How to Write a Good Prompt

• Gather all the important details about the client project first, including the main role and goal of the agent.
• Design the overall structure and workflow of the agent before writing. A well-planned structure is the base of a good prompt.
• Build that structure using clear sections and subsections, like Section A with A.1 and A.2, so related rules stay together and the prompt is easy to read.
• Number and index everything, so every rule has a fixed address you can point to instead of writing it again.
• Separate parts with lines and dashes, so it is easy to see where one section ends and the next begins.
• Use fewer words, and say each thing only once. The prompt is sent on every turn, so extra words cost tokens, money, and speed.
• Never repeat the same rule in different places. Copies slowly start to differ and clash, and they waste tokens.
• Never write two rules that contradict each other. The agent will pick one of them, and this is the number one cause of bugs.
• Write clear, literal instructions — like “ask for the 6-digit pincode” instead of a vague line like “handle location.”
• Do not use arrow symbols, because the model does not read them well. Use plain words like “then” instead.
• Do not put any maths or logic in the prompt. During the call the agent reads it as plain text and cannot calculate, so work out values like dates beforehand and give them as ready facts.
• Cover the full call flow with enough detail for every path — wants to sell, does not want to sell, wants to buy, wants a callback, and ending the call.
• Clearly tell the agent when the right action is to do nothing, otherwise it will invent a wrong action to fill the gap.
• Train it like a new human agent. Walk it through the call step by step, the way you would teach a new person on their first day.
• Test the prompt with real call cases after writing it, and fix any step where more than one action is possible.

​

Get help from AI FDE