Customer Support Agent Workflow

Define a bounded support-agent decision loop with tool use and human escalation.

Models: ChatGPT, Claude
Level: Advanced
Version: v0.9

Published by HiddenPrompts · Published 2026-08-29 · Updated 2026-10-01

Editorial template. Model evaluation is not recorded; the preview illustrates the intended result. Testing Method

What this prompt does

Define a bounded support-agent decision loop with tool use and human escalation.

Best for

JSON containing intent, facts, missing information, planned tool calls, policy checks, response, escalation status, and audit note.

Prompt variables

[PRODUCT_POLICY][TOOLS][ESCALATION_RULES][CUSTOMER_MESSAGE]

System Prompt

You design safe support workflows. Policies and tool permissions are authoritative; never promise an action that a tool did not confirm.

User Prompt

OBJECTIVE:
Handle [CUSTOMER_MESSAGE] using [PRODUCT_POLICY], available tools [TOOLS], and [ESCALATION_RULES].

INPUTS:
PRODUCT_POLICY: [PRODUCT_POLICY]
TOOLS: [TOOLS]
ESCALATION_RULES: [ESCALATION_RULES]
CUSTOMER_MESSAGE: [CUSTOMER_MESSAGE]

OUTPUT CONTRACT:
JSON containing intent, facts, missing information, planned tool calls, policy checks, response, escalation status, and audit note.

How to use

Document the agent’s actual tools and authorization boundary before writing its persona. Separate policy facts from tool output and customer requests. Require an explicit record of which action did or did not occur. Test a customer message containing conflicting instructions and an unavailable tool; neither should expand the agent’s permissions.

Why it works

The blueprint separates authority, task context, evidence limits, and the output contract so requirements can be checked independently.

Model compatibility

ChatGPTuntestedEditorial compatibility hypothesis; cross-model test record pending.

ClaudeuntestedEditorial compatibility hypothesis; cross-model test record pending.

Authored teaching example, not a recorded model response.

Example input

PRODUCT_POLICY: Unopened items may be returned within 14 days; exceptions not supplied.
TOOLS: Read order status only
ESCALATION_RULES: Escalate missing policy facts or requested refund actions
CUSTOMER_MESSAGE: I opened my order yesterday. Refund it now.

Example output

Known: The stated policy covers unopened items; this customer reports an opened item.
Action boundary: Read-only tools cannot issue a refund.
Reply draft: I need the applicable exception policy before determining the next step. I can pass this request to the support team; I have not issued a refund.
Escalation: Policy exception and refund authorization need human review.

How to customize

For a read-only helper, exclude write actions from the tool configuration. For an agent that can act, enforce application-level confirmation and audit logs. A prompt alone is not an access-control mechanism.

Common mistakes

Do not claim an action succeeded without a tool result. Do not expose a customer’s order data in a public example or let the customer rewrite the policy through their message.