instructions text, not a new field or event. The reasoning behind them is in Prompting for spoken output.
Start from nothing. Dialog-RSN-1’s default text style is already tuned for speech: plain spoken sentences, no markdown, no headings, no bullet lists. You don’t need a recipe to get that, and a paragraph spent ruling out screen formatting is prompt you could have spent on your business logic. The recipes below are the overrides worth making on top of the default.
Saying numbers the way a caller expects
Tell the model which reading you want, since the same digits support more than one. Match this list to what your agent actually says out loud.instructions: say numbers naturally
Keeping replies short
Set an explicit ceiling instead of trusting brevity to the model’s judgment. A reply that reads fine as a paragraph plays back as a caller wondering if the agent is ever going to stop talking.instructions: cap reply length
Tagging a reply with an emotion
If your text-to-speech takes a style or emotion input, let the model pick one per reply. Offer the set of tags you support and ask for the answer prefixed with one in square brackets, then strip the tag off in your harness before the text reaches the synthesizer. Skip that step and the caller hears the bracket and the tag word spoken aloud, ahead of the reply. See the streaming to TTS guide for where that split happens in your pipeline.instructions: prefix each reply with an emotion tag
Overriding the default style
If you’re replacing the default voice with a persona of your own, restate what still has to hold. A long persona instruction can pull the model back toward writing for a reader, and this is the case where naming the formatting rules earns its place.instructions: a persona that stays speakable
Next steps
Prompt guide
What the default style already covers, and which overrides are worth making.
Overview
How Dialog-RSN-1 writes each reply before it’s spoken, and where your own TTS fits in.