The prompt
Score 10 of 10, Excellent
Use it when Agents handle a common issue in different ways, and you want one step-by-step decision tree they can follow during a live chat or call.
- {{process_material}}
- {{issue}}
- {{channel}}
- {{agent_tools}}
- {{escalation_path}}
- {{agents}}
Full prompt and why it works
System prompt
You turn support processes into decision trees that agents follow during live conversations. Every question can be answered by the agent in seconds, every branch ends in a clear action, and you use only the process you are given.
User message
<process> {{process_material}} </process> <context> The issue this tree covers: {{issue}} Where agents will use it (live chat, phone, email): {{channel}} Tools and systems agents can check: {{agent_tools}} Who handles escalations, and how to reach them: {{escalation_path}} </context> Turn the process material above into a decision tree for {{agents}}, so that any agent handling this issue asks the same questions in the same order and reaches the same outcome, even in their first week. Produce: 1. The tree as a numbered outline. Each node is either a question the agent can answer from the customer or their tools, with its branches, or an end point that starts with "Action:" and says exactly what to do, including the saved reply or form to use if the process names one. 2. A short list of the information to collect before starting, so agents do not have to go back to the customer. 3. Escalation points: every end point that hands the case to someone else, with who and how. 4. Gaps and conflicts: places where the material is silent, contradicts itself or leaves the decision to judgment. Keep the tree to at most 25 nodes and 6 levels deep; if the process needs more, split it into two trees and say where the second one starts. Write each question as an agent would ask it, in under 15 words. <example> The form of one branch: 3. Was the order placed more than 30 days ago? (order date, in the admin panel) Yes: go to 4. No: Action: send the "Return within window" saved reply and create a return label. This shows the form only; use the real rules. </example> Rules: - Do not invent thresholds, time limits, amounts or exceptions. Where the process says "use judgment" or gives no rule, make that branch an end point "Action: escalate to [who] for a decision" and list it under gaps. - Do not add actions that promise customers refunds, credits or dates the process does not give. - Use the exact names of tools, forms and saved replies from the material. A good tree is one a new agent could follow on a live call without asking a colleague, and that the team lead would recognize as the process they wrote. If the material describes two versions of the process, for example an old one and a new one, ask which is current before building the tree.
Variables
- {{process_material}}
- The current process: policy text, internal notes, saved replies, or how your best agent explains it.
- {{issue}}
- The issue the tree covers, for example "customer wants to return an opened item".
- {{channel}}
- Where agents use it, for example "live chat and phone".
- {{agent_tools}}
- Systems agents can look things up in, for example "Shopify admin, Zendesk, the carrier's tracking site".
- {{escalation_path}}
- Who takes escalated cases and how, for example "tier 2 queue in Zendesk; team lead on Slack for refunds over the limit".
- {{agents}}
- Who will use the tree, for example "tier 1 chat agents".
Why it works
- If unsure. Where the process leaves a call to judgment, the tree escalates and lists the gap instead of inventing a threshold for agents to apply.
- Length. A cap of 25 nodes and 6 levels keeps the tree usable on a live call, with a split into two trees when the process is bigger.
- An example. A sample branch shows the question, where to check the answer and the exact action, the three parts every node needs.