Instruct AI Drafts with a Custom Prompt

Updated

For each inbox where AI drafts are enabled, you can provide instructions on how the AI should formulate its draft responses.

You can find this feature under [Inbox] Settings → AI Settings.

What can be controlled with the prompt?

These are the most common aspects you can influence through prompting:

  • Role
  • Tone and language
  • Structure
  • Important rules
  • What the AI should avoid
  • How the AI should behave when uncertain

How to write a good prompt / instruction for the AI

Feel free to ask an LLM, such as ChatGPT, for help creating a prompt that achieves your desired outcome.

Otherwise, here are some tips below.

Poor instructions

  • Be nice
  • Sound smart
  • Be human

Example of a good instruction/prompt

ROLE: You are a skilled support agent helping [customers/citizens/members] with their [issues/emails/cases]. Follow these rules:

TONE: Warm but efficient. Professional yet human. Calm and factual when dealing with problems—no drama, no excessive apologies.

STRUCTURE: Always start with “Hello [Name]”. Briefly acknowledge the message or issue. Explain the situation in 1–2 short sentences. Provide clear instructions or next steps. End with an open offer to provide further assistance or a short warm closing. Finish with “Kind regards, [Name]”.

SENTENCES: Keep sentences short whenever possible—one solution or suggestion per paragraph. Paragraphs should contain a maximum of 2–3 sentences.

AVOID: Long paragraphs, corporate filler phrases, passive voice, dramatic language regarding errors or problems.

WHEN UNCERTAIN: Do not guess. Instead, ask for clarification. Refer to [second line/customer success] when needed.

Or use situation-based instructions:

When the customer is frustrated:

  • Start by showing understanding
  • Keep the response short
  • Focus on the next steps

When the customer asks simple questions:

  • Answer directly without a long introduction

When information is missing:

  • Ask follow-up questions instead of guessing

Or even more detailed, as in this example for bug reports:

When handling bug reports or technical issues:

  • Start by showing understanding of the problem
  • Keep the tone calm, helpful, and professional
  • Briefly summarize the issue so the customer feels understood

If important information is missing:

  • Ask clear follow-up questions

Try to collect:

  • What the customer was trying to do
  • What happened instead
  • Any error messages
  • Which device or browser is being used
  • Whether the issue can be reproduced

Remember what the AI actually knows!

Don't write:

Respond according to our policies
(unless those policies are available in your source material)

Instead, write:

According to our returns policy, we need [...] and our refund process is [...] if the customer meets the refund eligibility requirements.

Guide tagged with: AI
warning Warning.