GuidesUpdated

ChatGPT Prompts for Restaurants: Reports, Rotas, Reviews and More

Seven copy-ready ChatGPT prompts for UK restaurants, cafes and bars: menu descriptions, Google review replies, margin reports, legal staff rotas, social posts, supplier emails and translations, with what to paste in and what to check.

Alex Riesenkampff

10 July 2026 · Updated 8 October 2026 · 16 min read · Markdown

The ChatGPT prompts that save a restaurant owner time give the model a role, the real facts of the job and the format you want back, and the answer gets checked before it reaches a guest or a member of staff. The ONS found that 29% of UK businesses used at least one AI technology in June 2026, up 8 percentage points in a year, and 17% used large language models to write text. Below are seven copy-ready prompts for an independent cafe, bar or restaurant: menu descriptions, Google review replies, margin reports, staff rotas, social posts, supplier emails and menu translations. Each comes with what to paste in, what to check, and what changes when Super44 already holds the venue's data.

Which restaurant jobs can ChatGPT do well?

Writing is where it works, and the National Restaurant Association's US figures show how unevenly: 19% of full-service and 15% of limited-service operators used AI for marketing, against 10% for admin and 6% of restaurants for taking orders. The ONS reports that accommodation and food businesses are among those that most commonly use free-to-use AI software, so for most venues this is a chat window and a copied prompt rather than a purchase. The table below maps seven routine restaurant jobs to the prompt, the inputs and the check each one needs.

The jobs that suit a copied prompt are occasional ones. If you find yourself exporting the same report every Monday, the last section covers when to connect the source instead.

How do you write a prompt that doesn't come back generic?

Give the model three things, a role, the facts and the format, because OpenAI's own prompt-engineering guide recommends stating identity, instructions and context, and a menu prompt with a 25-word limit and three variants comes back far more usable than "write a description". For a restaurant that means:

1. Role: "You are a menu copywriter for independent restaurants" or "You are the owner replying calmly to a critical review."
2. Facts: the dish, the ingredients, the prep, the guest, the tone. Without them the model fills gaps with plausible filler. With them it sounds like your place.
3. Format: length, number of variants, register. "Three variants, under 25 words each, warm with a little dry humour" leaves the model far less to guess.

Then iterate. A second instruction such as "shorter", "less salesy" or "closer to how we talk" nearly always beats rewriting the first prompt. OpenAI's guide also notes that the best content and order vary by model, so if one assistant keeps missing the point, move your facts to the end of the prompt and try again.

What is the best ChatGPT prompt for menu descriptions?

The best menu prompt forces specific ingredients and preparation into the copy, and its output must be checked against all 14 allergens UK law requires a food business to declare for loose food. "Grilled chicken sandwich with honey sauce" is a fact. "Crispy grilled chicken, hot honey glaze, melted pepper jack, house pickles, toasted brioche" is a reason to order it. The prompt has to supply that detail; the model cannot know it. If you want the whole menu laid out rather than the lines on it, our comparison of AI menu makers prices the design tools that turn the copy into something printable.

You are an experienced menu copywriter for independent restaurants.
Write a menu description for: [dish name].
Ingredients and prep: [e.g. "dry-aged beef, 48-hour braise, celeriac purée,
roasted mushrooms, red wine jus"].
Audience/tone: [e.g. "neighbourhood bistro, warm, no sales-speak"].
Length: max 25 words.
Give me three variants, each with a different focus (preparation / ingredient
sourcing / the eating experience).
Do not add any ingredient, garnish or origin claim I have not listed.

A cafe, a bar and a bakery should get different results from the same prompt, because their inputs differ. What does not change is the check. Models add plausible details, and a garnish of toasted sesame or a "nutty" brown butter that is not on the plate is an allergen statement you did not make. GOV.UK is plain about the rule: food businesses must give allergen information for any loose item containing one of the 14 allergens, and verbal information must tell the guest where to find written confirmation.

Unreviewed AI text causes trouble beyond allergens. In April 2026 a Cheltenham restaurant apologised after a blog post wrongly said a rival had closed, and said the error came from an AI-generated draft "that was not appropriately reviewed before publication." Read every line before it goes public. For what to do with the dishes once the words are right, our guide to menu engineering beyond the matrix covers placement and pricing.

How should you reply to Google reviews with ChatGPT?

Use ChatGPT to draft a short reply that names the specific point in the review, then read it before posting: in a US BrightLocal test of replies to a positive restaurant review, 58% of consumers unknowingly picked the AI-written reply as their favourite. A well-drafted reply does not give itself away; a stock one does. Google's own guidance asks owners to "Keep it short and simple" and to avoid "sending the same 'thank you' to everyone". The prompt below handles an ordinary review; for the ones that go wrong, the angry one-star with its facts muddled or a review naming a member of staff, our guide to replying to Google reviews with AI has a prompt for each.

