Give the AI your own words before it writes any: ten of your current menu descriptions, a voice note of you describing dishes to a regular, and a list of words you'd never use. Then let it draft only from dish facts you supply, and edit every line by reading it aloud. It offers options; the voice and facts stay yours.
Done this way, a full menu rewrite takes an evening rather than a week, and specials take minutes. The main risk isn't bland copy, which is easy to spot; it's invented detail, such as a "hand-picked" herb or a "house-cured" ham that isn't, so the method below builds in a fact check at every stage.
Why AI menu copy sounds the same everywhere
Ask any assistant to "write a mouth-watering description of our lamb flatbread" and you'll get something like this:
Indulge in our succulent slow-cooked lamb, nestled on a warm artisanal flatbread and drizzled with a refreshing mint yoghurt, finished with jewel-like pomegranate seeds.
That's what these tools produce when they know nothing about you: the average of every menu they've read. Every adjective is doing marketing work and none is doing information work. It also invents: "artisanal" is a claim about your bread that you never made. The fix isn't a cleverer prompt. It's giving the AI enough of your own material that the average stops being its only option.
Stage 1: capture how you actually talk about food (30 minutes)
Collect three kinds of raw material:
- Your existing menu. Pick the ten to fifteen descriptions you're happiest with. If there aren't ten, pick the ones regulars quote back to you.
- A voice note. Record yourself describing five dishes as you would to a regular at the pass or the counter: what's in it, why you like it, who orders it. Three or four minutes is plenty. Your phone's transcription, or pasting the audio into an assistant that accepts it, turns this into text.
- Your "never" list. Words you'd be embarrassed to see on your own menu. Most kitchens come up with a dozen in two minutes.
The voice note is the most useful of the three. People write menus more stiffly than they speak, and the spoken version is usually closer to the voice diners already know.
A stretch of a typical transcript, from a grill owner talking about the pork chop (illustrative):
So the chop, that's a thick one, proper thick. We brine it overnight, just salt and a bit of sugar, then it goes on the grill hard so the fat catches. People who order it have usually been before. It comes with the apple and mustard thing, which is basically a sharp relish, cuts through the fat. I tell people it's the thing I'd eat.
Three voice-card lines come straight out of that. "Proper" and "hard" go under words we use. "Cuts through the fat" shows the owner explains why a side is on the plate, so "say what the sharp thing is for" goes under how we sound. And "the thing I'd eat" is a line to use once on the whole menu, on the dish that deserves it. None of those would come out of a prompt, which is why the recording matters more than the old menu.
Stage 2: turn it into a one-page voice card
Paste all three into an assistant and ask it to draft a voice card, then correct the draft yourself. It should end up roughly like this:
VOICE CARD - [restaurant name]
Who we are: [one line, e.g. neighbourhood grill, open kitchen,
most customers are regulars]
How we sound: [e.g. short, plain, a bit dry. Name the ingredient,
then how it's cooked. No selling.]
Sentence length: [e.g. fragments are fine. Max 20 words a dish.]
Words we use: [e.g. charred, sharp, proper, overnight, our own]
Words we never use: succulent, mouth-watering, nestled, drizzled,
indulge, elevated, artisanal, infused, perfectly, jewel-like,
melt-in-your-mouth, symphony, [your own]
Suppliers: [name them only when the diner would recognise
the name, e.g. "bread from the bakery two doors down"]
Examples we love: [paste 5 of your best descriptions]
Never write: allergen, dietary or health claims; "homemade",
"fresh", "local" or "hand-made" unless the dish facts say so.Keep it to a page. A longer card gets followed less closely, not more. If you want to build a fuller voice guide for all your writing, building a brand voice guide AI can follow takes it further.
