OpenAI Voice Agent Prompt Generator: GPT-Live, Realtime

The OpenAI Voice Agent Prompt Generator creates prompts, session settings and request templates for GPT-Live 1, GPT Realtime 2.1, 2.1 Mini and 2, transcription models, GPT Realtime Translate and GPT Audio 1.5. It is free, runs in your browser and needs no sign-up.
Models last verified: 17 October 2026

Generate your prompt

Access 40,000+ AI Prompts for ChatGPT, Claude & Codex — Free!

Subscribe to get instant access to our complete Notion Prompt Library — the largest curated collection of prompts for ChatGPT, Claude, OpenAI Codex, and other leading AI models. Optimized for real-world workflows across coding, research, content creation, and business.

Access Free Prompt Library

Which OpenAI voice model should I choose?

Compare the OpenAI models covered by this generator, including their uses and pricing.

Model	Best use	Price
GPT-Live 1	Full-duplex voice conversations that can continue while a backend agent handles reasoning and tools.	$0.05 per minute, billed per second, plus backend model tokens
GPT Realtime 2.1	Speech-to-speech agents that need configurable reasoning, instruction following, and tool use.	$32.00 audio in / $64.00 audio out per 1M tokens
GPT Realtime 2.1 Mini	Faster, lower-cost realtime voice interactions using a distilled reasoning model.	$10.00 audio in / $20.00 audio out per 1M tokens
GPT Realtime 2	Earlier speech-to-speech model; OpenAI’s realtime prompting guide was written for it	$32.00 audio in / $64.00 audio out per 1M tokens
GPT Realtime Translate	Streaming speech-to-speech translation for live multilingual audio experiences.	$0.034 per minute
GPT Realtime Whisper	Low-latency transcript deltas from live audio.	$0.017 per minute
GPT Live Transcribe	Low-latency live transcript deltas.	$0.017 per minute
GPT Transcribe	High-accuracy transcription for completed audio files, streamed file transcripts, and committed Realtime turns.	$0.0045 per minute
GPT Audio 1.5	Audio input and audio output in an existing Chat Completions application.	$32.00 audio in / $64.00 audio out per 1M tokens

From OpenAI’s model and pricing pages, checked 17 October 2026.

How do I use my prompt?

Choose the mode for your OpenAI voice or speech model.
Enter the role, task and relevant details, such as tools, languages or keywords.
Generate and copy the prompt or settings.
Paste session instructions or settings into the OpenAI voice model you use, GPT-Live backend instructions into your backend model, and the GPT Audio 1.5 request into your Chat Completions code.

Prompting tips for OpenAI voice models

For GPT-Live, keep the front-line conversation prompt concise and place detailed procedures and tool configuration in the backend. Prompting GPT-Live
Use the Live template’s Backchannel, Interruption, and Delegation policies, and list backend capabilities plus handoff conditions. Prompting GPT-Live
Do not guess a backend result: wait before quoting a price, confirming a booking, or saying an action is complete. Prompting GPT-Live
For Realtime, begin with a minimal prompt, evaluate representative conversations, and add directions only for observed failures. Prompting realtime models
Use short labeled prompt sections and include only those that matter for the voice product. Prompting realtime models
Start most production Realtime agents at low reasoning effort and raise it only when task complexity, latency tolerance, and failure cost justify it. Prompting realtime models
Use short preambles only when they help the caller understand that reasoning or a tool action is underway. Prompting realtime models

Worked examples

Receptionist (GPT-Live)

Before: Create a voice receptionist for [BUSINESS] that handles appointment bookings, changes and cancellations.

After (made with this generator):

You are the receptionist for [BUSINESS]. Speak warmly and briefly, at a calm pace. If a caller sounds frustrated, acknowledge it once and move to solving the problem. Speak English unless the caller asks to switch. If a name, date or number is unclear, ask the caller to repeat or spell it. Before you book, change or cancel anything, read the details back and wait for a clear yes. Never quote a price or confirm a booking until the backend returns it.

Backchannel policy
Use short acknowledgements such as "mm-hm" only while the caller is giving details.

Interruption policy
If the caller interrupts, stop speaking and answer the new point first.

Delegation policy
Backend tools: check_availability, book_appointment, cancel_appointment.
Delegate to the backend when: the caller asks about free times, wants to book, move or cancel, or asks a price.
Do not delegate to the backend when: the caller asks for opening hours or the address: [HOURS], [BUSINESS ADDRESS].

Delegate before giving an answer that depends on backend work. Do not guess the result while waiting.

GPT-Live 1: $0.05 per minute, billed per second, plus backend model tokens.

What the generator added: The prompt adds rules for tone, unclear details, acknowledgements, interruptions and backend delegation. It requires a readback and a clear yes before booking, changing or cancelling, and a backend result before quoting prices or confirming bookings.

Billing support agent (Realtime)

Before: Create a billing voice agent for [PRODUCT] that checks accounts, issues refunds and hands over difficult cases.

After (made with this generator):

# Role and Objective
You are the billing support agent for [PRODUCT]. Resolve billing questions on the call and hand over anything else.

# Reasoning
Use low reasoning effort to start. Raise it only for steps that fail at low.

# Preambles
Before a tool call, say one short sentence, such as "Let me check that."

# Verbosity
Answer in one or two short sentences.

# Tools
Tools: lookup_account, issue_refund. Confirm before any action that changes data, and wait for a clear yes.

# Unclear Audio
If a name, date or number is unclear, ask the caller to repeat it. Do not guess.

# Entity Capture
Read codes, reference numbers and email addresses back character by character.

# Escalation
Hand over to a person if the caller asks twice or reports a wrong charge over [AMOUNT].

Model gpt-realtime-2.1, reasoning effort low to start.

What the generator added: The prompt adds low initial reasoning effort, short replies, tool-call preambles and confirmation before data changes. It also adds unclear-audio checks, character-by-character readbacks and handover rules for repeated requests or wrong charges over [AMOUNT].

Frequently asked questions

Is the generator free, and does it run a voice model?

The generator is free and needs no sign-up. It runs in your browser and generates prompt text and settings only. It never calls a model. OpenAI model usage has separate costs.

What does it generate for GPT-Live 1?

It generates a short conversation prompt in OpenAI’s fixed template, plus separate backend instructions. The conversation prompt covers acknowledgements, interruptions and delegation. Detailed workflows go into the backend instructions.

Which Realtime models does the voice-agent mode support?

It supports GPT Realtime 2.1, GPT Realtime 2.1 Mini and GPT Realtime 2. It uses the twelve sections from OpenAI’s Realtime prompting guide. You fill only the sections you need.

What does transcription mode produce?

It produces a context prompt, keyword list and language list for GPT Transcribe, GPT Live Transcribe and GPT Realtime Whisper. For live transcription, it also adds the delay setting. The form removes angle brackets and line breaks from keywords.

Can I write instructions for GPT Realtime Translate?

No. OpenAI documents no instructions for this model; it acts as an interpreter. The generator outputs the output-language setting only. Each target language needs its own session.

Can the generator control tone in GPT Audio 1.5?

OpenAI documents no way to steer tone in GPT Audio 1.5. The generator builds a Chat Completions request with text and audio output, a voice, a format and a system message for content.