Disclosure: this article is published by Relay. Author Raymond Zhu is a founding member. It is documentation-based, not a hands-on test. Claude features were checked against Anthropic’s help center on October 29, 2026. Sources: Claude: Use voice mode, Claude: Quick entry on Mac, Relay for macOS.
Speaking is a good way to give an AI more context.
Most weak prompts are short because typing the background is tedious. Speaking removes some of that effort: you can explain the situation, the audience and what you’ve already tried in the time it takes to type one line.
Relay believes voice can give AI more context with less strain than typing, because you can say a thought and keep developing it without first turning it into edited keystrokes. The point is simple: when the prompt decides the quality of the answer, the easier it is to say more, the better your odds.
Use the built-in dictation when you stay in one app.
ChatGPT offers dictation through the microphone in the message box: you speak, review the transcription and send it as text. That’s separate from Voice, ChatGPT’s spoken back-and-forth conversation mode. Options vary by app and platform, so check yours in the app.
Claude draws the same line. Anthropic’s voice mode guide describes dictation as turning speech into a text prompt and voice mode as a two-way spoken conversation. On a Mac, Claude Desktop’s quick entry can start dictation with Caps Lock; the shortcut is off by default and needs macOS 14 or later. Other assistants have their own voice options; check each app’s help pages.
If you’re deciding between talking to an assistant and dictating into one, ChatGPT Voice vs Mac Dictation vs Relay covers that choice. This guide is about dictating good prompts.
Sources: Claude: Use voice mode, Claude: Quick entry on Mac.
Use a system-wide layer when prompts move between apps.
Each assistant’s dictation works in that assistant. If you prompt ChatGPT in a browser, Claude on the desktop, an AI feature inside your notes app and a coding agent in the Terminal, a voice layer that works everywhere saves learning several shortcuts. Mac Dictation is one; the best dictation apps for Mac compares others.
Relay for macOS is one of those layers. It works in every app where you type with a keyboard and turns what you say into ready-to-send text. It’s an interface, like a keyboard, not a chatbot: it doesn’t answer your questions itself. It helps you get your thinking into whichever AI tool you choose.
Sources: Relay for macOS.
Structure a spoken prompt in four parts.
Goal: start with what you want. “I need a one-page summary of this contract for my manager.”
Context: say what the AI can’t see. Who it’s for, what’s already been decided, what you’ve tried, and what “good” looks like to you.
Constraints: say the limits. Length, what to leave out, the tone, any facts it must not change.
Format: say the shape of the answer. A table, bullet points, a draft email, three options to choose from.
Put together, that might sound like: “Summarize the attached contract for my manager, who hasn’t read it. She cares about cost and the notice period. Keep it under 200 words, don’t give legal advice, and finish with the three questions we should ask the supplier.”
Think out loud, then tidy it up.
Spoken prompts are messy: you restart sentences and change your mind. Built-in dictation types all of that, which an AI can usually still follow, though it may weigh your first version and your correction equally.
Voice apps that resolve self-corrections give you a cleaner prompt. Relay, for example, supports incomplete thoughts, changes of mind and spoken development of an idea, and shapes the result around what you finally meant.
Review the prompt before you send it.
Read it once. Names, figures and technical terms are where speech recognition slips, and a wrong number in a prompt produces a confidently wrong answer.
Remember there are two data paths: the dictation tool’s and the AI service’s. Check both before you speak confidential material. How Relay handles context and privacy describes Relay’s.
Type the parts that need precision.
Voice is best for the explanation around a prompt, not every character in it. Paste code, exact quotes, URLs and data rather than reading them aloud, then speak the instructions around them. A spoken prompt with a pasted excerpt is often the most reliable combination.