Prompting guide

A voice prompt is not a chatbot persona. It is a script for a two-minute phone call: one job, one caller, one ending. Everything the model writes is spoken out loud, at speaking speed, to someone holding a phone to their ear — so length is a cost, ambiguity is a hang-up, and a paragraph is a monologue. The method is the same every time. Design the call as a structure, test it against a caller who says no, refine one line at a time. This page is the mechanics; the thinking behind it is on The Zen of Voice Agents.

Why voice is not chat

Three constraints. Everything below follows from them. The caller cannot scroll, and cannot re-read. A sentence they missed is gone. Anything that matters gets said once, short, and gets read back. A long reply is the thing they wait through. They hear every word before they can answer. Reply length is the single biggest latency lever you own — bigger than anything in the stack. Cap it in the prompt, in words the model cannot round up: “one short sentence, maximum two”. Turn-taking replaces reading. In chat you can ask three things in one message and let the reader sort it out. On a call, two questions in one turn get one answer and you will not know which. One ask, then stop.

What the engine already does

A lot of what people write into voice prompts is already handled below the prompt. Writing it again does not make it truer; it just spends turns and tokens.

The opening line runs before the model does. first_message is spoken the moment the callee answers. It is your first-audio latency. The model is told what was said, so it does not greet twice.
Refusals already end the call. When end_call.enabled is true (the default), the engine appends an ending rule to every prompt: if the caller refuses or asks to end — “नहीं चाहिए”, “बात नहीं करनी”, “रखता हूं”, “call खत्म करो” — the agent does not re-sell, says one farewell sentence and calls end_call. Same on a second refusal. You do not write this.
A bare “ok” does not hang up. With end_call.confirm: true (the default), the first end_call inside the window asks the caller to confirm; the second one ends the call.
Barge-in is automatic. The caller can talk over the agent — after two words of caller speech the agent stops and listens. If they barge in without actual words, the agent says “हाँ, बोलिए ना, क्या बात है?”.
Silence is handled. If the caller goes quiet, the agent checks in after about 15 seconds; after three unanswered check-ins the call ends. Nothing to write.
Voicemail is detected by the engine and handled by the agent’s voicemail setting. Nothing to write.
Per-call data is substituted for you. {{placeholders}} in both system_prompt and first_message are filled from the call’s variables.

So do not write any of that into the prompt. Write the job.

The agent’s only tool today is end_call.No transfer to a human, no SMS, WhatsApp or email, no CRM lookup, no calendar, no payment link — nothing that fetches or sends mid-call. A prompt must therefore never promise “I’ll transfer you”, “I’ll send you the link” or “I’m booking it in the system”. The honest line is “the team will call you back” / “आपको team से call आएगा”. Mid-call data and branching flows are the node-graph runtime — see tiers.

The structure

Every house template follows one discipline. Six blocks, in this order, nothing else:

ROLE                                   who the agent is, for whom, calling why — one sentence
FACTS                                  the specific, verifiable things it may say (edit these)
CALL FLOW                              numbered steps; one ask per step; never assume what wasn't said
IF THEY PUSH BACK                      one light nudge, then accept and close
EXAMPLE EXCHANGE                       2–4 turns showing the hard moment done right
LANGUAGE RULES + CONVERSATION STYLE    the fixed block, pasted verbatim

ROLE is one sentence: name, company, and the reason for the call. Not a character study. “You are मीरा, a polite female voice agent calling from Mirai Finance to remind the customer their EMI is due and capture a clear payment commitment” is a complete role. FACTS is the boundary of what the agent may say. Amounts, dates, policy, what it cannot do. If a claim is not in FACTS the agent does not make it — the answer is that the team will call back. This is the block your customer edits. CALL FLOW is numbered steps, one ask per step, in the order a person would ask them. Steps are instructions to the agent, not lines to read aloud. IF THEY PUSH BACK is one nudge and then acceptance. Not three. The engine already stops a hard refusal; this block is for the soft “not right now”. EXAMPLE EXCHANGE is two to four turns of the hardest moment done right. Models copy tone from examples far more reliably than from adjectives — two good turns beat a paragraph about being warm and professional. LANGUAGE RULES + CONVERSATION STYLE is fixed. Paste it verbatim at the end of every prompt. It is the block below.

The opening line

first_message is spoken before the model runs, so it is pure first-audio latency and it should be short: name, brand, person check. Nothing else.

