AI Prompts for Restaurant Owners: Menu Copy, Review Replies and Social Posts

A one-star review lands at 11:40 pm: "Waited 50 minutes, food was cold." You're tired, you know the kitchen was short-staffed, and every instinct says to reply with the full story. AI prompts for restaurant owners can help here, mostly by making you pause. They can also draft menu descriptions and Instagram captions in minutes. The line to hold: the model writes the words, and you own every claim about food, ingredients and safety.

These are untested templates. I haven't run them in a restaurant, and the worked examples are illustrative. For image generation see AI food photography prompts for restaurants, and for the broader small-business toolkit see AI prompts for small business owners. This post covers words only.

What the model must never decide

Settle this before any prompt. These claims are the owner's, based on recipes, suppliers, labels and whatever rules apply where you operate. I haven't checked any food-safety regulation for this post, so nothing below states a legal requirement.

• Allergens and dietary status: nut-free, gluten-free, vegan, jain, halal, vegetarian. A model will write "gluten-free" because it sounds right.
• Origin and quality claims: "farm-fresh", "organic", "imported", "locally sourced", "no preservatives".
• Health claims: "healthy", "low-calorie", "good for digestion".
• Hygiene or safety statements: "kitchen inspected", "100% hygienic".
• Prices, offers, opening hours unless you supply them.

The practical way to enforce this is a rule you paste into every prompt:

Do not make any claim about allergens, dietary suitability, ingredient origin,
freshness, health benefits, hygiene or certification. If a description would
imply one, write [OWNER TO CONFIRM: what] instead. Describe only the
ingredients and methods I list.

Then keep a one-page master sheet of verified dish facts (ingredients, allergens, dietary tags) that your kitchen lead signs off. Paste from that sheet, not from memory.

How do I write menu descriptions with AI?

Short, concrete, true. Good menu copy tells a hungry person what they'll taste and one reason to choose it. Models drift toward adjectives, so give them nouns.

Write menu descriptions for a [CUISINE] restaurant in [CITY]. Voice: [plain,
a little warm, no hype].

For each dish I give: name, ingredients, cooking method, one taste/texture
word I'd use, and one thing regulars say.

Rules:
- 12 to 20 words each. Start with the main ingredient or method, not an
  adjective.
- Ban: mouth-watering, delectable, authentic, finest, to die for, symphony,
  burst of flavour, tantalising.
- Use only the ingredients I list. Never add garnish, sauce or spice I did
  not list.
- No allergen, dietary, origin, freshness or health claims. Where a line
  would imply one, write [OWNER TO CONFIRM: ...].
- Give two versions of each: one plain, one with a hint of personality.

DISHES:
1. [NAME] | [INGREDIENTS] | [METHOD] | [TASTE WORD] | [REGULARS SAY]
2. ...

The "never add garnish" rule matters. If you list paneer, tomato and cream, a model will happily add "finished with fresh coriander" and now your description is wrong. Read each line against the recipe card. A wrong ingredient on a menu is not a style problem. For a customer with an allergy it's a risk.

For a bilingual menu, ask for the second language in a separate pass and have a native speaker check it, because dish names and regional terms translate unevenly. The regional-language translation checklist covers what to verify.

Menu engineering: a question, not an instruction

You can ask a model to suggest how to reorder or rename items, but it can't see your sales or margins. Paste your own numbers (item, price, units sold, food cost) and ask for questions, not verdicts.

Here are my top 15 items by units sold last month with price and food cost
(I computed contribution margin myself, shown below).
List questions I should answer about the menu layout: which high-margin
items are buried, which popular low-margin items could use a small price
change or a smaller portion, which low sellers could go. Do not calculate
anything. Do not claim what customers will do.

The ban on calculation keeps the arithmetic yours. Any test of a new price should be run on a few weeks of real sales, and you should look at what actually happened.

How do I reply to Google reviews without making it worse?

A public reply is written for the next hundred readers as much as for the reviewer. The goal is to look like a restaurant that listens and acts. Use AI to draft, and your own records for the facts.

