Giving an AI tool your business's voice
Every post sounds like it could belong to any business. Same cheerful opening, same stock phrases, nothing that says it's you. That happens when the tool has been told almost nothing about you.
Cause one: the instruction is too thin
"Write a post about our sale" leaves the tool guessing at everything. Anthropic's guidance says that spelling out the output you want can improve results.1
Try describing the reader and the register. Who is this for? Regulars, new locals, other businesses? Is the tone chatty, dry, warm, plain? Name the things you'd never say.
Cause two: no role or house rules
Tools let you set standing instructions, so you don't repeat yourself every time. Anthropic notes that giving the model a role in its standing instructions focuses its behaviour and tone.1 OpenAI's guide says such instructions can cover tone and goals, and can include sample answers.2
In plain terms: write a short, reusable note and paste it at the top of each session. It might say:
- We're a family-run bakery. We write like a friendly neighbour.
- Short sentences. No exclamation marks.
- Never use the words "premium" or "luxury".
- Never state a price, date or opening hour. Leave a blank for the owner to fill in.
That last rule helps with accuracy too. A blank is easier to spot than a wrong number.
Cause three: no examples
Description only goes so far. Showing the tool your real writing works better. OpenAI's advice is to give a varied set of sample inputs alongside the outputs you want.2 For posts, paste a few past posts you were happy with, and a note on why.
Pick posts that differ. One about a product, one about a quiet Monday, one thanking a customer. A set made only of sales posts may pull every draft that way.
Anthropic suggests a quick test for any instruction. Hand it to a colleague who knows little about the task and ask them to follow it.1 If they'd be confused, the tool may be too.
Use only posts you wrote and have the right to reuse. Leave out customers' comments and messages. Australia's regulator suggests keeping personal and sensitive details out of public AI chatbots.3 The same caution is sensible elsewhere, though the legal rules differ by country.
Cause four: you accept the first draft
The first draft is a rough cut. Change it. Swap the stiff phrase for the one you'd actually say. Cut anything that sounds like an advert.
Then keep your edited version. It's your best example for next time.
When it still sounds off
Some voices are hard to describe. If the draft stays generic after you've tried the fixes, write the first two sentences yourself and ask the tool to continue. Or use it only for the dull parts, such as a list of opening hours, and write the human parts yourself.
Whatever you end up with, a person still has to read it. For what to look for, see the pre-posting checks. For what goes into each request, see prompt writing.
Where the facts come from
- Prompting best practices, Anthropic
- Prompt engineering, OpenAI
- Guidance on privacy and the use of commercially available AI products, Office of the Australian Information Commissioner