Vibe coding with voice: what it is, the tools, a workflow, and the risks

The short answer

Vibe coding is building software by describing what you want to an AI in plain words and letting the AI write the code. The name comes from AI researcher Andrej Karpathy, who used it in a post on X on 2 February 2025, and in that same post he said he talks to his AI coding tool instead of typing. People vibe code in tools like Cursor, Claude Code, Lovable, Replit, Bolt and v0. Voice fits because most of the work is explaining: what the app should do, what looks wrong, and what to change next. Speaking lets you give the whole idea, with its details, in a few minutes. A workflow that holds up: say the full idea, ask for a plan, build in small steps, run it and report what breaks, and save each working step. The risks are real, so read the code before others use it, keep keys and passwords out of prompts, and add tests.

What is vibe coding?

Vibe coding is a way of building where you stop writing code line by line and start describing. You say what you want, the AI writes it, you run it, and you describe what to change. You judge the result by how it works, not by reading every line.

The term comes from one post. On 2 February 2025, Andrej Karpathy wrote on X: "There's a new kind of coding I call 'vibe coding'". He described giving in to the AI so fully that you "forget that the code even exists." The detail most guides leave out comes a sentence later: he said he just talks to his AI coding tool, using a dictation app, instead of typing. So voice was part of vibe coding from the first day it had a name.

In practice, vibe coding covers a range. At one end, someone with no coding background builds a small app for themselves in an afternoon. At the other, an experienced developer describes features to an agent and reviews every change. Both are describing instead of typing code. The difference is how much they check.

Which tools do people use for vibe coding?

These are the tools people name most often, each in its maker’s own words from its own page. We have not tested them side by side, so this is not a ranking.

Vibe coding tools, quoted from each maker's own page, read 19 October 2026. Not a ranking and not a test.

Tool	In its maker’s words	Good to know
Cursor	"your coding agent for building ambitious software"	A desktop app to download, plus a command line tool and an agent on the web
Claude Code	"an agentic coding tool that reads your codebase, edits files, runs commands"	Runs in the terminal, code editors, a desktop app and the browser
Lovable	"a full-stack AI development platform for building, iterating on, and deploying web applications"	You build web apps by giving instructions in plain language
Replit	"the AI platform that helps you turn your ideas into real outcomes"	Its Agent writes the code and keeps changing it as you ask
Bolt	Build "apps & websites by chatting with AI"	You build the whole thing in a chat
v0	"an AI agent that helps anyone create real code and full-stack apps and agents"	Its page says it deploys on "infrastructure powered by Vercel"

Swipe the table to see every column

A quick way to choose: if you already have code on your computer, start with an editor like Cursor or with Claude Code in your terminal. If you are starting from nothing and want a working page in the browser, the browser builders (Lovable, Bolt, Replit, v0) ask the least of you up front.

Why does voice fit vibe coding so well?

Because vibe coding is mostly talking. You are not typing code anymore. You are explaining an idea, describing what you see on the screen, and saying what is wrong. That is what speech is good at.

• You give more context. When people type, they write the short version: "add dark mode". When they talk, they add the parts that matter: where the switch goes, what to remember, what must not break. Anthropic’s advice for Claude Code says it plainly: "The more precise your instructions, the fewer corrections you'll need."
• Long briefs stop being a chore. A real DanDin user dictated a 605-word brief to an AI tool in 5.6 minutes. Typed at 40 words a minute, the same brief takes about 15.
• You can talk while you look. Describing what is wrong on the screen is easier out loud, with your eyes on the app, than typing it in a chat box.
• Your own language works. If you think in Arabic, you can explain the idea in Arabic, and a dictation app with translation hands the AI English.

Voice is weaker for exact things: file names, function names, code you want copied as is, and error messages. Paste those. Talking the idea and pasting the exact bits is the mix that works.

A practical vibe coding workflow

This works in any of the tools above. Each step is short on purpose.

1. Say the whole idea in one go

   Who it is for, what it does, what the first screen shows, and what done looks like. Do not stop to tidy your words. A minute of talking beats three lines of typing.

