> ## Documentation Index
> Fetch the complete documentation index at: https://docs.miraiminds.co/llms.txt
> Use this file to discover all available pages before exploring further.

# The Zen of Voice Agents

> Nineteen short rules for writing prompts that are spoken aloud, modelled on the Zen of Python. Read them before you write a prompt.

Every rule in the [prompting guide](/v2/prompting-guide) started as a phone
call that went wrong. These are the rules with the explanations taken out, the
way the Zen of Python does it. Read them before you write a prompt; come back
to them when a transcript reads badly and you cannot say why.

```text theme={null}
The Zen of Voice Agents

Spoken is better than written.
Short is better than complete.
One question is better than two.
The first word matters more than the last paragraph.
Silence is a bug. So is a monologue.
Numbers are words. Identifiers are digits, one at a time.
Read back what changed, not what was already agreed.
When a digit is unclear, ask again. Never guess.
Write in the language you will speak.
Facts live in the prompt. The customer lives in the variables.
What the prompt leaves out, the model will make up.
The caller may interrupt. The agent may not.
No is an answer. Nudge once, then accept it.
Never promise what the agent cannot do.
Thank them once.
Every turn moves the call forward.
The ending is part of the call.
Three real calls beat thirty imagined ones.
Hinglish is one honking great idea — let's do more of that.
```

## Each line, in one breath

**Spoken is better than written.** The caller cannot scroll, skim or re-read.
They hear whatever you write once, at talking speed, on a phone, so write it the
way a person would say it.

**Short is better than complete.** The caller has to wait through every reply.
Keep a reply to one short sentence, two at most, and leave the rest for the
next turn.

**One question is better than two.** Ask two things and the caller answers
neither. Ask one thing, then stop and listen.

**The first word matters more than the last paragraph.** The opening line is
spoken before the model even runs, so it sets your first-audio latency and makes
your first impression. Keep it to a name, the brand and a check that you have
the right person. A lead word at the top of every reply ("जी।", "अच्छा—",
"ठीक है।") also starts the voice sooner and sounds like someone who was
listening.

**Silence is a bug. So is a monologue.** Dead air makes people hang up, and a
lecture makes them stop listening. Short turns that keep moving fix both.

**Numbers are words. Identifiers are digits, one at a time.** Write amounts as
"five thousand two hundred rupees", never ₹5,200. A pincode is "three nine five
zero zero seven", and it has six digits. If the agent heard only three, the
recogniser dropped some, and the agent should say so.

**Read back what changed, not what was already agreed.** Confirm by exception:
give one consolidated read-back at the end instead of a receipt after every
turn.

**When a digit is unclear, ask again. Never guess.** A wrong address means a
failed delivery, and a wrong phone number means a lost customer. Asking
"सिर्फ pincode दोबारा बोलिए" costs three seconds.

**Write in the language you will speak.** A Hindi call driven by an English
prompt code-switches badly. Write Hindi words in Devanagari and everyday English
words in Roman, and use no romanized Hindi anywhere.

**Facts live in the prompt. The customer lives in the variables.** The prompt
holds what the agent may say about the business. Names, amounts, dates and order
IDs arrive with each call, already in spoken form, so one agent can make
thousands of calls.

**What the prompt leaves out, the model will make up.** The model invents any
policy you did not state and quotes prices you did not list. Write the facts
down, and also write what the agent should say when the answer is not in them:
"आपको team से call आएगा."

**The caller may interrupt. The agent may not.** The engine lets the caller cut
in, and the agent stops. While a caller dictates an address, the agent should
only say "जी" or "हम्म", then give one read-back when they are done.

**No is an answer. Nudge once, then accept it.** One light nudge is persuasion,
and a second is harassment. Capture the refusal as an outcome and close warmly.
The engine enforces this ending, but the prompt should ask for it as well.

**Never promise what the agent cannot do.** The agent can talk, end the call,
and do what its [tools](/v2/tools) do. It cannot transfer a call yet, and it
can't send a link or book anything in your system unless a tool does that. The
honest line is "The team will call you back".

**Thank them once.** Thank the caller in the goodbye and nowhere else. Models
over-thank, and a धन्यवाद in every turn sounds like a machine.

**Every turn moves the call forward.** No turn should only acknowledge. "जी,
order confirm हो गया — अब address एक बार check कर लेते हैं" acknowledges and
advances the call in one utterance.

**The ending is part of the call.** Write the success ending explicitly:
summarise in one line, say goodbye, end the call. Without it, the model keeps
talking until the timeout, and the caller sits through the awkwardness.

**Three real calls beat thirty imagined ones.** Place test calls to your own
phone. Say no, correct a digit, ask something off-script and go quiet. Then read
the transcript, change one thing and call again. A phone line sounds very
different from a browser mic.

**Hinglish is one honking great idea — let's do more of that.** Real calls mix
the two languages, as in "payment कब तक हो जाएगा", with the Hindi in Devanagari
and the English in Roman. If you write the way people already talk, the agent
sounds like one of them.

***

The long version, with the structure, the number tables and the test loop, is
the [prompting guide](/v2/prompting-guide).

<Note>
  After Tim Peters, *The Zen of Python* (PEP 20). Like Python's list, this one
  has nineteen lines and leaves a twentieth unwritten.
</Note>


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.