Most people get AI review replies wrong for the same reason: they think they are talking to a person. They write "be professional but warm", or "you are a customer service expert with 20 years of experience", and expect the AI to know what that means. It doesn't.
An AI doesn't understand your instructions. It reads them, then writes whatever comes next most naturally. You are not briefing someone, you are filling in context a machine will use to complete a sentence. So instead of describing a vibe, give it material: real sentences it can reuse, real facts it cannot guess, and rules it can follow mechanically.
Test the default output before you write anything
Generate a few replies with no custom instructions and see what you are working with.
In Reviewflowz the default agent is already heavily prompted, and it follows the patterns in the replies you have written yourself. If you have been replying manually, it picks up your style, and most accounts need no custom instructions at all.
Either way: test first, prompt second. If the default is 80 percent right, a few targeted lines get you the rest of the way. Write a novel before you have seen the default and you are almost certainly making it worse.
What not to put in your prompt
This is where most prompts get wasted. These sound useful and add nothing.
Tone and personality. "Be professional but friendly." "Use a warm, empathetic tone." Vague adjectives, interpreted in the most generic way possible. You get the same vanilla output with or without them.
Personas. "You are a customer service expert with 20 years of experience." The AI does not become a different entity because you told it to. This is noise.
Anything the tool already does. If the default already greets the reviewer by name and keeps replies short, "address the reviewer by name" is clutter. Test first, then you know.
Abstract guidance. "Be nice but not too nice." "Strike the right balance." These mean nothing to a machine filling in text. You would not write "draw something beautiful" on a paint-by-numbers sheet, you would fill in the colours.
What to actually put in your prompt
Two things: concrete sentences the AI can reuse, and facts it has no way of knowing.
Example sentences, with an instruction to rotate
The single most useful thing you can do. Given similar input, a model produces similar output, so without variety your replies to fifty five-star reviews all sound the same.
Give it a bank of sentences and tell it to rotate:
Rotate between these openings:
"Thanks for taking the time to write this, [name]."
"Really glad you enjoyed it, [name]."
"[Name], this made our day."
Do the same for closings, for transitions, for ways to acknowledge a compliment. What does not work is "vary the tone and structure to keep things fresh, be creative and avoid repetition". That sounds reasonable to a human and gives the AI nothing. "Be creative" is not actionable. "Use one of these three sentences" is.
Facts the AI cannot invent
Your business name as you want it written. A support address. A booking link. Product names. Key policies.
We are Chez Laurent.
Bookings go to chezlaurent.com/book, anything else to [email protected].
Our tasting menu changes every six weeks.
Without this the AI either leaves it out or makes it up.
What never to say
Telling the AI what to avoid works better than describing what you want, because the rules are binary and it can follow them mechanically.
Never mention a competitor.
Never offer a discount or compensation.
Never write "we apologise for the inconvenience".
Never promise a timeline.
Compare that to "avoid sounding corporate", which is subjective and gets interpreted differently every time.
Keep it to three to eight lines
If your prompt runs longer than a short paragraph you are over-prompting. Custom instructions sit alongside a default system prompt that is already doing a lot of work, and a wall of vague text drowns out the useful parts.
Every line should be something the AI could not work out on its own. "Be polite" it already knows. "Our return policy is 14 days, no exceptions" it does not.
One agent per kind of review
A single prompt covering five-star praise and one-star complaints will be mediocre at both. Build a separate agent for each, then let a rule decide which one handles what.
In Reviewflowz that lives under Replies. AI Agents is where you write each agent, and Reply Rules is where you route reviews to them.
Rules run top to bottom and the first match wins, so order matters. The last rule is the default that catches everything else. A rule can reply automatically with a named agent, queue the review for a person, or ignore it. Automatic replies post on Google Maps, App Store, Play Store, Trustpilot and Facebook.
The pattern most accounts land on: five stars answered automatically, one and two stars queued for a human, everything else on the default.
Change one thing at a time
Do not write a prompt and turn it loose. The agent editor has a test panel beside the instructions: it pulls a real review from your account and drafts a reply with the prompt as it currently stands, so you can see the effect before anything is posted. Step through a few, positive and negative, and in another language if your customers write in more than one.
When something is off, resist rewriting everything. Add one example sentence. Remove one vague line. Add one negative rule. Test again. Small targeted changes are how you get from good to great, and full rewrites usually just trade one set of problems for another.
There is more on customising your AI replies, and on replying to reviews with AI generally.