The most common mistake in voice agent deployments is taking an instruction set written for a chat assistant and wiring it to a phone line. The two channels look the same but their behavioral rules are opposites: in chat the customer reads, on a call they listen. A long answer that is read is acceptable; a long answer that is listened to is torture.
The ten rules below frame an instruction set that works on a voice call.
1. Cap answer length in words
"Be brief" is not enough, because brief is relative. Give a number instead: stay under forty words on normal answers; on topics needing explanation, give a two-sentence summary first and go deeper only if the customer asks.
This single rule produces the largest single improvement in voice agent quality.
2. Build lists to be spoken
A six-item list is normal in chat. On a call, by the third item the customer has forgotten the first.
The instruction should be: name at most three options, then ask "which of those suits you?" If there are more options, group them into categories first.
3. Have numbers spoken in speakable form
Prices, dates, phone and order numbers must be intelligible when read aloud. Reading an order number in groups rather than in one run lets the customer write it down. Add a rule: "say long numbers in groups of three, slowly."
4. Spell out what to do on unknown topics
This is the single most critical item. The instruction must be this explicit: on anything not in the knowledge base, do not guess, do not approximate, do not generate from general knowledge. If you do not know, say so and hand off.
A model's helpfulness drive pushes it to fill gaps. If you do not suppress that with instruction, production produces invented prices, invented opening hours and invented promotions.
5. Have it state its identity up front
The agent must say it is not human in the greeting. Saying so when asked is not enough, and being found out destroys trust. Also write that it must not impersonate a human: not "I won't say I'm human," but "I won't conceal that I'm an AI assistant."
6. Put handoff rules in the system, not in the prompt
Do not leave the handoff decision to the model. "Hand off if necessary" is interpreted differently by every model. Define hard thresholds instead: hand off if the customer asks for a representative a second time, if the topic is a refund amount or a complaint, or if the call exceeds the set duration.
We cover this in detail in tomorrow's post.
7. Define the purpose of the call in one sentence
What the agent exists for belongs at the very top of the instruction, in a single sentence: "The purpose of this call is to create the caller's appointment or change an existing one."
An agent with a vague purpose cannot steer a conversation and gets pulled into every topic. An agent with a clear purpose gently returns the conversation to its goal.
8. Do not overdo personality
Instructions for a witty, chummy or over-energetic personality usually backfire on voice. A customer trying to get something done finds jokes from the other side irritating.
The personality description that works is plain: calm, clear, respectful, no wasted words. Brand voice is built through word choice, not personality.
9. Define how a call ends
When and how the agent closes a call must be written. Without it, one of two failures occurs: it drags the conversation out, or it hangs up before the customer is ready.
The pattern that works: summarize what was done, confirm, ask whether there is anything else, deliver the closing line.
10. Revise the prompt using call recordings
An instruction set is not a document you write once and leave. At the end of the first week, listen to fifteen recordings and turn every problematic moment into an instruction line.
That loop raises quality more than anything else in the first month, and it requires no technical investment.
Three common mistakes
Letting the prompt keep growing. Instruction sets that gain a new line for every problem eventually contradict themselves. Before adding a line, check whether an existing one can be fixed instead.
Confusing the knowledge base with the prompt. The prompt describes behavior; the knowledge base carries content. Putting your price list in the prompt makes updates harder and bloats the instruction.
Writing in one language and running in several. If the agent will also speak English, test that the behavioral rules produce the same result in that language. Politeness patterns and levels of directness work differently across languages.
You can review the agent setup flow and the digital worker model on the digital workers page.