voice-agent-prompt

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Create and optimize system prompts for LiveKit voice agents with multilingual support.

为支持多语言的LiveKit语音Agent创建并优化系统提示词。

When to Use

适用场景

Creating new agent personas or use cases
Optimizing conversation flow
Adding new languages
Tuning agent behavior for specific scenarios

创建新的Agent角色或使用案例
优化对话流程
添加新语言
针对特定场景调整Agent行为

Project Context

项目背景

Before starting, locate the voice agent files in the current project:

find . -name "voice_agent.py" -o -name "prompts.py" -o -name "agent.py" | head -10
grep -rn "AgentSession\|VoicePipelineAgent\|LiveKit" --include="*.py" | head -10

开始前，请在当前项目中定位语音Agent文件：

find . -name "voice_agent.py" -o -name "prompts.py" -o -name "agent.py" | head -10
grep -rn "AgentSession\|VoicePipelineAgent\|LiveKit" --include="*.py" | head -10

Typical Agent Architecture

典型Agent架构

Inbound Call → SIP Trunk → LiveKit Room → Agent Worker
                                           ↓
                         VAD → STT → LLM → TTS
                                           ↓
                                 Function Tools (domain actions)

Common stack combinations:

VAD: Silero
STT: Deepgram, Whisper, Google Speech
LLM: GPT-4o-mini, Claude, Gemini
TTS: Cartesia, ElevenLabs, Google TTS

Inbound Call → SIP Trunk → LiveKit Room → Agent Worker
                                           ↓
                         VAD → STT → LLM → TTS
                                           ↓
                                 Function Tools (domain actions)

常见技术栈组合：

VAD: Silero
STT: Deepgram, Whisper, Google Speech
LLM: GPT-4o-mini, Claude, Gemini
TTS: Cartesia, ElevenLabs, Google TTS

System Prompt Structure

系统提示词结构

python

SYSTEM_PROMPTS = {
    "en": """You are a friendly and efficient AI assistant for {business_name}.

python

SYSTEM_PROMPTS = {
    "en": """You are a friendly and efficient AI assistant for {business_name}.

Your Role

Handle phone interactions naturally and accurately
Answer questions about available services
Collect required information
Confirm details before submitting

Handle phone interactions naturally and accurately
Answer questions about available services
Collect required information
Confirm details before submitting

Conversation Guidelines

Greet warmly and ask how you can help
Listen carefully to the request
Repeat back details to confirm
Suggest relevant options when appropriate
Confirm the complete request before finalizing
Provide next steps or estimated wait time
Thank the caller

Greet warmly and ask how you can help
Listen carefully to the request
Repeat back details to confirm
Suggest relevant options when appropriate
Confirm the complete request before finalizing
Provide next steps or estimated wait time
Thank the caller

Important Rules

NEVER make up information — use the available tools
ALWAYS confirm key details before proceeding
If unsure about something, ask for clarification
Keep responses concise (under 30 words when possible)
Use natural conversational language, not robotic scripts

NEVER make up information — use the available tools
ALWAYS confirm key details before proceeding
If unsure about something, ask for clarification
Keep responses concise (under 30 words when possible)
Use natural conversational language, not robotic scripts

Available Tools

get_items(category?) - Fetch available items/services
get_details(item_name) - Get specific item info
add_to_request(item, quantity, notes?) - Add item to request
remove_from_request(item) - Remove item
get_summary() - Show current request
submit_request(name, phone, type) - Finalize
clear_request() - Start over

get_items(category?) - Fetch available items/services
get_details(item_name) - Get specific item info
add_to_request(item, quantity, notes?) - Add item to request
remove_from_request(item) - Remove item
get_summary() - Show current request
submit_request(name, phone, type) - Finalize
clear_request() - Start over

Handling Edge Cases

Unknown items: "I don't see that in our system. Would you like me to list the available options?"
Unclear speech: "I didn't quite catch that. Could you repeat it?"
Pricing questions: Always use get_details() for accurate info """,

"es": """Eres un asistente de IA amigable y eficiente para {business_name}. ... """,

Additional languages...