2. Ask for a plan before any code

   Say "plan first, no code yet". In Claude Code, plan mode does this: Claude reads and proposes without changing files. Read the plan and fix it out loud before you approve it.

3. Build one small piece at a time

   One screen, one feature, one fix. Small steps are easy to check and easy to undo.

4. Run it, then say what you see

   Describe what happened in plain words: "the button does nothing on a phone". For errors, paste the exact message instead of reading it out.

5. Save every working step

   Commit to git, or use your tool’s history if it has one. When a later change breaks things, you go back one step instead of starting over.

6. Review before anyone else uses it

   Read the changes, or have someone who codes read them. Ask the AI to explain anything you do not understand, and ask it for tests.

A worked example: a spoken brief becomes a prompt

This is a real brief in a mix of Saudi Arabic and English, the way many builders talk, written as said and run through DanDin’s engine on 11 October 2026, in its "What I meant" mode with English output, with Cursor as the app in front. It is the same example as on our homepage.

As said: «عندي تطبيق مهام بـ React، أبي أضيف dark mode، يكون له زر في الـ settings، ويحفظ اختيار المستخدم في الـ localStorage، ولو الجهاز نفسه dark يشتغل لحاله، وانتبه لا تكسر الـ tests الموجودة.»

DanDin’s real output:

I want to add a dark mode feature to my React task application.

Please implement the following:

1. Toggle Switch: Add a dark mode toggle within the Settings page.
2. Persistence: Save the user's preference in localStorage.
3. System Preference: Automatically enable dark mode if the user's OS is set to dark mode.
4. Regression Testing: Ensure that all existing tests remain passing and are not broken by these changes.

Look at what survived. Every requirement the speaker said is there, including the warning at the end, which became its own step about tests. The mixed Arabic and English became one clean English prompt, so the AI never has to guess between two languages.

What is still missing is the part only you know: which files hold the settings page, and how you run the tests. Say that next, or type it, and the prompt is ready to send.

What are the risks of vibe coding?

Vibe coding makes it easy to build something that works on your screen and fails for everyone else. Three habits cover most of the risk.

Read the code before others rely on it

An AI can write code that looks right and misses the hard cases. Anthropic calls this "the trust-then-verify gap" in its Claude Code best practices, and the fix it gives is blunt: "If you can't verify it, don't ship it." For anything other people will use, read the changes, ask the AI to explain them, or ask a second AI session to review them with fresh eyes.

Keep secrets out of prompts and out of the code

• Never paste or say an API key, a password or a customer’s data in a prompt. Speech goes through a server too: Claude Code’s /voice sends audio to Anthropic, and DanDin sends your voice to its own servers.
• Keep keys in an environment file that git ignores, and check the AI did not write a key straight into the code.
• If a key ever lands in a prompt, a commit or a screenshot, replace it with a new one.

Make the AI prove it works

Ask for tests with every feature, and ask the AI to run them. Anthropic’s first piece of advice for Claude Code is to give it "a check it can run: tests, a build, a screenshot to compare." Without a check, "looks done" is the only signal, and that is how bugs reach real users.

Vibe coding by voice with DanDin

DanDin is our product, so weigh this part with that in mind. DanDin (dandin.com) is a dictation app for Mac and Windows, built for Arabic and English. Hold Fn on a Mac, or Ctrl and the Windows key on a PC, talk, and let go. The text lands at your cursor in whatever you build with: Cursor’s chat, Claude Code in your terminal, or a builder in your browser like Lovable, Bolt, Replit or v0.

It drops the "ums", applies your mid-sentence corrections and adds punctuation. Speak Arabic, English or both. Turn on Translate to English and Arabic speech arrives as English, or pick "What I meant" for a rewrite like the example above. One dictation can run up to 30 minutes. The free plan is 2,000 words every seven days. DanDin needs the internet, and there is no phone app.

To try it with nothing to install, our free voice to prompt tool turns a spoken idea into a clear English prompt in your browser.

For Claude Code in particular, including its own /voice command and what to do about Secure Keyboard Entry, see how to talk to Claude Code. If you keep repeating the same rules to your AI, put them in a skill: our Claude Skills guide has a complete one to copy. And for mixed Arabic and English speech, see Arabic and English in one dictation.