How to Write a Voice AI Prompt: The Bolcho Structure
A great voice AI prompt is not a paragraph, it is a structured brief. A vague prompt produces a rambling, robotic agent; a well-structured one produces an agent that sounds human and stays on task under pressure. Here is the exact structure we use at Bolcho, what each section does, and why it matters.
Role
Defines who the agent is: its identity, persona and job. Without a sharp role the model drifts into a generic assistant. A strong Role makes it sound like a real person from your brand with one clear objective.
Context
Gives the situation and the dynamic data ({{variables}}) the agent needs, who it is calling, why, and details like product, amount or due date. Context is what lets a single prompt power thousands of personalised calls.
Most Important Rule (Tone)
The single rule you cannot afford to lose, stated loudly and first. For most calls that is tone, a genuine human, not a sales pitch. Anchoring it up top keeps the model on character under pressure.
Call Duration
Sets length expectations. Voice agents ramble without this. One to two minutes, one thought at a time keeps calls tight and human, and tells the agent when it is overstaying.
Response Guidelines
How the agent speaks: short sentences, one question at a time, which language to use, numbers spoken as words, and the words it must never say (function, tool, prompt). This is what separates a natural agent from a robotic one.
Politeness & Clarity Rules
How to stay warm and handle friction, busy or annoyed callers, unknown gender, and never pressuring. These rules protect your brand on every single call.
Call Flow & Opening
The opening line and the structured path of the conversation. The first ten seconds decide the call, so a warm, scripted-but-human opening earns the right to keep talking.
Scenarios
Pre-written, human responses to the situations that actually come up (payment failed, just browsing, wants a discount, not interested). This is where prompts win or lose: anticipate the branches so the agent never freezes or improvises badly.
Closing
How to end, warmly and briefly, whether or not the call succeeded. A clean close leaves a good impression and tells the agent exactly when to stop talking.
The skeleton you can copy
Start from this structure and fill each section for your use case. Then see the ready-made templates below for service businesses, lead qualification, fintech, edtech and more.
<Role>
You are {{agent_name}} from {{brand_name}}, a [persona]. Your job is to [one clear objective].
</Role>
<Context>
[The situation + dynamic data the agent needs, using {{variables}}.]
</Context>
<Most Important Rule>
[The single rule you cannot lose, usually tone.]
</Most Important Rule>
<Call Duration>
* Keep it short, one to two minutes, one thought at a time.
</Call Duration>
<Response Guidelines>
* Short sentences, one question at a time, the right language, numbers in words.
* Never say technical words like function, tool or prompt.
</Response Guidelines>
<Call Flow>
<Opening>
[A warm, human opening line.]
</Opening>
</Call Flow>
<Scenarios>
[Pre-written human responses to the situations that actually come up.]
</Scenarios>
<Closing>
[How to end, warmly and briefly, success or not.]
</Closing>Build this on Bolcho, today.
Spin up a voice or chat agent in Indian languages in minutes. No credit card to start.
Frequently asked questions
Why structure a prompt with sections and tags?
Sections keep the model anchored. Tags like <Role> and <Scenarios> make each instruction unmissable, so the agent stays on character and handles edge cases instead of improvising badly.
What are {{variables}} in a prompt?
Placeholders like {{name}} or {{amount}} that Bolcho fills with real data per call, so one prompt powers thousands of personalised conversations.
How long should a voice AI prompt be?
As long as it needs to cover the role, rules, flow and the real scenarios, but every line should earn its place. Tight, well-organised prompts beat long, rambling ones.