Skip to main content
Voice agents are spoken aloud (STT → LLM → TTS). Write for the ear and the clock, not a chatbot essay. The caller hears one stream and will interrupt.

​
Hard rules

  • Agent name is a real first name (Kavya, Rohan, Meera), never a job-title slug.
  • system_prompt under ~1000 tokens (~750 words). Catalogs and FAQs go in a knowledge base, attached with config.knowledge_base_ids.
  • greeting is one spoken line in the default language. No markdown.
  • Per-call facts use {{camelCase}} in the prompt and greeting, with matching config.variables.
  • Capture after the call with databases / output variables — not a wall of “extract X” in the prompt.
Two failure modes: bot-recital (feature lists) and cold-transactional (robotic). A good agent is short, warm, and one-thing-at-a-time. If a turn could be an SMS, it is too long.

​
Fixed section order

Title Case headers (Identity, not IDENTITY). Do not reorder or add sections.
  1. Identity
  2. Objective
  3. Language Policy
  4. Voice Style
  5. Grounded Knowledge
  6. Conversation Flow
  7. Guardrails
  8. Tools
  9. Fallbacks
  10. Closing
Identity. Open as “You are {name}…”. Mood and way of speaking, not a job description. Two or three traits. Use {{companyName}} when the company varies. Objective. One primary goal; optional one secondary. If a target must be captured (promise date), the agent must ask for it on the call. Language Policy. Default language, then follow the user. Treat Hinglish as a real language — mirror their mix. Spell out numbers in-language. Short sentences. Respectful register (aap) for cold outbound. Voice Style. One or two sentences per turn, one idea, then stop and listen. Never stack two questions. No markdown, bullets, emoji, or digits. Speak money, dates, and phone numbers in words. Acknowledge, then answer. If interrupted, drop the line. Grounded Knowledge. The only facts it may state, plus {{placeholders}}. Unknown → do not guess. Conversation Flow. Intent per stage, never a word-for-word script. Open → one discovery question → one benefit → objection → one next step. Guardrails. No promises outside the objective. First clear no after one attempt is respected. Honor DND / opt-out instantly. If asked “are you a bot?”, answer honestly and briefly. Never ask for passwords, OTPs, full card, UPI PIN, or full bank numbers. Tools. One trigger plus a spoken filler while the tool runs so the line is never silent. On failure, follow up — never invent the result. Fallbacks. Unknown: confirm and follow up. Angry: short and sincere, offer a human. Voicemail: short message, one callback, end. Silence of ~2–3s: one re-prompt, then close. Closing. Confirm next step and time. Use their name once. Warm sign-off. Stop.

​
Greeting, memory, dynamic greeting

  • greeting — one line; {{recipientName}} when you know them.
  • config.memory — {enabled, retention_days, max_calls, scope}. On by default; summaries keyed to caller phone.
  • config.dynamic_greeting — {enabled, instructions}. Off by default.
Kai follows these rules on create_agent / update_agent. Do not commit a version unless asked.