Operators feel the tension. At a 2025 Restaurant Online roundtable, Heath Ball, owner of The Red Lion & Sun pub, asked: "Isn't there a danger that people will realise they're being replied to by a computer?" Mark Selby, co-founder of Wahaca, said in the same discussion: "I'd hesitate to let AI reply to a customer without a human sense-checking it."

Review text is also data about your venue. An International Journal of Hospitality Management study of 935,386 Google Maps reviews across 5,010 restaurants in London, Birmingham and Manchester linked food, service, atmosphere and value to the odds of a five-star rating. Guests read those reviews through AI too: a CGA by NIQ and Reputation survey found 26% of consumers in Britain used tools like ChatGPT to learn about venues, close to the 27% who used Google Maps.

You are the owner of a [cafe/restaurant/bar] replying to a Google review,
polite but personal, not corporate.
Review: [paste the review text].
Star rating: [number].
What happened on our side: [e.g. "kitchen was two chefs down that night"].
Our tone: [e.g. "casual and honest" or "formal and understated"].
Name the specific issue from the review, apologise only for what went wrong,
and, if relevant, offer one concrete next step (an invitation to talk, not just
"we'll follow up").
Max 60 words, no stock phrases like "we appreciate your feedback."

A Super44 analysis for a café in Cologne read its full Google review history and estimated that 1- and 2-star experiences were costing around €4,000 a month. The resulting plan found roughly €3,600 a month across weekend staffing, service training and reservation confirmations, with reply drafts for open reviews. Those figures describe an opportunity found, not revenue recovered. How many reviews you need in the first place is a separate question, answered with numbers in our guide to how many Google reviews a restaurant needs. If pasting each review into a chat is the step that keeps slipping, our comparison of tools for handling Google reviews covers what the paid options add.

How do you upload a POS report to ChatGPT or Claude and calculate margins?

Export the report as CSV or XLSX, remove personal data, define every column, and make the model calculate cost of goods sold as opening stock plus purchases minus closing stock, because purchases alone can overstate food cost by several points. Here is the arithmetic on an illustrative month:

The shortcut makes this month look four points worse than it was, because £1,600 of the purchases is still in the fridge. In a month where you run stock down, it errs the other way. Gross margin and operating margin are different measures again, and net sales cannot be set against costs that include VAT. Our food GP calculator does the same sum per dish.

Before uploading, remove guest and employee names, email addresses, phone numbers and free-text notes. If you are still choosing which assistant to use, our comparison of the AI tools worth using to run a restaurant sorts them by whether they can see your venue's data at all.

I run a [restaurant/cafe/bar]. I have attached a [daily/weekly/monthly] report.
The currency is [GBP/EUR/USD], sales are [including/excluding] VAT, and each row
represents [explain the rows]. Relevant columns:
- Net sales: [column]
- Opening inventory, purchases, and closing inventory: [columns]
- Wages plus employer NICs, pension, and other on-costs: [columns]
- Other operating costs: [columns]

First check whether the data is complete enough. Calculate cost of goods sold
as opening inventory + purchases - closing inventory; do not treat purchases
alone as food cost. Then calculate gross-margin, food-cost, fully loaded
labour-cost, and operating-margin percentages. If an input is missing, say so
instead of estimating it. Then:
1. show every formula;
2. calculate the metrics by week and for the full period;
3. identify the three largest changes;
4. cite the exact rows behind each conclusion;
5. suggest three questions I should investigate next.

The same structure works for a Z-report, a labour-cost check or a month-on-month food cost comparison. The answer is only as current as the file, so next month the export, clean-up and upload start again.

Can ChatGPT write a staff rota that follows UK law?

ChatGPT can draft a usable rota only if the prompt states the UK limits, because the model will not apply a 20-minute break after 6 hours, 11 hours' rest between working days or the 48-hour average week unless you ask for them. Young workers have stricter rules, and a weekend team often includes them. A chat window is the free route; our comparison of AI rota tools prices the apps that draft the week themselves for a team of twelve.

Help me draft a staff rota for [period, e.g. "next week"].
Team: [roles, e.g. "FOH 1 (full-time), Kitchen 1 (part-time 20h),
FOH 2 (weekend casual, aged 17)"].
Availability: [who can't work when].
Expected footfall by day/daypart: [rough estimate].
Hard limits to respect for workers 18 and over: no more than an average
48-hour week, at least 11 hours' rest between shifts, an uninterrupted
20-minute break for anyone working more than 6 hours, 24 hours off each week.
For anyone under 18: a 30-minute break after 4.5 hours, 12 hours' rest between
shifts, 48 hours off each week, no more than 8 hours a day or 40 a week.
Do not schedule anyone under 18 between 22:00 and 07:00 in this draft.
Flag any requested exception for manual checking; midnight to 04:00 is
prohibited even under the hospitality exception.
Do not invent availability. Show the result as a table with start and end
times per person and day, list any uncovered hours, and flag any shift that
sits close to one of these limits.