Draft a public reply to this review of [RESTAURANT]. Review text (name
removed): "[PASTE]"

What I know from our records (use only this): [e.g. a Saturday-night
staff shortage, a delayed order, a known issue with the oven that day]
What I will offer: [e.g. invite them to contact me at EMAIL/PHONE]

Rules:
- Under 90 words. Start with thanks, not "We're sorry you feel...".
- Acknowledge the specific complaint in their words.
- State one concrete step we are taking, only if it is on my list.
- Do not argue, explain at length, blame staff or the customer, or reveal
  order details, names or times.
- Do not admit or deny legal responsibility. Do not promise refunds or
  compensation unless I wrote it.
- Move the conversation offline in the last sentence.
- Use the same warmth for positive and negative reviews.

"Under 90 words" and "move offline" do most of the work. Long defensive replies are what screenshots are made of. If a review mentions illness, injury or an allergic reaction, stop. Don't use a template. Contact the customer directly, record what happened, and speak to your food-safety lead and, if needed, a professional before replying in writing.

Positive reviews: don't let them all read the same

Replies that start "Thank you for your kind words!" thirty times in a row look automated. Ask for variation in structure by supplying one detail from the review:

Reply to this 5-star review in under 40 words. Mention one specific thing
they wrote: "[QUOTE ONE PHRASE]". Do not use "valued customer", "thrilled",
or "look forward to serving you". Sign off as [FIRST NAME, ROLE].

Fake reviews are off the table

Don't use a model to write reviews of your own restaurant, to reply as a fake customer, or to flood a competitor. It's dishonest, it harms customers, and platforms and regulators take it seriously. What each platform allows, including whether you may offer an incentive for reviews, is in the platform's own policy. I haven't verified those pages for this post, so read them directly. A legitimate use is drafting the ask: "Write a two-line message inviting guests to leave honest feedback, no incentives, with a link placeholder."

How do I plan a week of social posts?

The hard part of social media isn't the caption. It's deciding what's true and interesting this week. Start with a list of real things, then ask for formats.

Real things happening at [RESTAURANT] this week (the only content you may
use): [e.g. new monsoon special, Thursday live music at 8 pm, chef
returning from a trip with a new technique, a photo I took of the dosa
station]
Platform: Instagram / WhatsApp status / Google Business post.
Audience: [neighbourhood families, office lunch crowd, college students].

Write 5 posts: caption (max 40 words), a one-line idea for the photo or
reel, and the best posting day with a reason. Rules: no invented offers,
prices, awards or guest names; no hashtag walls (max 4, specific to our
area); no "foodie heaven" language; every dietary or origin claim marked
[OWNER TO CONFIRM].

For a Hindi, Marathi or other local-language caption, ask for one version and have someone who reads it check the tone. Don't assume the translation matches your voice. For competitor and customer research that feeds these posts, the review mining prompts show a way to find what customers keep mentioning, using reviews you have the right to read and quote.

An illustrative example

An invented dish and review, showing output shape rather than a real result.

Dish input: Malvani fish curry | kingfish, coconut, kokum, red chilli | simmered | tangy | regulars order it with rice.

Plain: Kingfish simmered in coconut and kokum with red chilli. Tangy, best
with steamed rice. [OWNER TO CONFIRM: fish source and any allergen
labelling for fish and coconut]
Personality: Our tangiest curry: kingfish, coconut and kokum, slow and
spicy. Regulars order it with rice.

Review: "Waited 50 minutes, food was cold." Records: Saturday, two staff absent. Draft reply:

Thank you for telling us. A 50-minute wait and a cold plate is not what we
want anyone to leave with. We were short-staffed that evening and we are
adjusting weekend rostering. I'd like to hear more: please message me at
[EMAIL] so I can look at your visit properly. - [FIRST NAME], owner

Note what it leaves out: no order details, no excuses beyond one true line, and no promised refund. Remove "we are adjusting weekend rostering" if you aren't.

Where this goes wrong

What not to automate

• Allergen and dietary labelling. The kitchen lead and your records decide it. A model has no way to know.
• Responses to illness, injury or legal complaints. Handle directly and carefully, with a professional if needed.
• Pricing and discount decisions. You know your food cost; the model doesn't.
• Replies to every review without reading them. The real reviews are free market research.
• Anything posted in your name that you haven't read. If a regular wouldn't believe it came from you, rewrite it.

For why the "list only what's true" pattern works, see avoiding hallucinations, and for ways to double-check any AI output, use the fact-checking checklist. The prompt library has more templates you can copy.