UndieGravity LogoUndiegravity

What Is Vibe Coding? Where It Works and Where It Fails

Vibe coding means building software by prompting AI and running the result without reading the code. Where it is useful, the risks, and how to do it safely.

By UndieGravity, , 2 min read

Andrej Karpathy coined the term "vibe coding" in a post on X in February 2025. He described describing what you want to an AI, accepting the code it writes, and running it without really reading it, and he noted this was fine for throwaway weekend projects. The phrase spread fast, and today people use it loosely for almost any AI-assisted coding.

Where vibe coding works

  • Prototypes you will throw away, to test an idea or a layout.
  • Personal tools only you will use.
  • Learning: seeing a working example, then studying how it works.
  • Small scripts where you can check the output directly.

Where it fails

The problems start when vibe-coded software is used by other people or handles their data. Code nobody has read is code nobody can maintain or vouch for.

  • Security: missing permission checks, secrets in client code, and unvalidated input are common in unreviewed AI output.
  • Maintenance: the second change is harder than the first when nobody understands the structure.
  • Hidden bugs: code can look like it works in a demo and fail on edge cases such as empty lists, slow networks, or unusual input.
  • Dependencies: assistants sometimes invent package names. Installing a look-alike package can pull in malware.

A safer way to work fast with AI

  1. Start from a known base. A template or starter gives the AI clear patterns to follow.
  2. Ask for small changes and read each diff before accepting it.
  3. Keep the type checker and linter on. They catch a large share of AI mistakes for free.
  4. Review the risky parts yourself: authentication, payments, database access, and anything with user data.
  5. Test in a real browser on a phone, not only the happy path in the demo.

Prototype versus product

A good rule: vibe code the prototype, engineer the product. When a prototype proves an idea worth keeping, read through it, restructure what needs it, and add tests before real users depend on it.

Templates are a practical middle ground: you get a reviewed structure to build on, and the AI fills in your content and features within it.

Keep reading