AI Generation

Describe what your assistant should do in plain language, and have a system prompt written for it — for the model it actually runs on.

VoiceDock includes AI-powered generation tools that help you configure your voice assistants faster. Instead of writing a system prompt or a workflow integration from scratch, describe what you need in plain language and let the AI generate a production-ready configuration for you.

AI Generation turns a natural-language description into an assistant prompt written for the model your assistant runs on.

Overview

The prompt generator is available directly in the VoiceDock dashboard when creating or editing an assistant: Generate with AI, in the system prompt section.

Tools and analysis templates have their own AI generators, on their own pages. See Custom tools and Call analysis.

Prompt Generator

The Prompt Generator creates a complete system prompt for your assistant based on a natural-language description of what the assistant should do.

It writes for the model your assistant runs on. A pipeline assistant, Gemini Live, GPT-Live and Grok Realtime differ in whether there is a separate speech step and in who decides the turn, which changes what a prompt can steer at all. The generator follows those differences and leaves out rules that the selected model would not carry out. Prompting per model explains the reasoning, and applies just as much if you write the prompt by hand.

How It Works

Open the Prompt Generator

Navigate to the assistant editor in the HMS Sovereign dashboard. Click the Generate with AI button in the system prompt section.

Describe Your Assistant

Write a description of what your assistant should do. You can type or use the built-in voice input to dictate your description. Be as detailed or as brief as you like -- the generator will fill in the gaps with best practices for voice AI.

Example input:

I need an assistant for a dental clinic. It should answer questions about our
services (whitening, implants, cleanings), check appointment availability,
and book appointments. If someone has an urgent issue, transfer them to our
front desk. The clinic is open Monday to Friday, 9am to 5pm. The assistant
should speak Dutch.

Set the Conversational Dynamics (GPT-Live only)

If the assistant runs on GPT-Live, three extra choices appear: how often the assistant makes listening sounds while the caller talks, whether it yields to every interruption or only to the caller, and what it does with a voice in the background.

They appear only on that model because only that model decides its own turns. On a pipeline and on the other realtime providers, interruptions are handled by the platform rather than by the prompt, so there is nothing to choose.

Review the Output

The generator returns:

System prompt -- written for the model the assistant runs on, following the rules that apply to it
Explanation note -- what to check, what was assumed, and anything you asked for that this model cannot carry out

Accept or Refine

Review the generated output. You can:

Accept it as-is to apply it to your assistant
Retry with feedback -- provide specific feedback and regenerate for a better result
Edit manually after accepting to make fine-tuned adjustments

Voice Input

The Prompt Generator supports voice input via the microphone button. Click it to start recording, speak your description naturally, and click again to stop. The transcription is added to the text input in real time, so you can combine voice and typed input.

Tip: Pro Tip Voice input works especially well for this feature -- describe your assistant the same way you'd explain it to a colleague. The AI handles the translation into a structured voice assistant prompt.

What Makes Generated Prompts Special

Generated prompts follow two sets of rules: the ones that hold for any phone assistant, and the ones that belong to the model this assistant runs on.

Always:

Short sentences -- kept under 15 words for natural phone delivery
One question at a time -- and the question is the last thing in its turn, because a caller starts answering the moment they hear one
Human fillers -- natural phrases like "Let me check that for you" distributed throughout
Signposting -- the assistant says what it is doing before a lookup or a tool call
No visual formatting -- no markdown, URLs, bullets, quotation marks or special characters that cannot be spoken
Structured flow -- identity, context, personality, boundaries, conversation flow, tool instructions, edge cases and sign-off
Restraint about email addresses -- the assistant prefers a phone number, and is given a way out after two failed attempts at spelling an address

Depending on the model:

Model family	What the prompt gets
Pipeline	Writing rules for the speech engine: everything spelled out, no digits or symbols in running text, commas as the only pause. No rules about interrupting, because the assistant's speech settings decide that.
Gemini Live	Persona first, rules in execution order, tool instructions in separate sentences, and the output language stated twice. Pace described in words, since there is no speech engine to slow down.
GPT-Live	The above, plus a backchannel policy, an interruption policy and instructions for background voices, because this model decides its own turns.
Grok Realtime	Persona and rules as above, without the pipeline writing rules and without turn-taking instructions.

If you ask for something the selected model cannot carry out, the generator leaves it out of the prompt and says so in the note, rather than writing a rule that quietly does nothing. Prompting per model has the full reasoning.

Tools and Analysis Templates

The prompt generator writes the prompt and nothing else. Tools and analysis templates have their own AI generators on their own pages, where they are created once and attached to any assistant that needs them:

Custom tools -- function tools, end call and transfer call
Analysis templates -- the schema and prompts used to extract structured data after a call

Describe in your prompt description what the assistant should be able to do, and the generator will write the prompt around the tools that are already attached. It will also tell you, in the note, when you describe a capability that has no matching tool yet.

Example

The same description, generated for two different models. The business logic is identical; what differs is everything the model can actually carry out.

Input:

A customer support assistant for an e-commerce store. It should look up order
status, process return requests, and escalate complex issues to a human.
Speak English. Track whether the customer's issue was resolved.

Generated for a pipeline assistant (excerpt):

You are a customer support assistant for [Company Name]. You help callers
with order inquiries, returns, and general questions.

Your tone is warm and professional. Keep answers to one or two sentences.

How you speak
Write every word out in full. Never use digits, symbols or abbreviations:
say "two hundred and forty euro", not "EUR 240". Read an order number one
digit at a time, with a comma between the digits.
Use commas as your only pause. Do not use double commas or trailing dots.

When a caller asks about an order, say that you are looking it up, then use
the check order status tool. Read the status back and ask one question at a
time. Wait for the answer before you continue...

Generated for a GPT-Live assistant (excerpt):

You are a customer support assistant for [Company Name]. You help callers
with order inquiries, returns, and general questions.

Personality
Speak warmly and at an unhurried pace. Be clear and direct, not overly
cheerful. If the caller is frustrated, acknowledge it briefly and move to
the next helpful step.

Backchannel policy
Stay silent while the caller is speaking. One short acknowledgement is
allowed only when the caller pauses inside a longer story and is clearly
waiting. When in doubt, stay silent.

Interruption policy
Stop speaking the moment the caller interrupts you, and listen. Do not
finish your sentence, and do not repeat the whole interrupted message
afterwards: summarise and continue from where it still matters.

When a caller asks about an order, say that you are looking it up, then use
the check order status tool...

Note what is absent from each. The pipeline prompt has no rule about interruptions, because its speech settings decide that. The GPT-Live prompt has no spelling rules, because there is no speech engine to spell for.

Best Practices

Prompt Generation - - Set the model before you generate. The prompt is written for the model the assistant has at that moment, so switching afterwards means generating again.
- Be specific about the assistant's purpose, language, and personality
- Attach the tools first. The generator writes around the tools that are already there, and tells you when you describe something that has no tool yet.
- Include business rules and constraints (hours, policies, limits)
- Use the retry-with-feedback feature to iteratively refine the output
- Always review and edit the generated prompt before going live

**Tip: ** A generated prompt is a starting point, not a finished product. Always review and customize it before deploying to production.

Costs

Generating a prompt uses AI processing and counts toward your organization's AI usage. The cost is displayed in the generator dialog before you confirm.

Note: Generation costs are separate from per-call costs. You are only charged when you click Generate, not when reviewing or editing the output.

AI Generation

On this page