SpunkArt
Long reads and site policies

What vibe coding means, and how to review AI-written code

Vibe coding is an informal name for building software mostly by describing what you want to an AI coding assistant and accepting what it writes, often without reading every line. It can produce working prototypes fast. It can also produce code nobody understands. This guide covers both sides and how to check the output.

Where the term comes from

The phrase spread in programming circles to describe a relaxed style: you talk to the assistant, run the result, paste back any error, and repeat until it works. The emphasis is on outcome rather than on understanding the code. People now use the term loosely for any heavy use of AI to write software.

A spectrum, not a single practice

StyleHow the person worksFits
Pure vibeAccepts changes without reading them; judges only by whether it runs.Throwaway prototypes, personal experiments.
GuidedReads the plan and skims the changes; tests the main paths.Small tools with low stakes.
ReviewedTreats AI output like a colleague’s pull request and reviews it properly.Anything other people rely on.

The risk grows with how many people depend on the software and what data it touches.

What AI assistants do well

  • Writing boilerplate: configuration, form handling, routine data conversions.
  • Explaining unfamiliar code or error messages.
  • Drafting tests, once you say what behaviour matters.
  • Suggesting several approaches to compare.
  • Small, well-described changes in a codebase they can read.

Where it goes wrong

  • Plausible but wrong. Code that looks right and passes a quick try, but fails on edge cases such as empty input, time zones or large files.
  • Invented interfaces. Calls to library functions or options that do not exist, or that exist in a different version.
  • Security gaps. Missing input validation, secrets written into code, overly broad permissions, queries built by joining strings.
  • Silent scope creep. A request to fix one thing quietly rewrites five others.
  • Unmaintainable growth. Each fix adds another layer, and after weeks nobody can explain how the parts fit.

How to review AI-written code

Read the change as a diff, not as a finished file. Then work through this list:

  1. Does it do what was asked, and only that? Unrelated edits are a reason to reject and ask again.
  2. Can you explain each part in one sentence? If not, ask the assistant to explain, then check the explanation against the code.
  3. Are the imports and functions real? Look up anything unfamiliar in the library’s own documentation.
  4. What happens with bad input? Try empty values, very long values, wrong types and missing files.
  5. Where does outside data enter? Form fields, URLs, files and API responses must be validated and escaped.
  6. Are there secrets in the code? Keys and passwords belong in environment variables, never in the repository.
  7. Do the tests test anything? A test that only checks that a function runs, without checking the result, proves little.
  8. Did anything get deleted? Removed checks, error handling or tests are easy to miss in a large diff.

A small example

You ask for a function that averages a list of numbers. The assistant returns:

function average(nums) {
  return nums.reduce((a, b) => a + b) / nums.length;
}

It works for [2, 4, 6], returning 4. With an empty list, reduce without a starting value throws an error. A reviewer asks: what should happen for an empty list? Then adds a check and a test for that case.

Questions

Can a non-programmer build real software this way?

They can build working prototypes. Running software that handles other people’s data or money safely still needs someone who can review the code and understand the risks.

Does AI-written code need a different review from human code?

The same standards apply. The difference is volume: an assistant can produce more code than you can comfortably read, so ask for smaller changes.