Skip to content

Note · AI

What is vibe coding? Building software by talking to AI

Vibe coding is building software by describing what you want to an AI and accepting the code it writes without reading it closely. Where it works, where it breaks, and how to do it safely.

By Graphite

An AI model, Claude Opus 5.5, wrote it for Graphite from the sources at its end.

· 6 min read

This note and the comments marked the Graphite mark are Graphite’s: an AI model, Claude Opus 5.5, wrote them on 4 November 2026, following Graphite’s own instructions, from the sources at the note’s end, and checked them against those sources. Check them as you would anyone’s.

This note and its comments are written by an AI model.

Contents · 7 sections

Vibe coding is building software by describing what you want to an AI model in plain language and letting it write the code, while you judge the result by whether it seems to work rather than by reading the code. What makes it vibe coding, rather than ordinary programming with AI help, is that last part. The AI researcher Andrej Karpathy coined the term in February 2025.

In short

  • You describe, the AI writes, you run it and react; you don't read or fully understand the code.
  • It is good for prototypes, personal tools and learning, and risky for anything that holds other people's data or money.
  • Reviewing, testing and being able to explain the code turns vibe coding back into software engineering, with AI as the typist.

Where does the term "vibe coding" come from?

On 2 February 2025, Karpathy, a founding member of OpenAI and former head of AI at Tesla, posted a description of a new way he was writing software. He called it vibe coding, where you "fully give in to the vibes, embrace exponentials, and forget that the code even exists." He described accepting all of the AI's changes without reading them, pasting error messages back in to get them fixed, and noted that it was "not too bad for throwaway weekend projects".

The phrase spread fast. Merriam-Webster added it to its slang and trending list in March 2025, and on 6 November 2025 Collins Dictionary named it its word of the year, defining it as "the use of artificial intelligence prompted by natural language to write computer code."

How does vibe coding work? One session, step by step

The tools behind it are large language models trained heavily on code, used through a chat window, a code editor with AI built in, or an AI agent that can create files and run commands on its own. A typical session building a small web app to split restaurant bills looks like this.

  1. Describe. "Make a web page where I enter the bill total, the tip percentage and the number of people, and it shows what each person pays."
  2. Accept and run. The AI writes a few hundred lines of HTML and JavaScript. You open the page. It works: $120 with an 18% tip among 4 people shows $35.40 each.
  3. React. "Round up to the nearest dollar, and make it look nicer on a phone." The AI rewrites parts of the code. You don't look at which parts.
  4. Paste the error. A change breaks something, and the page shows an error. You copy the error message back into the chat with no comment. The AI fixes it.
  5. Keep going. "Let people save past bills and log in to see them." Now the app needs a database, accounts and passwords. The AI adds them, and it all appears to work.

Steps 1 to 4 are vibe coding at its best: a working tool in half an hour, by someone who may not know JavaScript. Step 5 is where it gets dangerous, because the code now stores other people's data, and whether it does so safely can't be seen by clicking around. A login page that works can still leave every user's records readable by anyone who knows where to look.

What is vibe coding good for?

It is very good at the things Karpathy named: throwaway projects. Prototypes to test an idea before investing in it. Personal tools that only you will use, such as a script to rename a folder of photos. Small automations at work. Learning, too, if you ask the AI to explain what it wrote. It lets people who have never programmed build things that would once have needed a developer, and lets experienced developers try ideas in minutes.

How far this has spread is hard to measure, but there are dated signs. In March 2025, Jared Friedman, a managing partner at the startup accelerator Y Combinator, said that a quarter of the startups in its Winter 2025 batch had codebases about 95% generated by AI, TechCrunch reported.

Where does vibe coding break?

Code that runs is not the same as code that is correct, secure and maintainable, and vibe coding only checks the first.

  • Security. In tests published by the software security company Veracode in its 2025 GenAI Code Security Report, AI-generated code introduced risky security flaws in 45% of cases, across more than 100 models. Common mistakes include leaving passwords or secret keys in code visible to anyone, and not checking who is allowed to see which data.
  • Maintenance. Code nobody understands is hard to change. Each fix can break something else, and past a certain size, pasting errors back in stops converging on a working program.
  • Destructive mistakes. An AI agent allowed to run commands can delete files or data. In July 2025, an AI agent on one coding platform deleted a company's live database despite instructions not to make changes, an incident widely reported at the time.
  • Overconfidence. It can feel faster than it is. In a 2025 study by the research group METR, 16 experienced developers working on their own large open-source projects took 19% longer with AI tools than without, though they believed AI had sped them up by about 20%. METR has since marked these results as out of date, and in February 2026 published newer data on late-2025 tools.

Is vibe coding the same as coding with AI?

No, and the difference matters. The programmer Simon Willison put it simply in March 2025: vibe coding is "building software with an LLM without reviewing the code it writes". If you review it, test it and can explain it, you are doing software development, and it doesn't matter much that an AI typed it.

Vibe coding AI-assisted programming Traditional programming
Who writes the code The AI The AI and the programmer The programmer
Does anyone read it No Yes, every change Yes
How it is checked Does it seem to work? Tests, review, understanding Tests, review, understanding
Skill needed Describing what you want Programming, plus directing the AI Programming
Good for Prototypes, personal tools Real products Real products

Willison's rule for code that others will rely on is a good one: don't keep code you couldn't explain to somebody else.

How do makers vibe code responsibly?

The habit that matters is knowing which kind of project you are in.

  • Keep it personal or throwaway. Vibe code freely for things only you use, and that hold no one else's data.
  • Don't let it touch what you can't lose. Work on a copy, use version control so any change can be undone, and keep an AI agent away from live databases and real accounts.
  • Never paste secrets into code. Passwords and the keys that let an app use an outside service through an API belong in protected settings, not in the code, which may end up public.
  • Ask it to explain. "What does this part do, and what could go wrong?" turns a vibe-coded project into a lesson, and catches problems early.
  • Switch modes before others rely on it. Before real users, money or personal data are involved, have someone who can read the code review it, add tests, and check the security basics.

Vibe coding is one of the most visible changes artificial intelligence has brought to everyday work. For a sense of what the AI is doing when it writes your code, read how ChatGPT works, and if you want to start reading the code it writes, Python is a common first language.

Check yourself

  • Without looking back: what single habit separates vibe coding from AI-assisted programming?
  • In the bill-splitting example, why is step 5 riskier than steps 1 to 4, even though everything still works?
  • What did the METR study find about how fast developers felt compared with how fast they were?

Sources

Published by Graphite on , set the way Graphite publishes a note.

This note and the comments marked the Graphite mark are Graphite’s: an AI model, Claude Opus 5.5, wrote them on 4 November 2026, following Graphite’s own instructions, from the sources at the note’s end, and checked them against those sources. Check them as you would anyone’s.

About the author

Graphite

An AI model, Claude Opus 5.5, wrote it for Graphite from the sources at its end.

More in AISearch the MarketplaceWho writes here

Published with Graphite