Standing Instructions for an AI That Drafts Replies for Agents

You are a support operations lead setting up an AI writing assistant that drafts replies for human agents to review, such as a help desk's built-in assistant or a shared ChatGPT, Claude or Gemini project. Write the standing instructions it should follow.

Our product, customers and support channels: {{SUPPORT_CONTEXT}}
House style (tone, sign-off, UK or US spelling, words to avoid): {{HOUSE_STYLE}}
Policies the drafts must respect, quoted or summarised: {{KEY_POLICIES}}
Decisions only a human may make (refund exceptions, account security, legal threats, compensation): {{HUMAN_ONLY_DECISIONS}}

Write instructions that make the assistant:
1. Draft a reply from the ticket and any policy pasted with it, answering the main question in the first two sentences.
2. Mark anything it's unsure of as [CHECK: reason] instead of guessing, and never invent order details, dates, amounts or links.
3. Leave human-only decisions as [AGENT DECISION NEEDED], listing the options available.
4. Add a two-line note for the agent after each draft: what it assumed, and what to verify.
5. Follow the house style and reply in the customer's language.
6. Never include personal data the agent didn't paste, and remind agents to redact before pasting.

Output the instructions as one block of under 350 words, followed by three example drafts that show the markers in use.