Your prompt doesn’t just decide what the assistant says — it decides how spoken details land. A phone number typed as digits and a phone number spelled out for speech can sound very different once a voice reads them aloud. A few habits make the difference.
This page is about how you write the words in your prompt and knowledge base. To fix how a specific name or term is pronounced everywhere, use the pronunciation dictionary instead — the two are complementary.
The most reliable technique is also the simplest: instead of relying on the model to reformat digits on the fly, give it the exact words to say.
Phone numbers — spell out each digit and group them for clarity:
Email addresses — spell every character, and say the symbols as words:
Websites — say recognizable words normally, and spell out any segment made only of letters:
Times and dates — use a spoken format, never a raw timestamp:
Pace long strings of digits
Long numbers — account numbers, IBANs, confirmation codes — are easier to follow when they’re broken into short groups with a natural pause between them, rather than read as one unbroken string. A few short groups, separated by commas, are usually enough. Keep the grouping consistent within a prompt, and listen to the result before relying on it — pacing on punctuation can vary slightly by voice.
Don’t overdo it. Long chains of pauses make speech sound unnatural — reserve them for information the caller genuinely needs a moment to write down.
Giving the caller a moment
If a caller says they need a second — to grab an account number, for example — tell the assistant explicitly how to respond rather than leaving it to guess:
Test what it actually sounds like
Wording that reads fine on screen doesn’t always sound right out loud. Use the assistant editor’s Test button to place a real voice call and listen — especially after changing a voice, since pacing and pronunciation both depend on it.