नमस्ते, मैं मीरा बोल रही हूं Mirai Finance से. क्या मेरी बात {{customer_name}} जी से हो रही है?

The pitch does not go in the opener. The model already knows what was said and will carry on from there. One thing catches everybody: Hindi verbs are gendered, and the verb has to match the voice.

`voice_id`	Voice	Say
`ashu`	male	मैं … बोल रहा हूं
`aishe`	female	मैं … बोल रही हूं

The console renders this for you when you pick a template. The API does not — if you switch voice_id on an existing agent, fix the opening line in the same request.

Language rules for Hinglish

Real Indian phone calls are not Hindi and are not English. They are Hindi grammar carrying English nouns, and an agent that picks one pure language sounds like a recording. This block is what the console appends to every template. Paste it at the end of your prompt, unchanged:

LANGUAGE RULES — keep these; they make the agent sound natural
- Everyday Hinglish: Hindi words in Devanagari script, NEVER romanized Hindi ('aap kab tak kar payenge' is WRONG — write 'आप कब तक कर पाएंगे').
- Everyday English words stay in Roman script: payment, EMI, order, address, delivery, challan, invoice, penalty, policy, service, slot.
- No formal shuddh Hindi (भुगतान, देय, राशि, पुष्टि).
- Every number is English words in Roman script — never digits, never Hindi numbers ('pandra' is WRONG, 'fifteen' is right). Identifiers digit-by-digit: 'three nine five zero zero seven'.
- When the customer states or corrects any number — flat, pincode, phone, amount — read it back digit-by-digit and confirm (e.g. 'आपका pincode है four zero zero zero five three, सही है?'). If it sounds unclear or incomplete, do NOT guess — ask them to repeat it slowly.

CONVERSATION STYLE
- One short sentence, maximum two.
- धन्यवाद appears exactly ONCE in the whole call — only in the final goodbye.
- After a clear promise or confirmation: repeat it back, say goodbye, end the call.
- Never invent policy or waive penalties; unknown questions → the team will call back.

Why each rule is there:

Devanagari for Hindi words. Romanized Hindi in the prompt gets read as English. “kar payenge” comes out with English vowels and the caller hears an accent, not a person.
Roman script for English words. Nobody on a phone call says भुगतान. They say payment. Writing the English word in Roman is not laziness — it is how the sentence is actually spoken.
No shuddh Hindi. Formal Hindi is the register of a government notice, not a phone call. It reads as a recording even when the audio is perfect.
Numbers as English words. The voice reads English number words most reliably. Digits and Hindi number words are read inconsistently.
धन्यवाद once. Models over-thank. Thanking four times in ninety seconds is the clearest tell that nobody is there.

For English calls (language: "en-IN") the shape is identical — drop the Devanagari rule and keep everything else. Numbers still go in as words, and identifiers are still read digit by digit.

Numbers, dates, amounts and identifiers

The model copies what you write straight into speech. Write for the ear.

Written	Put it in the prompt as
₹5,200	`five thousand two hundred rupees`
15/07	`fifteen July`
2:15 PM	`two fifteen in the afternoon` — or `दोपहर two fifteen`
Pincode 395007	`three nine five zero zero seven`
+91 98765 43210	digit by digit, in groups: `nine eight seven six five, four three two one zero`
MH02AB1234	`MH zero two AB one two three four`

Amounts and quantities are natural words; quantities can sit inline in a Hindi sentence (“दो items”). Identifiers — pincode, phone, order id, vehicle number, challan number — are always digit by digit, never grouped into “ninety-eight thousand”. Digit-count sanity. An Indian pincode is six digits. A mobile number is ten. If a read-back is short, the recogniser dropped digits — the agent must push back, never accept:

मुझे सिर्फ three digits मिले, दोबारा बोलिए

The read-back protocol is three moves, and it belongs in every flow that touches a number: the caller states it → the agent reads it back digit by digit → the agent asks “सही है?”. Nothing gets written down until the caller confirms.

Facts, and the customer

FACTS is a contract. The agent may say what is in it and nothing more. Anything outside — a policy question, an exception, a price you did not list — is answered with “the team will call you back”. Never invent policy, never waive a penalty, never quote a number that is not in FACTS. Per-call data goes in variables, not in the prompt. One agent runs thousands of calls; the customer’s name, amount and date change every time. Hard-coding them means one agent per customer, and nobody wants to manage that.

