I Vibe-Coded a Project by Voice. Here's What Worked.

What Is Vibe Coding?

The term comes from Andrej Karpathy, OpenAI co-founder. The idea: instead of writing code line by line, you describe what you want in natural language and let an AI model write it. You guide, review, and iterate — but the AI handles syntax.

Tools like Cursor, Claude Code, Warp, and GitHub Copilot have made this practical. Millions of developers now vibe code daily.

But here's the thing nobody talks about: vibe coding prompts are natural language. And the fastest way to produce natural language isn't typing — it's talking.

Why Voice Makes Vibe Coding Faster

We tested dictation against typing at 90 WPM. Speaking consistently wins for anything longer than a sentence. For vibe coding prompts — which are often 2–5 sentences of context-heavy description — the gap widens.

Three reasons voice works better for prompts specifically:

• You give more context when speaking. Typing feels like effort, so developers write terse prompts. Speaking feels free, so you naturally explain why you want something, not just what. More context = better AI output.
• You think in sentences, not keystrokes. Vibe coding is about communicating intent. That's what speech evolved for.
• Your hands stay on the keyboard for code review. Speak the prompt, review the diff, type corrections. Your hands never leave the code.

The Setup: EmberType + Any AI Coding Tool

EmberType is a system-wide dictation app for Mac. It types into whatever text field is focused — including Cursor's prompt bar, Claude Code's terminal, Warp's AI input, VS Code's Copilot Chat sidebar, or Windsurf's Cascade panel. One app covers every tool.

Setup takes two minutes:

1. Download EmberType (free 7-day trial)
2. Pick a Whisper model (Large v3 Turbo recommended for developers)
3. Set your global shortcut (I use Option+Space)
4. Optional: add project-specific terms to the custom dictionary — framework names, variable conventions, API names

Now press your shortcut in any app, speak, and EmberType types the transcription directly into the focused field. No copy-paste. No browser extensions. It just works.

Tool-by-Tool: What the Workflow Looks Like

Cursor

Open the AI prompt bar (Cmd+K for inline, Cmd+L for chat). Press your EmberType shortcut. Speak: “Refactor this function to use async/await instead of promise chains, and add error handling that logs to our Sentry instance with the user's session ID as context.” EmberType types it into Cursor's prompt. Hit Enter.

That prompt took 6 seconds to speak. Typing it would have taken 20–25 seconds — and I probably would have written a shorter, vaguer version.

Claude Code

Claude Code runs in the terminal. EmberType works in Terminal.app, iTerm2, Warp, and any terminal emulator. Press your shortcut, speak your prompt, and it appears at the cursor. Particularly useful for multi-step instructions: “Find all API endpoints that don't have rate limiting, add a rate limiter middleware using the token bucket algorithm, and write tests for each endpoint.”

Warp

Warp's AI command input is a text field. Same workflow — focus the input, press your shortcut, speak. Warp's natural language-to-command feature pairs well with voice: “Find all TypeScript files modified in the last week that import from the auth module.”

VS Code + Copilot Chat

Open Copilot Chat (Cmd+Shift+I), press your EmberType shortcut, speak. Works identically to Cursor's chat. You can also dictate into inline suggestions by focusing the Copilot inline input.

Windsurf

Cascade's chat panel is a standard text input. EmberType dictates directly into it. The workflow is the same as Cursor.

A Real Example: Building a Feature by Voice

I built a dark mode toggle for a React dashboard entirely by voice prompts in Cursor. Here's the actual sequence:

Prompt 1 (spoken, 8 seconds):

"Add a dark mode toggle to the settings page. Use a React context provider so the theme persists across page navigations. Store the preference in localStorage. Default to the system preference using prefers-color-scheme."

Cursor generated the ThemeContext, provider component, and toggle button. About 80% correct on the first pass.

Prompt 2 (spoken, 4 seconds):

"The toggle should be an animated switch, not a button. And the transition between themes should fade over 200 milliseconds on the body element."

Cursor refined the UI component and added CSS transitions.

Prompt 3 (spoken, 3 seconds):

"Write a test that verifies the theme persists after a page reload using React Testing Library."

Total voice time: 15 seconds of speaking. Total typing time for the same prompts: estimated 60–70 seconds. And critically, the spoken versions were more detailed than what I would have typed.

Why Offline Matters When Your Prompts Describe Code

Here's something developers should think about: when you vibe code by voice, your spoken words describe your application's logic, architecture, and implementation details. If your dictation tool sends audio to a cloud server, those descriptions are being transmitted externally.

For personal projects, maybe you don't care. For proprietary codebases — work projects, client code, anything under NDA — it's a real consideration.

EmberType runs Whisper AI entirely on your Mac's Apple Silicon chip. Your spoken prompts are transcribed locally and never leave your device. The AI coding tool still sends the text prompt to its model (that's how Cursor/Claude Code work), but the audio of your voice — which contains ambient sounds, background conversations, and vocal patterns — stays on your machine.

It's not paranoia. It's hygiene.

Tips From a Week of Voice-First Vibe Coding

• Add your stack to the dictionary. EmberType's custom dictionary prevents transcription errors on framework names. Add entries like "Next.js", "Prisma", "Supabase", "tRPC" so they transcribe correctly.
• Speak in complete thoughts. Don't dictate one sentence at a time. Speak the full prompt — context, instruction, and constraints — in one breath. More context per prompt = better AI output.
• Use voice for prompts, keyboard for edits. I found the best rhythm is: speak the prompt, review the generated code, then use keyboard for small corrections. Voice for intent, hands for precision.
• Power Mode helps. EmberType's Power Mode auto-adjusts formatting based on the active app. In terminal apps, it strips unnecessary punctuation. In editors, it keeps it.

---