How to Use Voice Input with Claude Code and Codex CLI

Claude Code and Codex CLI have changed how developers write and maintain code. Instead of manually writing every line, you describe what you need and the AI handles implementation. But there is an irony: the better these tools get at generating code, the more time you spend typing prompts instead of code.

Voice input fixes this imbalance. This guide covers how to set up voice-driven prompting with PromptPaste and get the most out of it with both Claude Code and Codex CLI.

Setup

Install PromptPaste from the Microsoft Store. On first launch, it will download the local transcription model — this happens once and takes a minute or two depending on your connection. No account is required.

The default hotkey is Ctrl+Shift+Space. Hold it to record, release to transcribe and insert. The text appears wherever your cursor is focused — your terminal, editor, or any other input field.

That is the entire setup. There is no configuration required to start using it with Claude Code or Codex CLI.

Prompting Claude Code by voice

Claude Code runs in your terminal as an interactive session. When it is waiting for input, your cursor is in the terminal prompt. Hold your PromptPaste hotkey, speak your instruction, and release. The transcribed text appears in the Claude Code input, ready to send.

Voice works especially well for the kinds of prompts that make Claude Code most effective: detailed context, specific requirements, and multi-step instructions. Instead of typing out three paragraphs of context, speak it in thirty seconds.

Example: instead of typing a long prompt about refactoring a module, you might say: "Refactor the authentication middleware in src/lib/server/auth.ts to separate token validation from session creation. Keep the existing HMAC-SHA256 signing but extract the cookie handling into its own function. Add error types for expired tokens versus invalid signatures."

That level of detail takes significant typing effort but flows naturally in speech.

Working with Codex CLI

Codex CLI follows the same pattern. It accepts natural language input in the terminal, so voice insertion works identically. The key difference is that Codex CLI often works in a more autonomous mode, so your prompts tend to be higher-level descriptions of what you want accomplished.

Voice is particularly useful for Codex CLI's planning-oriented prompts, where you describe a feature or fix at a high level and let the tool figure out the implementation details.

Tips for effective voice prompting

Be specific. Voice makes it easy to add detail, so take advantage of it. Mention file paths, function names, and specific behaviors you expect. The transcription handles technical terms well — it will correctly capture terms like "useState", "kubectl", "middleware", and most programming jargon.

Speak in complete thoughts. Instead of dictating fragments and editing, pause briefly to collect your thought, then speak the full instruction. This produces cleaner transcriptions and better prompts.

Use voice for context, keyboard for corrections. After dictating a prompt, scan it quickly and use the keyboard to fix any transcription errors before sending. This hybrid approach is faster than either pure typing or pure dictation.

For multi-file changes, describe the full scope upfront. Something like: "I need to add a new API endpoint for user preferences. Create the route handler in app/api/preferences, add the types to contracts.ts, and update the entitlement check to include the new feature flag." This kind of scoped instruction is tedious to type but effortless to speak.

Common patterns

Developers using voice input with AI coding tools tend to converge on a few patterns. Bug reports become more detailed because it costs nothing to add context. Code review comments get more thorough. Commit messages become more descriptive because dictating a sentence is faster than deciding which words to cut.

The meta-pattern is that voice input removes the tax on verbosity. When communicating intent is cheap, developers communicate more intent, and AI tools perform better as a result.