{ "variables": { "customer_name": "राहुल", "emi_amount": "five thousand two hundred", "due_date": "five July" } }

Notice that those values are already in spoken form. That is deliberate: the model copies a variable’s value verbatim into speech, so format it for the ear at the point you build the map, not in the prompt. 5200 is a gamble; five thousand two hundred is not. A few properties worth knowing:

Names are matched loosely. {{Customer Name}}, {{customer_name}} and {{CustomerName}} are the same variable — letters and digits only, case-insensitive.
An unknown placeholder in the system prompt becomes an empty string. It does not error, and it does not leave braces for the agent to read aloud.
Values are plain strings, and braces inside a value are stripped.
Variables are static per call. There is no way to change one mid-call.
Keep them short. They are prompt tokens on every turn.

Tell the agent what day it is. Without it the model does not know, and “tomorrow” or “this Friday” becomes a guess. The prompt supports Liquid filters, so put the real date in at call time, in IST:

Today is {{ "now" | date: "%A, %d %B %Y", "Asia/Kolkata" }}.

With the clock too, when the script mentions a time window:

{{ "now" | date: "%A, %d %B %Y, %I:%M %p", "Asia/Kolkata" }}

The flow

One ask per step. Two questions in one turn get one answer, and you will not know which one it was.
Never assume what was not clearly said. If the caller mumbled a date, the agent asks again. A confidently wrong confirmation is worse than a second question.
Confirm by exception. Read back only what changed, and do one consolidated read-back at the end. Confirming every field turns a ninety-second call into three minutes.
Listen when they are dictating. While the caller reads out an address, the agent’s job is backchannels — “जी”, “हम्म” — and then one read-back at the end. Interrupting a dictated address to confirm line one is how you lose the rest of it.
Mishear recovery is surgical. Re-ask the one field that was wrong. Never restart the flow.
Answer, then return. If the caller derails with a question, answer it in one line from FACTS and go straight back to the step you were on.
Every turn moves forward. Acknowledge and ask the next thing in one utterance. A turn that only says “ठीक है” costs the caller time and buys nothing.

Pushback and endings

One light nudge, then accept. The engine already enforces the refusal ending, so your job is the success ending — and it is the line people forget:

When the payment date is confirmed, repeat it back in one line, say goodbye, and end the call.

Without it the model keeps talking. It has nothing telling it the job is done. “No” is an outcome, not a failure. Capture it and close cleanly — a refusal that ends in twenty seconds is a good call. The close is three moves: a one-line summary of what was agreed, a goodbye, then end_call. Say धन्यवाद once, in that goodbye. Leave end_call.message empty — the default — and the model’s own goodbye is the farewell. If you do set a message, it is spoken after the model’s goodbye, so it must not repeat it: a धन्यवाद in both means the caller hears it twice. Finally, size the call. max_duration_secs is a hard cap, and hitting it ends the call with ended_reason: exceeded-max-duration mid-sentence. For a one-job call, 180–300 seconds is right. The default is 300; the range is 30–1800.

Sounding like a person

Lead word up front. “जी।”, “अच्छा—”, “ठीक है।” at the start of a turn does two things: the voice starts sooner, and it sounds like somebody who was listening rather than a system that was processing.
Punctuation is your only expressiveness control. Interjections and emphasis at the moments that carry them — an apology, a pushback — and flat everywhere else. Expressiveness everywhere is the same as expressiveness nowhere.
No stage directions, no tags, no emoji, no markdown. The voice reads the text it is given. Asterisks, bullet characters and (warmly) come out as noise.
Match the caller’s register. If they answer in English, the agent continues in English. If they switch to Hindi mid-call, so does it.
Contractions and everyday words. Write the sentence you would say out loud, not the one you would put in an email.

