ع
Start Topics Teams Reference What's new Saved
playbook

Draft a reply that sounds like a person

Turn a frustrated customer message into a warm, human reply that owns the mistake, gives one clear next step, and never invents details only you can confirm.

easy ~15 min
when to reach for this

A customer is upset and you have to reply in five minutes without sounding like a form letter. The temptation is to reach for corporate filler — "we sincerely apologize for any inconvenience" — which reads as a wall the customer has to climb. This system gets you a reply that owns the mistake plainly, gives one clear next step, and leaves [placeholders] for the private details only you can verify — so you ship a human reply fast without ever letting Claude guess an order number or a refund.

gather this first
  • The customer's message — paste it straight into the chat in Claude Desktop, or drop it in as ticket.txt. Strip their full name, email, and account number first; you'll fill those back in yourself.
  • Your voice in 3 lines: how you say sorry, how formal you are, and 2 phrases you'd never use ("per our policy," "as a one-time courtesy").
  • What you can actually offer here — refund, replacement, store credit — and what's off the table, so Claude never promises beyond it.
the workflow

  1. Read the message back before drafting

    Don't ask for a reply yet. Make Claude prove it understood the upset and the ask — that catches a misread before it becomes a tone-deaf reply.

    you ask
    Read ticket.txt. In 3 bullets, tell me: what actually went wrong for this customer, how upset they sound on a scale of mildly annoyed to furious, and the one thing they most want from us. Don't write a reply yet.

    what you get back A short read-back — "order arrived with a cracked screen; furious, second time this has happened; wants a replacement shipped today, not a refund." If that's wrong, you fix it now, not after a draft sets the wrong tone.

    This "read it back to me first" move is the single habit that keeps replies from missing the real complaint.

  2. Draft the reply in your voice, with placeholders

    Now ask for the reply — but pin the voice and force placeholders so Claude can't fill in private details it has no way of knowing.

    you ask
    Now draft a reply in my voice: warm and plain, no corporate jargon, no "we apologize for any inconvenience." Own the mistake in the first sentence, give one clear next step, and keep it under 120 words. Use [name], [order number], and [what we're doing] as placeholders. Do not invent any order number, refund amount, ship date, or policy.

    what you get back A short, human reply — "Hi [name], you're right to be frustrated, and I'm sorry — a cracked screen twice is not okay. Here's what I'm doing: [what we're doing]..." — with every private detail left as a bracket for you to fill.

  3. Pressure-test the tone and the promise

    Have Claude check its own draft against the two ways a reply fails: sounding robotic, or promising something you didn't authorize. Cheaper to catch here than in the customer's inbox.

    you ask
    Read your draft back critically. Does any line sound like a form letter? Did you promise anything I didn't tell you we offer? Flag both, and give me one alternative for the weakest sentence.

    what you get back A short critique — "line 3 reads canned; the phrase 'rest assured' is filler" — plus a flag if it slipped in a promise, and a warmer rewrite of the flat line.

  4. Fill the placeholders and send

    This is the human step that can't be skipped. You paste the real details from your own approved system, give it one read aloud, and own the send.

    you ask
    I've filled in [name], [order number], and the offer is a replacement shipping tomorrow with a prepaid return label. Slot those in and read me the final version exactly as it'll send.

    what you get back The finished reply with your real details in place — ready for one last read-aloud before you, not Claude, press send.

    Read it aloud once. If it sounds like something you'd actually say to the person, it's ready.

make it your own
  • Build the repeatable version: when the same situation keeps coming up, graduate the best of these into the Build a canned-response library you'll actually reuse playbook so you start from a template instead of a blank box.
  • Save the move (Power Track): once your drafting prompt is dialed in, store it as a /reply custom command (see the Playbook's Capabilities tab) so a fresh draft is one command, not three prompts. Custom commands are an opt-in Power-Track step — on Desktop you can keep the dialed-in prompt saved in a note and paste it instead.
  • Tier the apology: keep a short note on how you respond to mildly-annoyed vs. furious so the warmth scales with the upset instead of being one flat register.
watch out for
  • Never let Claude state an order number, refund amount, ship date, or policy — those are [placeholders] you fill from your own system. A confident wrong detail to a customer is worse than a slow reply.
  • Strip the customer's full name, email, card, and account number before pasting — model good [placeholder] hygiene and put the real details back in your approved tool, not the chat.
  • Claude drafts the shape and the tone; a human owns the send. The apology has to be yours, and only you can confirm what you're actually offering.

you'll end up with A warm, on-voice reply with the mistake owned and one clear next step — placeholders for everything private, and nothing promised that you didn't authorize — in about the time it takes to read the ticket twice.

Questions people ask

How do I stop Claude from inventing an order number or refund amount?
Make it use `[placeholders]` for every private detail and tell it explicitly not to invent an order number, refund, ship date, or policy. You fill those in from your own approved system — a confident wrong detail to a customer is worse than a slightly slower reply.
What should I strip from the ticket before pasting it in?
Remove the customer's full name, email, card, and account number first, replacing them with placeholders. You model good hygiene and put the real details back in your approved tool, not the chat.
How do I keep the reply from sounding like a form letter?
Give Claude your voice in three lines and ban corporate filler like "we apologize for any inconvenience," then have it own the mistake in the first sentence and critique its own draft for canned lines. Read the final version aloud — if it sounds like something you'd actually say, it's ready.
Does Claude send the reply?
No. Claude drafts the shape and the tone, but a human fills the placeholders, confirms what you're actually offering, and presses send. The apology has to be yours.