One wrong path or number can change a coding request. Prepare one authorized task, check its captured text, send deliberately and inspect the result. We checked Anthropic's voice manual on October 25, 2026. BIG CHANGE did not test the client or a project.

The big change

  • What changed: Hold mode offers a reviewable prompt before Enter.
  • Why it matters: You can compare the words on screen with a written task before giving the agent an instruction.
  • What to watch: A checked prompt is only an input. Review Claude's answer and any project changes separately.

Prepare one real task

Write a short task card before recording. Name the project and exact target, the change or question, any boundaries on what Claude may do, and an observable result. For a code change, include the project's normal checks and who accepts the work. Keep this card beside the terminal; it gives you something precise to compare with the transcript and output. Do not speak secrets or protected project details without permission.

Voice needs local Claude Code, Claude.ai sign-in and a microphone. API-key, Bedrock, Google Cloud Agent Platform, Microsoft Foundry, SSH and cloud sessions are excluded; WSL needs WSLg. Anthropic receives the audio. Check permission before speaking.

Prove the prompt will wait

Check the settings that apply to this session. Anthropic's settings guide explains that managed and session settings can override project and user files. In Claude Code, inspect loaded sources:

Snippet
/status

Inspect these user and project files, plus any higher-priority setting:

Snippet
~/.claude/settings.json
.claude/settings.json
.claude/settings.local.json

Merge this valid JSON into the appropriate writable scope, preserving every unrelated key and any existing voice keys. It is an example object, not a replacement settings file:

JSON
{
  "voice": {
    "autoSubmit": false
  }
}

If the effective value remains uncertain, type the task. A file being loaded does not establish which value won. With an empty prompt, start hold mode:

Snippet
/voice hold

Hold Space until listening… appears, speak at least three harmless words, then release. Confirm the text waits for Enter; clear it. Stop if it sends. Tap mode sends three-word transcripts on its second tap, so avoid it here. If hold fails, use the keybinding guide and retest, or type.

Speak, review, send, check

  1. From an empty composer, hold Space, speak your actual authorized task, release and wait for final text. Compare every path, name, number, command, requested action and permission boundary with your task card. Edit errors before proceeding.
  2. Press Enter only when the text matches your intended request. Observe Claude's actual answer and any proposed or performed changes. Run the project's appropriate checks and record what passed, failed or remains uncertain. A correct transcript does not establish a correct result.
  3. When finished, turn recording off and check its indicator:
Snippet
/voice off

Esc or Ctrl+C cancels recording or processing, never submitted work. Inspect earlier action. Voice mode may persist; retest next time.

If something fails

For missing /voice, check account eligibility; for silence, microphone access; for poor recognition, language:

Snippet
/config

Correct the text or type. See troubleshooting. Transcription uses no Claude messages or tokens; submitted work follows plan limits. We measured no result or cost.

For a separate Ubuntu workflow, see How to use Myna local dictation on Ubuntu.