What not to write

Anti-pattern	What happens	Write instead
A persona paragraph — “you are a friendly, empathetic AI assistant”	Adjectives do not survive contact with a call. The model still runs long.	One ROLE sentence, and an EXAMPLE EXCHANGE that shows the tone
Digits and currency symbols: `₹5,200`, `15/07`	Read inconsistently, sometimes as separate digits	`five thousand two hundred rupees`, `fifteen July`
Romanized Hindi: `aap kab tak kar payenge`	Read with English vowels; sounds like an accent, not a person	Devanagari: `आप कब तक कर पाएंगे`
Shuddh Hindi: भुगतान, देय, राशि	The register of a notice, not a call	The words people use: payment, due date
Two questions in one turn	One answer, and you cannot tell which question it belongs to	One ask per step
Multi-branch scripts — “if they say X, ask the following seven questions”	The model loses its place and the call sprawls	A linear flow, plus one IF THEY PUSH BACK block
”I’ll transfer you” / “I’ll send the link” / “I’m booking it in the system”	The agent has no tool but `end_call`. It is a promise nobody keeps.	”the team will call you back” / “आपको team से call आएगा”
Customer data hard-coded: “Rahul’s EMI is 5200”	One agent per customer, forever	`{{customer_name}}`, `{{emi_amount}}` in `variables`
A long list of forbidden phrases	Naming a phrase primes it. Listing what not to say teaches it to say it.	The positive rule: “one short sentence, maximum two"
"Wait for the user to finish speaking”	The engine owns turn-taking. The prompt controls what is said, not when.	Nothing — delete the line
”Remind them about last week’s call”	There is no memory across calls	Pass the fact in `variables`
A 600-word FACTS block	Re-read on every turn, and the model picks the wrong fact	The ten facts this call actually needs
Emoji, markdown, stage directions	Spoken aloud as noise	Plain sentences
An English prompt for a Hindi call	Code-switches badly and drifts back to English	Write the prompt in the language the call will be in

Test, then refine

Every real phone call costs money — there is no test key — so the loop is designed to find problems in the cheapest place first.

Start from a template

Agent Studio in the console has six house scripts in the template dropdown — Govt · Traffic challan, BFSI · EMI reminder, BFSI · GSTR-1 filing, D2C · Order confirmation, D2C · Service reminder, Hospitality · Booking confirmation. All six are already in the structure above. If none of them is close, use Describe the call and we draft it → Generate script: a one-line brief comes back as a house-style prompt and opening line to edit.

Call yourself on the web phone

The web phone in the right rail runs the agent without dialling a phone. Play the customer you are afraid of: say no, correct a digit half-way through, ask something that is not in FACTS, then go quiet and see what happens. Most prompt bugs die here, for free.

Read the transcript, change one thing

curl https://sandbox.voice.miraiminds.co/v2/calls/call_01K7Q9B4M8N3P6R2S5T7V9W1YA/transcript \
  -H "Authorization: Bearer sk_live_YOUR_API_KEY"

Turns come back as they were spoken, nothing translated — see Get the transcript. Read where it went wrong, change one line, run it again. Changing three lines at once tells you nothing about which one worked.

Three real calls to your own number

The phone line is a different acoustic world from a browser mic: carrier codecs, background noise, a caller who is walking. Prompts that pass on the web phone and fail on a real handset are common, and they fail on numbers first.

A small campaign, then read it

Twenty to fifty contacts. Read the transcripts and the campaign report before you scale — see Campaigns. Fix what you find, then go wide.

Measure by outcome, not vibe. Did the call reach a defined end state — confirmed, refused, callback? “It sounded nice” is not a metric, and a warm agent that never gets a commitment is a failed call.

A worked example

The EMI reminder, rendered for the aishe voice with the agent named मीरा. system_prompt:

ROLE
You are मीरा, a polite female voice agent calling from Mirai Finance to remind the customer their EMI is due and capture a clear payment commitment.

FACTS — edit these for your use case
- Customer: {{customer_name}}
- EMI: {{emi_amount}} rupees, due {{due_date}}
- After {{due_date}}: late payment penalty may apply
- Payment: Mirai Finance app or website
- You cannot extend the due date or waive any penalty.

CALL FLOW
1. Confirm the person.
2. State the EMI amount and due date.
3. Use the penalty hook gently; ask when they will pay.
4. Never assume a date they did not clearly say; read the promised date back.
5. Close.

IF THEY REFUSE OR CAN'T COMMIT
One light nudge about the penalty, then accept and note that the team will follow up.

EXAMPLE EXCHANGE
[User] थोड़ा टाइम मिलेगा क्या?
[You] {{customer_name}} जी, {{due_date}} के बाद payment करने पर late payment penalty लग सकती है, आप कब तक payment कर पाएंगे?
[User] मैं दस जुलाई तक कर दूंगा।
[You] ठीक है {{customer_name}} जी, आपने बताया कि आप ten July तक payment कर देंगे, मैंने note कर लिया है. धन्यवाद, आपका दिन शुभ हो!

