Guide
How to give your support team one reply tone with AI
A support team's AI replies sound like one company when every agent's AI uses the same short tone guide and the same three real replies, instead of each agent's own prompt. Write the guide as do-and-don't pairs, pick three replies that represent your best, and share them as a skill or shared instructions every agent's AI reads. A related finding: in the NBER study of support agents given a generative AI assistant that suggested replies in real time, issues resolved per hour rose 14% on average and 34% for novice and low-skilled agents. The study measured that assistant, not a tone guide, but it shows what sharing the best agents' way of answering can do.
Short answer
- One tone guide: five or six do-and-don't pairs.
- Three real replies that show it.
- Shared, so every agent's AI reads the same version.
- Reviewed weekly: fix the guide, not the agents.
Write the tone guide
| Do | Don't |
|---|---|
| Answer the question in the first sentence | Start with "Thank you for reaching out" |
| "We got this wrong, and here's the fix" | "Unfortunately, due to a system issue…" |
| Say when they'll hear back, with a day | "As soon as possible" |
| One way to reach a person | Three links to help articles |
| Sign with a first name | "The Support Team" |
An example guide; write your own from your best replies.
A worked example
Illustration: a made-up business. A 7-person support team at a meal-kit company answers about 1,200 emails a week. Each agent uses ChatGPT Business with their own prompt; customers get seven different tones. The team lead writes a six-line guide, picks three replies that earned thank-you notes, and shares them as one skill in the workspace.
A ticket arrives: "My box came without the chicken for Tuesday's recipe. Again." The draft from the shared skill:
Draft reply
Hi Ana, we left the chicken out of your box, and this is the second time, which isn't good enough. I've refunded Tuesday's recipe in full and added a credit for next week's box. I've also flagged your address to the packing team so your next box is checked by hand. If anything's missing again, reply here and it comes straight to me. Leo
Leo checks two things before sending: the refund and credit actually went through, and "checked by hand" is a promise packing can keep. It is; he sends it.
Roll it out
- Collect ten replies customers thanked you for; pick the three that sound most like you.
- Write the guide as five or six do-and-don't pairs, plus banned words and your sign-off.
- Share it once, as a skill or shared instructions every agent's AI reads; retire personal prompts.
- Keep checks with the agent: refunds, dates and promises are confirmed before sending.
- Review ten replies a week and change the guide when the drafts drift.
Next step
Once the reply skill works, the next ones follow the same pattern: an FAQ from tickets, an incident notice. If your team builds them in Claude Code, see how to create Claude skills in Claude Code; for the rollout beyond support, getting your team to use AI the same way.
Skills for this
- Customer successCustomer reply in our toneDrafts a support reply in your team's tone: answers the question first, owns mistakes, drops jargon, and ends with one clear next step.
- Customer successFAQ from support ticketsTurns a batch of tickets or chats into an FAQ: questions in customer words, ranked by count, short answers, and flags for answers to confirm.
- Customer successIncident notice for customersWrites customer-facing incident notices by stage, from investigating to resolved, and a postmortem summary with a plain apology and what changes.
- Customer successSupport ticket triageTriages incoming tickets into a table with category, S1 to S4 severity, routing, and a first reply for each, sorted by what to answer first.
Questions
How do I get consistent AI replies from a support team?
Give every agent's AI the same short tone guide and three real replies that represent your best, as a shared skill or shared project instructions, instead of each agent's own prompt. Review a sample of replies weekly and fix the guide, not the agents.
What goes in a support tone guide?
Five or six do-and-don't pairs (answer first; own mistakes without blaming; no 'per my last email'), words you never use, how you sign off, and three real replies that show it.
Can customer details go into the AI?
Only in a company account whose plan doesn't train on your data by default, and only what the reply needs. Keep payment details out entirely.
AI skill
Have your AI answer customers in your team's tone
Put this skill in AI Skill Pod and everyone on your team gets it in their Claude, Cursor or ChatGPT (Pro, Business, Enterprise): one invite link, free to start. Customer reply in our tone has your AI draft each support reply in your team's tone from a short guide or three past replies: the answer first, mistakes owned without blame, no filler.
14 days free · No card needed
Sources
- Brynjolfsson, Li and Raymond: Generative AI at Work (NBER Working Paper 31161) — agents given a generative AI conversational assistant resolved 14% more issues an hour on average, 34% for novice and low-skilled agents (read 2026-11-05)
- Claude Help Center: Using skills in Claude — Team and Enterprise owners deploy skills to the organization (read 2026-11-05)
- OpenAI Help: Skills in ChatGPT — shared in Business, Enterprise, Healthcare and Edu workspaces (read 2026-11-05)
- ChatGPT pricing — Business: no training on business data by default (read 2026-11-05)