Three people reply to comments on your behalf. One is chirpy, one is corporate, one uses the word "folks". A customer reading the thread meets three different companies.
"Sound like us" is the instruction, and it's unfollowable. Everyone thinks they're doing it.
The fix is to write the voice down in a form that can actually be checked. The useful discovery is that it takes about six fields, not a brand book.
The six fields
Everything a person or a model needs to reply as you, without reading three years of your posts.
1 · Tone: three or four words, not ten
"Warm, direct, a bit dry." Stop there.
Ten adjectives cancel out to nothing, exactly like they do in a creative brief. If your list contains both "playful" and "authoritative", you haven't described a voice, you've described a committee. Pick the three that would still be true on a bad day.
2 · Audience: who you're talking to
One or two sentences about a real person. "Small manufacturers who've been burned by a cheap supplier and are now suspicious of anything that sounds like marketing."
This one field does more work than tone. Voice is largely a function of who you believe you're addressing: the same sentence reads differently written at a sceptic than at an enthusiast.
3 · Do: the specific moves
Not virtues. Moves.
- "Answer the question in the first sentence, then add context."
- "Use the customer's own words back."
- "Admit it when we don't know."
"Be helpful" is not a do. It's a hope. A do is something you could observe someone failing to follow.
4 · Don't (and be specific)
This is where most of the value is, because it's where the accidents happen.
- "Never say 'we're sorry you feel that way'."
- "No exclamation marks in complaint replies."
- "Don't use 'guys' or 'folks'."
- "Never promise a delivery date."
Every good don't-list is a scar. Write down the thing that went wrong last time.
5 · Signature phrases: the small handful
Things you actually say. Not a tagline. The recurring phrasing that makes your writing recognisable.
Keep it short. A long list turns a voice into a tic, and readers notice repetition faster than they notice consistency.
6 · Sign-off (if you have one)
Whether replies are signed, and how. "Sam" changes the temperature of a reply more than most of the tone words above it. Decide once.
Voice under pressure
The interesting cases are the hard ones, and a document that only covers the easy ones is decoration.
Complaints. Most brand voices quietly collapse here into corporate defensiveness. Decide in advance: do you apologise plainly? Do you explain? Do you take it to DM, and after how many exchanges? Comment triage covers which complaints need this at all. But the ones that do are where a voice is actually visible.
Praise. The failure mode is being embarrassed by it. A flat "thanks!" wastes the warmest moment you'll get all week.
Questions you can't answer. "I don't know, I'll find out" in your voice, rather than silence, is a real decision worth writing down.
Someone being funny. Can you play along? How far? This one's worth an explicit answer, because it's where junior people either freeze or go too far.
Scale without becoming a machine
Templates are fine. Templates sent verbatim are not.
The distinction: a template is a starting shape, and the specific detail is what makes a reply land. "Thanks for the kind words!" is a template sent verbatim. "Thanks, and yes, the seal is the bit everyone underestimates" is the same template with the one detail that proves a person read the comment.
That detail costs about four seconds and it's the entire difference.
Don't automate sending. A model can draft; a human should send. Not for craft reasons. Because the failure mode of automated replies is catastrophic and rare, which is the worst combination. The one time it fires an upbeat auto-reply at somebody describing a serious problem, you'll wish there had been a person in the loop.
The honesty line
Everything above is about consistency, which is legitimate. There's a line past it that isn't.
- Don't fabricate specifics. A model asked to sound warm and knowledgeable will happily invent a delivery timeframe or a product detail. Every factual claim in a reply is your responsibility, whoever drafted it.
- Don't pretend a bot is a person where it matters. Signing an automated reply with a real employee's name is a small lie that costs a lot when discovered.
- Don't use a voice document to dodge. If the honest answer is "we got this wrong", no amount of tone calibration improves it.
Where Acumin fits
Brand Voice stores exactly those six fields: tone, audience, do, don't, signature phrases, sign-off. It can draft a first version by reading your existing content, which is genuinely useful as a starting point and should be edited, not accepted: what the model infers is what you've done, which isn't always what you'd choose.
Once saved, the voice shapes everything Acumin writes for you: engagement replies, generated copy, the assistant. One document, not a setting repeated per surface.
And the constraint that matters: Acumin drafts, you send. Every reply goes out under your hand.
How to use this tomorrow
Write the don't-list first. Five items, each one a mistake that has actually happened.
It's the fastest of the six fields to write, it's the one that prevents the most damage, and once it exists the other five are much easier, because you've already worked out what you're not.
Related: Comment triage decides which comments get a reply at all. The 24-hour DM window is the clock you're writing against in DMs.