Answer these questions before writing a single line of prompt. These answers determine almost every design decision downstream.

​
Purpose

Write the agent’s goal in one sentence. If you cannot do this, the prompt will be confused. Example: “This agent calls customers who missed an EMI payment and either collects the amount or schedules a callback.”

​
Call Direction

OUTBOUND

You control when the call happens and why. Pre-load context (name, account, amount due). Identity verification is mandatory before sharing data.

INBOUND

The user controls timing — handle anything. Assume you know very little until the user speaks. Classify intent before doing anything else.

​
Language

  • What language does the user prefer to open in?
  • Does the agent need to switch mid-call if the user responds differently?
  • Is the register formal or conversational? (For Indic languages: jor gaaru? Female or male persona?)

​
Edge Cases — List the Top 5 Before You Start

Common edge cases:
  • Wrong person picked up
  • User is busy / asks to call back
  • User disputes the information (amount, date, record)
  • User becomes hostile or uses profanity
  • User speaks a different language than expected
Why do this first? Every common scenario that breaks the happy path should be listed before writing instructions. Gaps discovered after the prompt is written are harder to close cleanly.

​
Data Flow

WHENQUESTION TO ANSWER
Start of callWhat do you already know? (Name, account, due amount, ticket ID)
During the callWhat does the user tell you that gates a decision? (Intent, complaint location, reference number)
End of callWhat needs to go somewhere? (Ticket created, work order, analytics, summary)