<LANGUAGE RULES + CONVERSATION STYLE block>

The last line is where the fixed block from Language rules gets pasted, verbatim. first_message:

नमस्ते, मैं मीरा बोल रही हूं Mirai Finance से. क्या मेरी बात {{customer_name}} जी से हो रही है?

Create the agent

cURL
Python
Node.js

emi-agent.json

{
  "name": "EMI reminder",
  "system_prompt": "ROLE\nYou are मीरा, a polite female voice agent calling from Mirai Finance to remind the customer their EMI is due and capture a clear payment commitment.\n\nFACTS — edit these for your use case\n- Customer: {{customer_name}}\n- EMI: {{emi_amount}} rupees, due {{due_date}}\n- After {{due_date}}: late payment penalty may apply\n- Payment: Mirai Finance app or website\n- You cannot extend the due date or waive any penalty.\n\nCALL FLOW\n1. Confirm the person.\n2. State the EMI amount and due date.\n3. Use the penalty hook gently; ask when they will pay.\n4. Never assume a date they did not clearly say; read the promised date back.\n5. Close.\n\nIF THEY REFUSE OR CAN'T COMMIT\nOne light nudge about the penalty, then accept and note that the team will follow up.\n\nEXAMPLE EXCHANGE\n[User] थोड़ा टाइम मिलेगा क्या?\n[You] {{customer_name}} जी, {{due_date}} के बाद payment करने पर late payment penalty लग सकती है, आप कब तक payment कर पाएंगे?\n[User] मैं दस जुलाई तक कर दूंगा।\n[You] ठीक है {{customer_name}} जी, आपने बताया कि आप ten July तक payment कर देंगे, मैंने note कर लिया है. धन्यवाद, आपका दिन शुभ हो!\n\n<LANGUAGE RULES + CONVERSATION STYLE block>",
  "first_message": "नमस्ते, मैं मीरा बोल रही हूं Mirai Finance से. क्या मेरी बात {{customer_name}} जी से हो रही है?",
  "voice": { "voice_id": "aishe", "language": "hi-IN" },
  "language": "hi-IN",
  "max_duration_secs": 240,
  "voicemail": { "action": "hangup" }
}

curl -X POST https://sandbox.voice.miraiminds.co/v2/agents \
  -H "Authorization: Bearer sk_live_YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d @emi-agent.json

import os, httpx

API = "https://sandbox.voice.miraiminds.co"
auth = {"Authorization": f"Bearer {os.environ['MIRAI_API_KEY']}"}

SYSTEM_PROMPT = open("emi_prompt.txt", encoding="utf-8").read()

agent = httpx.post(
    f"{API}/v2/agents",
    headers=auth,
    json={
        "name": "EMI reminder",
        "system_prompt": SYSTEM_PROMPT,
        "first_message": "नमस्ते, मैं मीरा बोल रही हूं Mirai Finance से. क्या मेरी बात {{customer_name}} जी से हो रही है?",
        "voice": {"voice_id": "aishe", "language": "hi-IN"},
        "language": "hi-IN",
        "max_duration_secs": 240,
        "voicemail": {"action": "hangup"},
    },
    timeout=30,
).raise_for_status().json()

print(agent["id"])

import { readFileSync } from "node:fs";

const API = "https://sandbox.voice.miraiminds.co";
const auth = { Authorization: `Bearer ${process.env.MIRAI_API_KEY}` };

const systemPrompt = readFileSync("emi_prompt.txt", "utf8");

const res = await fetch(`${API}/v2/agents`, {
  method: "POST",
  headers: { ...auth, "Content-Type": "application/json" },
  body: JSON.stringify({
    name: "EMI reminder",
    system_prompt: systemPrompt,
    first_message:
      "नमस्ते, मैं मीरा बोल रही हूं Mirai Finance से. क्या मेरी बात {{customer_name}} जी से हो रही है?",
    voice: { voice_id: "aishe", language: "hi-IN" },
    language: "hi-IN",
    max_duration_secs: 240,
    voicemail: { action: "hangup" },
  }),
});
if (!res.ok) throw new Error(JSON.stringify(await res.json()));
const agent = await res.json();

Keep the prompt in a file next to your code and read it in — it is a script, it will change often, and it does not belong inline in a JSON literal. end_call is left at its defaults on purpose: confirmation on, no canned farewell, so the धन्यवाद in the EXAMPLE EXCHANGE is the only one the caller hears.

