Most people get AI review replies wrong for the same reason: they think they're talking to a person. They write things like "be professional but warm" or "you are a customer service expert with 20 years of experience" and expect the AI to understand what that means. It doesn't.
An AI doesn't understand your instructions. It reads them and then writes whatever comes next most naturally — like a game of Mad Libs (fill in the blanks). You're not briefing someone. You're filling in context that a machine will use to complete a sentence. The more specific and concrete that context is, the better the output. The vaguer it is, the more generic the result.
This distinction matters because it changes what you should write in your prompt. Instead of describing a vibe, you give the AI actual material to work with: real sentences it can use, real information it can't guess, and clear rules it can follow mechanically.
Start by testing the default output
Before writing anything, generate a few replies with no custom instructions. See what you're working with. Some tools will give you decent output out of the box, others will give you something painfully generic — either way, you need a baseline before you start prompting.
On ReviewFlowz, the default AI agent is already heavily prompted and instructed to follow the patterns from your previous replies. If you've been replying to reviews manually, the agent picks up on your patterns and matches your existing style. So most users don't need custom instructions at all.
But regardless of the tool you use: test first, prompt second. If the default output is already 80% right, a few targeted lines will get you the rest of the way. If you start writing a novel-length prompt before seeing the default, you're almost certainly making things worse.
What NOT to put in your prompt
This is where most people waste their prompt. These instructions sound useful but add nothing — or actively hurt output quality:
Tone and personality descriptions. "Be professional but friendly." "Use a warm and empathetic tone." "Be concise but thorough." These are vague adjectives that an AI interprets in the most generic way possible. You'll get the same vanilla output with or without them.
Persona and role-play instructions. "You are a customer service expert with 20 years of experience." "You are a brand ambassador for our company." The AI doesn't become a different entity because you told it to. These instructions are noise.
Things your tool may already handle. Check what the default output does before duplicating it in your prompt. If the tool already greets the reviewer by name, analyzes sentiment, or keeps replies concise — adding "address the reviewer by name" or "keep it short" is just noise that clutters your prompt.
Abstract guidance. "Be nice but not too nice." "Strike the right balance between empathy and professionalism." These mean nothing to a machine filling in text. You wouldn't write "draw something beautiful" on a paint-by-numbers sheet — you'd fill in the actual colors.
What to actually put in your prompt
Your prompt should contain two things: concrete examples the AI can reuse and rotate between, and factual business information the AI has no way of knowing.
#1 Sentence examples with rotation instructions
This is the single most impactful thing you can do. AI models are deterministic — given similar input, they produce similar output. If you don't provide variety, your replies to 50 five-star reviews will all sound the same.
The fix is simple: give the AI a bank of example sentences and tell it to rotate.
Good prompt — opening sentences:
Use one of these openings and rotate between them:
"Thanks for sharing your experience, [name]."
"Really appreciate you taking the time to write this, [name]."
"Great to hear from you, [name]!"
"This means a lot to us, [name] — thanks for the kind words."
"[Name], thank you for the detailed feedback."
Good prompt — closing sentences:
Rotate between these closings:
"Looking forward to seeing you again soon."
"Don't hesitate to reach out if you ever need anything."
"Hope to welcome you back before long!"
Bad prompt (what NOT to write):
"Vary the tone and structure of each reply to keep things fresh. Use different openings and closings. Be creative and avoid repetition."
That last example is exactly the kind of instruction that sounds reasonable to a human but gives the AI nothing to work with. "Be creative" is not actionable. "Use one of these five sentences" is.
The same logic applies to transition phrases, ways to acknowledge a compliment, or ways to address a complaint. If you want variety, provide the actual variants. Don't describe variety in the abstract.
#2 Business information the AI can't invent
The other half of a good prompt is factual context. Things like:
Your business name (as you want it written — "Chez Laurent", not "the restaurant")
A support email address: "If a reviewer reports an issue, direct them to [email protected]"
A phone number or booking link: "For reservation issues, include our booking link: https://example.com/book"
Specific product or service names: "Our main product lines are X, Y, and Z"
Key policies: "We offer a 30-day return policy on all items" or "We do not offer refunds on custom orders"
This is information the AI would otherwise have to make up or leave out. Giving it these facts means it can include them naturally when relevant.
#3 Negative instructions (what to avoid)
Telling the AI what not to say is often more effective than trying to describe what you want. AI follows mechanical rules well.
Good negative instructions:
Never mention competitor names in a reply.
Do not offer discounts or compensation in review replies.
Never say "we apologize for the inconvenience" — it's generic and overused.
Do not promise specific timelines for fixes (e.g., "this will be resolved within 24 hours").
Never use the words "noted" or "we strive."
These are specific, binary, and easy for an AI to follow. Compare that to "avoid sounding corporate" — which is subjective and will be interpreted differently every time.
Keep it short
A good custom prompt is 3 to 8 lines long. If yours is longer than a short paragraph, you're almost certainly over-prompting.
Why does length matter? In some AI reply systems, your custom instructions replace or override part of the default system prompt. The more you write, the more default behavior you may be displacing. In others, your prompt is just appended — but a wall of vague text still drowns out the useful parts.
Think of it this way: every sentence in your prompt should be something the AI literally could not figure out on its own. "Be polite" — it already knows. "Our return policy is 14 days, no exceptions" — it doesn't.
Use separate prompts for different review types
A single prompt trying to handle five-star praise, one-star complaints, and everything in between will be mediocre at all of them. If your tool supports it, create separate configurations:
One for positive reviews (short, warm, varied)
One for negative reviews (empathetic, includes contact info, doesn't make promises)
One per platform or region if your tone differs across them
In ReviewFlowz, these are called Agents, and they can be linked to automation Flows so the right agent handles the right type of review automatically. But regardless of the tool you use, the principle is the same: narrower focus means better output.
Test and iterate in small steps
Don't write your prompt and publish it. Test it on real reviews first — ideally a mix of positive, negative, and mixed reviews in different languages if relevant.
When something isn't right, resist the urge to rewrite everything. Change one thing at a time. Add an example sentence. Remove a vague instruction. Add a negative rule. Then test again.
Small, targeted adjustments are how you get from good to great. Full rewrites usually just trade one set of problems for another.
TL;DR
Put in your prompt:
3–5 example opening sentences + "rotate between them"
3–5 example closing sentences + "rotate between them"
Your business name, support email, phone number, or booking link
Key policies (returns, refunds, warranties)
Specific things to never say or do
Leave out of your prompt:
Tone descriptions ("be warm", "be professional")
Persona instructions ("you are a 20-year expert")
Things your tool's default already handles (test first)
Abstract creativity instructions ("vary your replies", "be original")
Anything longer than a paragraph