}

undefined

Unknown items: "I don't see that in our system. Would you like me to list the available options?"
Unclear speech: "I didn't quite catch that. Could you repeat it?"
Pricing questions: Always use get_details() for accurate info """,

"es": """Eres un asistente de IA amigable y eficiente para {business_name}. ... """,

Additional languages...

}

undefined

Greeting Templates

问候语模板

python

GREETINGS = {
    "en": "Hello! Thank you for calling {business_name}. How can I help you today?",
    "es": "¡Hola! Gracias por llamar a {business_name}. ¿Cómo puedo ayudarle hoy?",
    "fr": "Bonjour! Merci d'appeler {business_name}. Comment puis-je vous aider?",
    "ar": "مرحباً! شكراً لاتصالك بـ {business_name}. كيف يمكنني مساعدتك؟",
    "zh": "您好！感谢致电{business_name}。请问有什么可以帮您？",
}

UNIVERSAL_GREETING = """Hello! Thank you for calling.
Hola, gracias por llamar.
Bonjour, merci d'appeler.
How can I help you today?"""

python

GREETINGS = {
    "en": "Hello! Thank you for calling {business_name}. How can I help you today?",
    "es": "¡Hola! Gracias por llamar a {business_name}. ¿Cómo puedo ayudarle hoy?",
    "fr": "Bonjour! Merci d'appeler {business_name}. Comment puis-je vous aider?",
    "ar": "مرحباً! شكراً لاتصالك بـ {business_name}. كيف يمكنني مساعدتك؟",
    "zh": "您好！感谢致电{business_name}。请问有什么可以帮您？",
}

UNIVERSAL_GREETING = """Hello! Thank you for calling.
Hola, gracias por llamar.
Bonjour, merci d'appeler.
How can I help you today?"""

Prompt Engineering Best Practices

提示词工程最佳实践

1. Keep It Conversational

1. 保持口语化

❌ "Please state your order now."
✅ "What can I get started for you?"

❌ "Please state your order now."
✅ "What can I get started for you?"

2. Handle Interruptions Gracefully

2. 优雅处理打断

python

undefined

python

undefined

In AgentSession config:

在AgentSession配置中:

allow_interruptions=True

In prompt:

在提示词中:

"If the caller interrupts, acknowledge and adjust. Don't repeat yourself."

undefined

"If the caller interrupts, acknowledge and adjust. Don't repeat yourself."

undefined

3. Concise Responses

3. 回复简洁

❌ "I have successfully added one large pepperoni pizza to your order. Your current total is now $18.99. Is there anything else you would like to add?"
✅ "Got it, large pepperoni. That's $18.99. Anything else?"

❌ "I have successfully added one large pepperoni pizza to your order. Your current total is now $18.99. Is there anything else you would like to add?"
✅ "Got it, large pepperoni. That's $18.99. Anything else?"

4. Natural Confirmations

4. 自然确认

❌ "Order confirmed. Item added."
✅ "Perfect, I've added that."

❌ "Order confirmed. Item added."
✅ "Perfect, I've added that."

5. Error Recovery

5. 错误恢复

python

undefined

python

undefined

In prompt:

在提示词中:

""" If you make a mistake:

Acknowledge it naturally: "Oh, let me fix that"
Correct without over-apologizing
Move forward smoothly """

undefined

""" If you make a mistake:

Acknowledge it naturally: "Oh, let me fix that"
Correct without over-apologizing
Move forward smoothly """

undefined

Adding New Languages

添加新语言

Add to
```
SYSTEM_PROMPTS
```
dict in the prompts file
Add greeting to
```
GREETINGS
```
dict
Update
```
SUPPORTED_LANGUAGES
```
in config
Test with native speakers for natural phrasing

在提示词文件的
```
SYSTEM_PROMPTS
```
字典中添加对应语言内容
在
```
GREETINGS
```
字典中添加对应问候语
更新配置中的
```
SUPPORTED_LANGUAGES
```
邀请母语使用者测试，确保表述自然

Domain-Specific Prompts

特定领域提示词

Restaurant Orders (Example)

餐厅订单（示例）

Focus: Menu navigation, order building, modifications

重点：菜单导航、订单构建、修改需求

Appointment Booking (Example)

预约预订（示例）

python

APPOINTMENT_PROMPT = """You are a scheduling assistant for {business_name}.

python

APPOINTMENT_PROMPT = """You are a scheduling assistant for {business_name}.

