Is vibe coding bad?

01What vibe coding actually means

Andrej Karpathy coined the term in a post on X in February 2025: you describe what you want, accept whatever the AI writes, paste error messages back in and "forget that the code even exists". He framed it as fine for throwaway weekend projects.

That definition matters, because most arguments about whether vibe coding is bad are really arguments about two different things. One is never reading the code. The other is using AI heavily while still reviewing, testing and understanding what ships. The second is how most professional developers now work. The first is what critics mean.

02Where vibe coding works well

When the cost of the code being wrong is low, vibe coding is a genuinely good way to work. You get a working thing in an afternoon instead of a week, and you learn quickly whether the idea is worth more effort.

• Prototypes and demos you will show, test with users and then rebuild properly.
• Personal tools: a tracker, a script, an internal dashboard that only you or your team touches.
• Exploring an unfamiliar stack, where the AI gets you past the boilerplate.
• One-off data jobs, where you can check the output directly.

03Where it goes wrong

The risks show up in three places: security, maintenance and speed.

Security is the clearest. In Veracode's 2025 GenAI Code Security Report, models given a choice between a secure and an insecure way to write code picked the insecure one in 45% of 80 test tasks, and the rate did not improve as the models got better at writing code that runs. If nobody reads the code, nobody catches that.

Maintenance is the slow problem. Code you do not understand is code you cannot safely change. Developers describe a pattern where each new prompt fixes one thing and breaks another, until a rewrite is the only way forward. The same thing happens with AI website builders when people keep prompting instead of deciding; the most common AI builder mistakes are mostly versions of it.

Speed is the surprise. In METR's randomized trial of experienced open-source developers working in their own repositories, tasks took about 19% longer with AI tools, even though the developers believed they had been faster. That study measured AI assistance on large, familiar codebases, not vibe coding a new app, but it is a useful warning that feeling fast and being fast are different.

04How to vibe code without the downsides

The developers who report good results treat the AI as a fast junior colleague rather than an oracle. The habits are simple and they work.

1. 1Decide what the code is forA prototype you will throw away can be vibe coded freely. Code you intend to keep needs review and tests from the start.
2. 2Commit oftenSmall git checkpoints let you roll back when the AI wanders off, instead of untangling a broken project.
3. 3Write the plan down firstA short spec of what the feature does and what it must not do keeps the AI on track across prompts.
4. 4Read anything touching data or moneyAuthentication, payments, file uploads and database queries are where AI code most often goes wrong.
5. 5Ask for testsTests are the cheapest way to know that the next prompt did not break the last feature.

05Is vibe coding bad for learning to code?

It can be. If the goal is to become a developer, accepting code you do not understand skips the part where you learn. Several developers in the Hacker News discussion said they learned far more on projects they wrote mostly by hand. The useful middle path is to vibe code to get something working, then read it line by line and ask the AI to explain anything you cannot follow.

For a website rather than an app, there is a third option: tools built for websites that make the design and structure decisions for you, so you are not debugging generated code at all. Vibe coding vs AI website builders covers when each makes sense.