Cookies

We use what's needed to run the site. With your yes we also count visits and remember which partner sent you. Privacy notice

Only necessaryAccept all
Cookie details
Login session Necessary
Keeps you logged in
rocket_consent Necessary
Remembers this choice, on every site
theme Necessary
Light or dark, as you picked
Visitor statistics (no cookie) Statistics
Counts page views; IP and browser hashed with a daily key, never stored
rk_ref Marketing
Remembers which partner recommended us
Statistics only, no marketing
Rocket Journal
ThemeSystemLightDark
Product

What vibe coding actually solves, and what it doesn't

AI coding tools collapse the cost of a first version. They don't touch the reason most products fail.

By Rocket Editorial

2 min read

Part of our series on taste and shipping fast. The opening essay: Anyone can ship now. Taste decides who wins.

What it solves

Be honest about the gains, because they are large. With a coding agent you can:

  • Go from an idea to something clickable in hours instead of weeks.

  • Try three versions of a screen and throw two away without guilt.

  • Put a real product in front of real people before you have raised money or hired anyone.

  • Build the boring parts (sign-in, billing pages, admin tables) without spending your best weeks on them.

All of that is about one thing: the cost and time of making something. Time-to-market has stopped being a privilege of well-funded teams.

What it doesn't solve

Most products do not fail because they were built too slowly. When CB Insights studied 431 venture-backed companies that shut down since 2023, 43% of those with a known cause had poor product-market fit: they built something not enough people wanted badly enough. Only running out of money was more common, and that is often what a weak fit turns into. Faster building does not fix product-market fit. It lets you reach the same wrong answer sooner, with more code to maintain.

There is also a trap in how fast it feels. In a randomised study by METR in early 2025, experienced open-source developers expected AI to make them 24% faster. On their real tasks they were 19% slower, and afterwards they still believed they had been about 20% faster. METR is careful to say this was one setting with early-2025 tools, and that its newer data looks different. The lesson that holds is not about the number. It is that the feeling of speed and actual progress can come apart, and you won't notice from the inside.

Developers themselves are cautious. In Stack Overflow's 2025 survey, more developers said they distrust the accuracy of AI tools than trust it. The tools are used everywhere and trusted by few. That is a sensible place to be.

The useful framing

Vibe coding is not a way to create products. It is a way to ship fast. Those are different things. Creating a product means deciding who it is for, what it must do well, what it must never do, and why anyone would switch to it. A model will happily build any answer to those questions. It can't tell you which answer is right, because the answer isn't in the code. It is in the people you are building for.

Use AI for what it is excellent at: making the cost of trying almost zero. Then make the trying count.