If the rota lives in a spreadsheet, attach the XLSX or CSV and explain the columns. Ask for a list of proposed changes before a rewritten file: "move Tuesday's late from FOH 1 to FOH 2" is easy to check, while a silently regenerated rota is not. Role labels reduce direct identification, but ages, hours and availability can still identify someone in a small team. Treat the file as personal data: remove unnecessary details and check the account's data handling before uploading it.

The model will optimise for whatever you emphasise, so do not let cost be the only goal. Heath Ball said his managers would welcome a computer-made rota: "It stops them getting manipulated by staff who want certain shifts or don't want certain shifts." Fairness still needs a human eye: who keeps getting the closes, who never gets a Saturday off. Our guide to building a fair staff rota sets out the four-week fairness check. The money matters too: at £25,000 monthly revenue, one percentage point of labour cost is about £3,000 a year.

Which prompts work for social posts, supplier emails and translations?

These three jobs need the least checking, but each still needs one concrete fact in the prompt, such as "the price rose 12% with no notice after three years as a customer", or the draft comes back vague. Marketing is where restaurant operators already use AI most, on the National Restaurant Association's figures above the job table.

Write an Instagram caption for [today's special/promotion] at our [cafe/bar/
restaurant]. Tone: [e.g. "casual, minimal exclamation marks, no emoji
overload"].
Give one concrete reason it's worth trying (ingredient, seasonality, limited
run), not just "come try it!" with no reason attached. Max 3 sentences plus 3
relevant hashtags.

Write a polite but firm email to our supplier [name/product]. Situation: [e.g.
"price for X rose 12% with no advance notice"].
Goal: [e.g. "ask for justification, propose a volume discount or price hold"].
Tone: businesslike, not confrontational, but clear about what we expect. Max
120 words.

Translate this menu item into [language]: [paste the original].
Keep internationally recognised dish names in their original form and briefly
explain them in parentheses. Keep every allergen named in the original.
Tone: appetising, not a literal, technical translation.

Dom Jacobs, managing director of The Cadogan Arms, uses AI for clearer staff handover notes and adds a warning about internal writing that applies to captions too: "You don't want everything in your business to read like it was written by a robot." Edit the caption until it sounds like the person who serves the special.

What restaurant data is safe to paste into ChatGPT?

Public review text is low-risk, but staff names, rotas, guest contact details and complaint notes are personal data, and 51% of UK and Irish operators in an Access Hospitality survey named data security and privacy among their biggest AI concerns. The ICO says data protection law applies to personal data in generative AI even when that processing is "incidental" or unintentional.

A practical rule set for a small venue:

• Paste freely: public review text, your own menu, supplier prices, aggregated sales by day.
• Replace first: staff and guest names (use roles or "Guest A"), phone numbers, emails, booking references.
• Keep out: health details behind a sick call, pay and disciplinary notes, anything about a named guest that is not already public.
• Check the account: data and training settings differ by product and plan. Anthropic, for example, uses Claude Free, Pro and Max chats for training only if the user opts in, apart from chats flagged for safety review. Look in your own assistant's data controls before uploading anything with a name in it.

The ICO has been direct with the wider industry. Closing its investigation into Snap's My AI chatbot in May 2024, Stephen Almond, the ICO's Executive Director of Regulatory Risk, said organisations "developing or using generative AI must consider data protection from the outset". For a restaurant, that means the clean-up happens before the paste.

When should you stop pasting and connect your data instead?

Copy-paste prompts suit a job you do once a month or less; once the same export, clean-up and upload happens every week, a connected system is faster and keeps personal data out of a general chat window. Use this order:

1. Start with the prompts above for menu copy, one-off review replies, a supplier email or a single month's margin check. They are free and they show you every formula.
2. Switch when the job repeats. A weekly sales review, a rota rebuilt every Thursday or a review queue that never empties are signs the upload has become the work.
3. Do not automate what nobody reads. Allergen statements, anything posted under your name and any rota change that affects someone's pay need a human sign-off, whichever tool drafted them.
4. Reverse course if the checks keep finding errors. If drafts keep inventing ingredients or missing rest rules, tighten the prompt with those exact limits before trusting it with more.

Super44 keeps the restaurant's sales, reviews and staffing context current and can raise the next issue without waiting for an upload. For a decision that needs someone inside the venue, our guide to combining AI with a hospitality consultant shows where each one helps.