damngoodprompts
Menu

Customer support reply drafting

Draft replies that are grounded in policy, account facts, and the agent's actual authority.

Last reviewed: August 10, 2026ChatGPTClaudeGemini
Technique

Best suited to Claude - strong fit for careful customer-facing tone and separating draft copy from internal verification notes

The task

You need a customer reply draft. The model can help with clarity and tone, but it should not invent policy, promise refunds, or decide account actions.

Why the old approach is outdated

“Write a friendly support reply” optimizes for tone before truth. That is backwards. Support drafts need grounding first, then voice.

The current approach

Draft a support reply using only the ticket, account facts, and policy excerpt below.

Do not promise refunds, credits, cancellations, or timelines unless the policy excerpt
explicitly supports them. If a fact is missing, put it in "agent verification needed."

Return:
1. Customer reply draft.
2. Agent verification needed.
3. Policy sentence used.

If the same support workflow also needs categories, urgency, or CRM fields, split that into a structured data extraction pass instead of hiding it inside the reply draft.

When to use a different model

Use ChatGPT when the reply is part of a structured support workflow with categories, urgency, and CRM fields. Use Gemini when the support evidence includes screenshots or attached documents.

What to avoid

  • Asking the model to “make it right” without policy boundaries.
  • Letting apology language imply fault.
  • Mixing internal notes into the customer-facing reply.
  • Sending private customer data to a model surface that is not approved for that use.

Source set