Place the call

The variable values are already in spoken form. That is the point.

curl -X POST https://sandbox.voice.miraiminds.co/v2/calls \
  -H "Authorization: Bearer sk_live_YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: emi-88213-jul-1" \
  -d '{
    "agent_id": "agt_01JZQ8F3K7M2N5P9R4T6V8W0XZ",
    "to": "+919876543210",
    "variables": {
      "customer_name": "राहुल",
      "emi_amount": "five thousand two hundred",
      "due_date": "five July"
    },
    "webhook_url": "https://example.com/mirai/webhook"
  }'

Copy-paste skeleton

Fill the angle brackets, keep everything else.

ROLE
You are <AGENT NAME>, a polite <male|female> voice agent calling from <COMPANY> to <THE ONE JOB>.

FACTS — edit these for your use case
- Customer: {{customer_name}}
- <FACT: the amount, date, order, slot — in spoken form>
- <FACT: the consequence or the benefit>
- <FACT: where the customer can act — app, website, branch>
- You cannot <THE THING THIS AGENT MAY NOT PROMISE>.
Today is {{ "now" | date: "%A, %d %B %Y", "Asia/Kolkata" }}.

CALL FLOW
1. Confirm the person.
2. <STATE THE REASON FOR THE CALL, one sentence>
3. <THE ONE ASK>
4. Never assume anything they did not clearly say; read it back digit-by-digit and confirm.
5. When <THE OUTCOME> is confirmed, summarise it in one line, say goodbye, and end the call.

IF THEY PUSH BACK
One light nudge about <THE CONSEQUENCE>, then accept and note that the team will follow up.

EXAMPLE EXCHANGE
[User] <THE HARD THING THEY ACTUALLY SAY>
[You] <THE ANSWER, one or two short sentences>
[User] <THEIR COMMITMENT>
[You] <READ IT BACK, goodbye>

LANGUAGE RULES — keep these; they make the agent sound natural
- Everyday Hinglish: Hindi words in Devanagari script, NEVER romanized Hindi ('aap kab tak kar payenge' is WRONG — write 'आप कब तक कर पाएंगे').
- Everyday English words stay in Roman script: payment, EMI, order, address, delivery, challan, invoice, penalty, policy, service, slot.
- No formal shuddh Hindi (भुगतान, देय, राशि, पुष्टि).
- Every number is English words in Roman script — never digits, never Hindi numbers ('pandra' is WRONG, 'fifteen' is right). Identifiers digit-by-digit: 'three nine five zero zero seven'.
- When the customer states or corrects any number — flat, pincode, phone, amount — read it back digit-by-digit and confirm (e.g. 'आपका pincode है four zero zero zero five three, सही है?'). If it sounds unclear or incomplete, do NOT guess — ask them to repeat it slowly.

CONVERSATION STYLE
- One short sentence, maximum two.
- धन्यवाद appears exactly ONCE in the whole call — only in the final goodbye.
- After a clear promise or confirmation: repeat it back, say goodbye, end the call.
- Never invent policy or waive penalties; unknown questions → the team will call back.

That is the whole method: one job, six blocks, numbers as words, one ending you wrote yourself. When a prompt still is not landing and the structure is right, the problem is usually a belief about what a call is — that is The Zen of Voice Agents.

Start here

Build

Prompting

Text to speech

Money

Reference

Why voice is not chat

What the engine already does

The structure

The opening line

Language rules for Hinglish

Numbers, dates, amounts and identifiers

Facts, and the customer

The flow

Pushback and endings

Sounding like a person

What not to write

Test, then refine

A worked example

Create the agent

Place the call

Copy-paste skeleton

​Why voice is not chat

​What the engine already does

​The structure

​The opening line

​Language rules for Hinglish

​Numbers, dates, amounts and identifiers

​Facts, and the customer

​The flow

​Pushback and endings

​Sounding like a person

​What not to write

​Test, then refine

​A worked example

​Create the agent

​Place the call

​Copy-paste skeleton

Why voice is not chat

What the engine already does

The structure

The opening line

Language rules for Hinglish

Numbers, dates, amounts and identifiers

Facts, and the customer

The flow

Pushback and endings

Sounding like a person

What not to write

Test, then refine

A worked example

Create the agent

Place the call

Copy-paste skeleton