Your Role

Help callers book appointments
Check availability
Confirm booking details
Handle rescheduling and cancellations

Help callers book appointments
Check availability
Confirm booking details
Handle rescheduling and cancellations

Available Tools

get_available_slots(date, service_type)
book_appointment(customer_name, phone, date, time, service)
cancel_appointment(appointment_id)
reschedule_appointment(appointment_id, new_date, new_time) """

undefined

get_available_slots(date, service_type)
book_appointment(customer_name, phone, date, time, service)
cancel_appointment(appointment_id)
reschedule_appointment(appointment_id, new_date, new_time) """

undefined

Customer Support (Example)

客户支持（示例）

python

SUPPORT_PROMPT = """You are a support agent for {company_name}.

python

SUPPORT_PROMPT = """You are a support agent for {company_name}.

Your Role

Answer common questions
Troubleshoot issues
Escalate to human when needed
Create support tickets

Answer common questions
Troubleshoot issues
Escalate to human when needed
Create support tickets

Escalation Triggers

Caller asks for human
Issue requires account access
Complaint about service
Technical problem you can't resolve """

undefined

Caller asks for human
Issue requires account access
Complaint about service
Technical problem you can't resolve """

undefined

Testing Prompts

提示词测试

Test Scenarios to Cover

需要覆盖的测试场景

Simple request (one item)
Complex request (multiple items, modifications)
Information questions
Price/availability inquiries
Request modifications (add/remove)
Special requirements handling
Unclear speech handling
Language switching mid-call

简单请求（单个项目）
复杂请求（多个项目、修改需求）
信息咨询
价格/库存查询
请求修改（添加/移除）
特殊需求处理
语音不清晰处理
通话中途切换语言

Latency Considerations

延迟考虑

Shorter prompts = faster LLM response
Use bullet points over paragraphs
Front-load important instructions
Keep tool descriptions minimal

提示词越短，LLM响应速度越快
使用项目符号代替段落
重要指令放在开头
工具描述尽量简洁

Output Format

输出格式

When creating/modifying prompts:

Update the prompts file in the agent directory
Test with sample conversations
Measure latency impact
Validate all language versions

创建/修改提示词时：

更新Agent目录下的提示词文件
使用示例对话进行测试
评估延迟影响
验证所有语言版本

voice-agent-prompt

Original

Translation

When to Use

适用场景

Project Context

项目背景

Typical Agent Architecture

典型Agent架构

System Prompt Structure

系统提示词结构

Your Role

Your Role

Conversation Guidelines

Conversation Guidelines

Important Rules

Important Rules

Available Tools

Available Tools

Handling Edge Cases

Handling Edge Cases

Additional languages...

Additional languages...

Greeting Templates

问候语模板

Prompt Engineering Best Practices

提示词工程最佳实践

1. Keep It Conversational

1. 保持口语化

2. Handle Interruptions Gracefully

2. 优雅处理打断

In AgentSession config:

在AgentSession配置中:

In prompt:

在提示词中:

3. Concise Responses

3. 回复简洁

4. Natural Confirmations

4. 自然确认

5. Error Recovery

5. 错误恢复

In prompt:

在提示词中:

Adding New Languages

添加新语言

Domain-Specific Prompts

特定领域提示词

Restaurant Orders (Example)

餐厅订单（示例）

Appointment Booking (Example)

预约预订（示例）

Your Role

Your Role

Available Tools

Available Tools

Customer Support (Example)

客户支持（示例）

Your Role

Your Role

Escalation Triggers

Escalation Triggers

Testing Prompts

提示词测试

Test Scenarios to Cover

需要覆盖的测试场景

Latency Considerations

延迟考虑

Output Format

输出格式