The natural move is to reuse: the agent already handles chat well, so just switch on voice. The first conversation shows the problem. The agent reads out a numbered list of five items, spells a URL, says “see the link below” and takes a while to start, because the huge prompt that worked in text now costs time the call does not have.
Voice guidelines need four changes from text ones: answers of two or three sentences; no formatting (lists, markdown, links); at most two or three items at a time, followed by a question back; and a lean prompt, because every extra instruction adds latency. The content can stay the same. The delivery cannot.
The ear does not scroll back
A reader can reread. A listener cannot. In a spoken answer, the caller has to hold everything they heard in memory until the sentence ends, which limits the useful length of an answer to far less than it seems. Five options read in a row become one: the last.
What to drop from the text prompt
- Long lists. Mention the top two or three and ask whether the person wants to hear the rest.
- Formatting. Bold, headings and bullets do not exist in audio, and depending on the voice the symbols end up read aloud.
- Links and long addresses. If the person needs a link, offer to send it by message.
- Jargon. Nobody pauses a call to look up an acronym; keep it simple.
What to add
- An explicit length rule: “answer in three sentences or fewer”.
- A question back at the end of the answer, to hand the turn over.
- Short transition phrases, such as “right” and “got it”, that give rhythm without padding.
- What to do when it does not understand: asking the caller to repeat is acceptable in voice; guessing is not.
A good voice prompt is shorter than the text one, not longer. Every extra line is processed on every answer, and in voice it costs perceived time.
Reasoning costs silence
Higher reasoning levels improve complex answers, and in text the wait is tolerable. In voice, every second of reasoning is a second of silence on the line. For most voice support, low reasoning backed by a well-organised knowledge base does better than the reverse. The knowledge base matters too: short, specific items are retrieved faster than an eighty-page PDF.
The voice is part of the prompt too
Timbre shapes perception as much as wording. A firm voice and a warm voice reading the same greeting give different impressions of the brand. Before choosing, listen to at least three voices reading your business’s real greeting, not the sample sentence.
How to test
At Take Sales, every voice has an audio preview in the agent form, and the test call rings your phone with the agent as configured. Ask questions you know produce lists, ask for an address, interrupt mid-answer. If the agent handles text and voice at once, context carries over when the visitor switches between them, so test the switch as well. Voice agent and response latency are in the glossary.
