keeps users excited and subscribed. "Looks finished" and "is finished" are very different products, and the gap between them is where an entire service economy is forming.
A new kind of client just appeared
For most of software history, the person who needed a developer was a business that had decided to build something. There was a budget, a plan, maybe an agency. The conversation started at zero.
The vibe-coding wave created a completely new customer: someone who already has a half-built product, has already spent money to get it, is emotionally invested in it, and is now stuck. This person is nothing like a cold lead. They are not wondering whether to build. They built. They just cannot get it over the line.
That changes everything about the work. You are not selling a vision. You are not arguing about whether the project should exist. You are walking into a situation where the pain is sharp, specific, and present, and the only question is who can make it stop. That is the easiest kind of client to help and, frankly, the easiest kind to charge.
And there are a lot of them. Every tool that markets "build an app without code" is, by design, manufacturing future stuck founders at scale. The marketing promise and the engineering reality do not match, and the difference is a person sitting in front of a half-working app at 1am wondering what went wrong.
Why this is a market, not a moment
It is tempting to write this off as a temporary glitch that the AI tools will fix in their next version. I do not think so, for three reasons.
First, the tools are not incentivised to fix it. Their growth comes from how fast they get a new user to a "wow." Production hardening is slow, unglamorous, and specific to each app. It is the opposite of a viral demo. They will keep optimising for the first 80% because that is what sells subscriptions.
Second, the volume keeps rising. Every month more non-technical people start building. The output of half-finished apps is not slowing down, it is compounding. A growing pile of "almost done" is a growing pile of demand.
Third, the gap is fundamentally about judgement, not typing. Knowing that an exposed API key is a problem, that a missing database index will kill you at 500 users, that your auth flow has a hole, that this code will be unmaintainable in a month: that is experience. An AI can write code faster than any human. It cannot yet care about the consequences of that code the way a seasoned developer does. The cleanup market is really a market for judgement, and judgement does not get cheaper just because code does.
What the cleanup actually involves
If you are curious what this work looks like in practice, it tends to cluster around the same handful of issues, over and over:
Authentication and access. Login that half works. Sessions that leak. The classic: any logged-in user can quietly see every other user's data because the queries were never scoped properly.
A real backend. The app works beautifully in the preview and collapses the moment more than one real person uses it, because there was never a proper database or server behind it.
Security. API keys committed straight into the frontend. No input validation. Data that anyone with basic curiosity can pull. This is the stuff that turns into a very bad day.
Deployment. It runs on the founder's laptop and refuses to live anywhere else. Getting it onto the real internet, reliably, is its own skill.
Maintainability. The code has grown into a tangle where every new feature breaks two old ones. The founder is now scared to touch their own product. This is the silent killer, because it does not show up in a demo. It shows up three months later as paralysis.
None of this is exotic. It is the ordinary, unglamorous craft of making software solid. The novelty is purely in who needs it now: not big companies with engineering teams, but thousands of individuals who got further than they ever expected and then hit a wall.
Who is positioned to win
Here is the part I find genuinely interesting. The people best placed to serve this market are not the purist engineers who roll their eyes at vibe coding. Those developers tend to want to throw the whole thing out and rebuild it "properly," which is expensive, slow, and insulting to a founder who is proud of what they made.
The people who win are the ones who can speak both languages. Who understand the fast, scrappy AI-building workflow and respect it, and who also have the real engineering depth to make it production-ready. The tone matters as much as the skill. "You got 80% there, that is the hard part, let me take it the rest of the way" lands very differently from "this is a mess, start over."
That combination, technical depth plus a non-judgemental attitude toward AI-built code, is rare today. It will not stay rare forever. But right now there is a clear window for developers who can sit comfortably on both sides of the line.
The shovel, this time, is judgement
Every major shift in how software gets made leaves a cleanup market behind it. Offshoring created a market for fixing offshored code. No-code tools created a market for migrating off no-code tools. AI builders are doing the same thing, just faster and at a scale we have not seen before.
The diggers are everywhere, and there are more every day. The interesting question is not whether AI can write code. It clearly can. The interesting question is who shows up when that code needs to actually work, in the real world, with real users and real money on the line.
That is the gold rush hiding in plain sight. Not building with AI. Finishing what AI started.
I do this kind of work myself, so I am admittedly biased. If you have a vibe-coded app that stalled out at 80%, I wrote up exactly how I approach unsticking it over on my Vibe Code Rescue page. And if you are a developer reading this: pay attention to this niche. The wall everyone is hitting is your opportunity.