"Voice prompts" means two different things, and people searching for it want one or the other. The first is prompting a voice assistant: speaking to ChatGPT Voice, Gemini Live or a similar mode so the spoken answer is useful. The second is prompting a voice generator: writing text and direction for a text-to-speech tool so the voice reads it the way a person would. The habits are different, and this guide takes them in turn.
Part 1: prompting a voice assistant
Spoken prompts fail for a reason that written ones do not: you cannot see what you said, so you ramble, and the model answers the last thing it heard. The fix is to speak the prompt the way you would write a good one, with the task first and the context after, and to set the format of the answer out loud, because the default spoken answer is long.
Set the rules once at the start
Voice modes keep the conversation, so a first instruction shapes everything after. Say something like:
For this conversation, keep answers under thirty seconds unless I ask for more. If I ask a question with several parts, answer the parts in order and say which one you are on. If you do not know something, say so instead of guessing.
Voice prompts that work spoken
- "I'm about to walk into a meeting about [topic]. Give me the three things I should know in one sentence each."
- "Interview me for a [role] job. Ask one question at a time, wait for my answer, then tell me what was missing before the next question."
- "Teach me [concept] while I drive. Explain it in two minutes, then ask me one question to check I got it."
- "I'm cooking. Walk me through [dish] one step at a time, and wait for me to say next."
- "Let's practise [language]. Speak only in [language], correct my mistakes after each reply, and keep it at a beginner level."
- "Brainstorm with me. I'll say an idea, you say a better version or a problem with it, in one sentence, and we go back and forth."
- "Read me the key points of what I just dictated and ask if I missed anything." (after dictating notes)
- "I'm nervous about a conversation with [person] about [topic]. Play them, be difficult but fair, and let me practise."
Two habits make a difference. Say "one at a time" for anything with steps; a voice answer that reads ten steps at once is useless by step three. And say "stop" or "wait for me" where you need a turn; otherwise the assistant fills the silence.
Part 2: prompting a voice generator
Text-to-speech tools read what you write, and they read it literally. A script that looks fine on the page comes out flat or rushed because the page does not carry the pauses, the stress or the pace a person adds without thinking. The prompt for a voice generator is the script plus the direction, and the direction is written into the text.
Write the script to be spoken
- Short sentences. A spoken sentence over twenty words loses the listener and the model's breath.
- Punctuation as timing. A comma is a short pause, a full stop a longer one, a new paragraph a breath. Use an ellipsis for a hesitation. Most generators respect these.
- Spell out what is read wrong: numbers as words where it matters ("twenty twenty-six"), abbreviations expanded, names with a phonetic hint in brackets if the tool supports it.
- Emphasis by word order, not by capitals. Put the important word at the end of the sentence. Capitals make some tools shout.
Direct the delivery
Tools that accept a style or direction field take short, concrete instructions. Direct the voice the way you would direct an actor: the character, the situation and the pace, not adjectives about sound quality.
Warm, unhurried, like a friend explaining something they find interesting. Slight smile in the voice. Pause before the key sentence. No announcer energy.
Clear and neutral, documentary narrator, steady pace, slightly lower at the end of each sentence, no emotion on the numbers.
Conversational and quick, podcast host, occasional natural hesitation, a bit of amusement in the second paragraph.
Voice prompts for common jobs
- Explainer video: script in short sentences with a new paragraph per scene; direction "clear, friendly, medium pace, pause at each paragraph for the visual to change."
- Audiobook sample: keep the author's sentences but add paragraph breaks at every shift of scene; direction "calm, measured, a touch slower on dialogue, distinguish the two speakers by pace rather than pitch."
- IVR or product voice: one sentence per line; direction "neutral, warm, slightly slower than conversation, same energy on every line."
- Ad read: direction "confident, upbeat without shouting, land the product name, pause before the offer"; write the offer as its own sentence.
- Meditation: very short sentences, ellipses for breaths; direction "soft, slow, low, long pauses, no rising intonation."
Generating the script with a chat model first
If you want ChatGPT, Claude or Gemini to write the voice script, say it will be spoken and give the length in seconds. One prompt covers it:
Write a script to be read aloud by a text-to-speech voice, [length] seconds at a natural pace, for [purpose and audience]. Short sentences, one idea each. Use punctuation for timing: commas for short pauses, full stops for longer ones, a new paragraph for a breath. Spell out numbers. No headings, no stage directions in the text. Then, on a separate line, write a one-sentence delivery direction for the voice.
Fixing a voice that sounds wrong
Flat: the sentences are too long and the direction is missing; cut the sentences and add a character to the direction. Rushed: add commas and paragraph breaks; some tools accept a speed setting, but punctuation works everywhere. Wrong stress: move the important word to the end of the sentence. Mispronounced name: add a phonetic spelling, or respell it in the text the way it sounds. Robotic numbers: write them as words.
The scripts and storytelling prompts in the library write for the ear, which is most of the work for part two, and the video ad script prompts produce the short spoken formats. For the assistant side, the tutoring prompts are built one question at a time, which is exactly the shape a voice conversation needs.
