Best Practices for Tabbly Voice AI Agents

​

Crafting Effective Prompts

​

Core Prompting Principles

• Define exactly what your agent should accomplish
• Use simple, direct language in your instructions
• Avoid ambiguous terms or vague objectives
• Start with a concise prompt and add details iteratively

• Specify the agent’s role clearly (e.g., “You are a customer support representative”)
• Define what topics the agent should and shouldn’t discuss
• Set expectations for conversation length and depth
• Example: “Keep responses under 2 sentences unless asked for more detail”

• Include relevant background about your business
• Explain industry-specific terminology
• Define your brand voice and tone
• Share common customer scenarios

​

Prompt Structure Template

Role: You are a [specific role] for [company name]

Objective: Your goal is to [clear objective]

Constraints:
- Keep responses concise and under [X] sentences
- Always [required behavior]
- Never [prohibited behavior]

Key Information:
[Relevant business details, policies, or data]

Conversation Flow:
1. [Opening approach]
2. [Main interaction guidelines]
3. [Closing procedure]

​

Critical Voice-Specific Considerations

​

Numbers and Formatting

• Wrong: “Your order number is 12345”
• Correct: “Your order number is one two three four five”
• Wrong: “The total is $91.50”
• Correct: “The total is ninety-one dollars and fifty cents”

• Phone numbers: “nine one four, two three five, six seven eight nine”
• Prices: “twenty-nine dollars and ninety-nine cents”
• Dates: “March fifteenth, twenty twenty-five”
• Times: “three thirty PM” or “fifteen hundred hours”
• Percentages: “fifteen percent” not “15%“

​

Multi-Language and Script Considerations

• Write Hindi words in Devanagari script: “आपका नाम क्या है?”
• Don’t use romanized Hindi in prompts: Avoid “aapka naam kya hai”
• Mix naturally: “Hello, आपकी मदद के लिए मैं यहाँ हूं”
• Ensure your TTS voice supports Devanagari rendering

• Arabic: Use Arabic script (العربية) not transliteration
• Chinese: Use appropriate characters (中文) not pinyin
• Korean: Use Hangul (한국어) not romanization
• Japanese: Use proper mix of Hiragana, Katakana, Kanji

• Keep each language in its native script throughout
• Don’t switch scripts mid-sentence for the same language
• Test pronunciation with your chosen TTS voice
• Provide pronunciation guides for brand names or unique terms

​

Pronunciation and Clarity

• If you want it pronounced as letters: “F A Q” or “U S A”
• If it’s a word: “FAQ” (pronounced “fak”) or “NASA”
• For clarity, add pronunciation hints: “FAQ (F-A-Q)” in your knowledge base

• Add phonetic spellings for difficult words
• Example: “Tabbly (TAB-lee)” in your training data
• Create a pronunciation dictionary for industry jargon
• Test how TTS handles technical vocabulary

• Commas create brief pauses
• Periods create longer pauses
• Question marks raise intonation at the end
• Exclamation marks add emphasis (use sparingly)
• Use ellipses (…) for thoughtful pauses

​

Conversation Design Best Practices

​

Keep It Natural

• “Let me connect you with our billing team”

• Occasional: “Hmm, let me check that for you”
• Don’t overuse: “Um,” “uh,” “like” in every response

• Program responses for when users cut in
• Example: “Oh, sorry—go ahead” or “Yes, I’m listening”

​

Context and Memory

• “As you mentioned earlier about…”
• “Based on what you told me…”
• This shows the agent is truly listening

• Track what information has been collected
• Don’t ask for the same details twice
• Build on previous answers logically

​

Error Handling and Fallbacks

• ✅ “I didn’t quite catch that. Could you rephrase?”
• ✅ “Just to clarify, are you asking about [interpretation]?”

• Define clear triggers for human handoff
• Example: “If user says ‘speak to a person’ more than once, transfer immediately”
• Set frustration detection: “I can tell this is urgent. Let me connect you with someone who can help right away”

​

Prompt Testing and Refinement

​

Iterative Improvement Process

• Begin with a basic prompt covering core functionality
• Test thoroughly in the playground
• Identify gaps and confusion points

• Layer in additional instructions one at a time
• Test after each addition
• Remove anything that doesn’t improve performance

• Unusual requests
• Ambiguous phrasing
• Multiple questions at once
• Background noise scenarios
• Various accents and speaking speeds

• Review call transcripts regularly
• Identify common misunderstandings
• Update prompts based on actual user behavior
• Track improvement metrics over time

​

Common Pitfall Avoidance

• Too many instructions confuse the agent
• Prioritize the most important behaviors
• Use knowledge base for detailed information

• “Be brief but comprehensive” creates confusion
• Choose one priority: brevity OR detail
• Make trade-offs explicit

• Explain context the agent can’t infer
• Provide definitions for industry terms
• Include relevant background information

​

Language-Specific Prompting Tips

​

English Variants

• Use American spelling and terminology
• “elevator” not “lift,” “apartment” not “flat”
• Date format: “March 15th, 2025”

• British spelling and phrasing
• “lift” not “elevator,” “mobile” not “cell phone”
• Date format: “15th March 2025”

• May include local terms: “lakh” (hundred thousand), “crore” (ten million)
• Write these clearly: “two lakh” or “two hundred thousand”

​

Non-English Languages

• Spanish: Choose between “tú” (informal) and “usted” (formal)
• Japanese: Define keigo (politeness) level
• German: Specify “du” vs “Sie”
• Korean: Set appropriate honorific level

• Spanish: Spain vs Latin America differences
• Portuguese: Brazil vs Portugal variants
• Arabic: MSA (Modern Standard) vs dialect
• Chinese: Simplified vs Traditional characters

​