Stage 3: give it facts, not adjectives
For each dish, fill in one row of a simple sheet. The AI writes from these columns and nothing else:
| Column | Example: lamb flatbread |
|---|---|
| Main components | Lamb shoulder, flatbread, pickled red onion, mint yoghurt, pomegranate |
| How it's cooked | Lamb slow-roasted overnight; flatbread made each morning and cooked on the grill to order |
| Where it comes from (if worth saying) | Flatbread is our own |
| Texture and taste notes | Rich, sharp from the onions, cool yoghurt |
| Who orders it | Lunch crowd, people sharing |
| Where the text will appear | Printed menu and online ordering |
Allergens are deliberately missing. They come from your allergen records, set out in whatever format the rules where you trade require, and never pass through the AI. Whether an AI should answer allergen questions at all covers the wider risk.
Stage 4: the drafting prompt
Put the voice card into a project (ChatGPT and Claude both let you store standing instructions and files in one) so every chat starts with it. Then, for each batch of dishes:
Using the voice card, write menu descriptions for the dishes below.
For each dish give 3 options:
A: printed menu, 8-15 words
B: online ordering, 20-35 words (the reader can't ask a server,
so say what's in it and roughly how big it is)
C: specials board, under 8 words
Use ONLY the facts in the rows. Do not add ingredients, methods,
origins or quality claims that aren't there. If a fact seems
missing, list it as a question instead of guessing.
No word from the "never" list.
[paste dish rows]The "list it as a question" line is the one that stops invention. You'll get back questions like "Is the yoghurt made in-house?" which are easy to answer and far better than a confident guess.
Here's a full reply for the pork chop, whose row said: thick-cut chop, brined overnight, grilled; apple and mustard relish (sample output, illustrative):
A: Thick pork chop, brined overnight and grilled hard.
Apple and mustard relish.
B: A proper thick pork chop, brined overnight and grilled over
high heat until the fat crisps, with a sharp apple and mustard
relish to cut through it. Served with perfectly golden
hand-cut chips.
C: Brined pork chop, apple relish
Questions:
1. Roughly how heavy is the chop, for the online description?
2. Is the relish made in-house?Option A is ready to use. Option B's last sentence breaks two rules at once: "perfectly" is on the never list, and the chips aren't in the facts row at all, so the model has added a side dish and a claim about it ("hand-cut") despite being told not to. Delete the sentence, answer question 1 ("about 350g before cooking"), and rerun option B only. The questions are the most valuable part of the reply: the weight of the chop is exactly what a diner ordering online wants to know and can't ask.
For the lamb flatbread, a good option A comes back like this:
Lamb shoulder, roasted overnight, on our own flatbread. Sharp pink onions, mint yoghurt, pomegranate.
Fourteen words, no selling, and every word traceable to the facts sheet. Compare it with the generic version at the top.
Matching the length to where the words appear
The same dish needs different words in different places, and AI is genuinely good at producing the variants once the facts are fixed. A rough guide:
| Where | Length | What the reader needs |
|---|---|---|
| Printed menu | 8-15 words | What it is and one reason to pick it. A server can fill in the rest. |
| Online ordering and delivery apps | 20-35 words | Everything on the plate, rough portion size, spice level, whether it travels well. Nobody can ask. |
| Specials board | Under 8 words | The hook. "Crab on toast, brown butter" is enough. |
| Set or tasting menu | 3-6 words a course | Main ingredient and technique; the server tells the story. |
| Children's menu | Very short, plain words | What it is, written for the parent reading it aloud. |
| Drinks list | 5-12 words | Taste, not tasting-note poetry. "Dry, citrusy, good with the fish." |
Online ordering is where most kitchens under-write. A diner scrolling a delivery app at 8pm is deciding between you and three other places on the strength of a few lines, and "Chicken burger" loses to a description that says what's in the bun. Ask the AI to add portion and spice information to option B specifically.
The difference on a delivery app, before and after (illustrative). Before: "Chicken burger". After, from a filled-in facts row:
Buttermilk-fried chicken thigh in a toasted brioche bun with pickles, shredded lettuce and chipotle mayo. Medium heat. One large burger; chips sold separately. The bun is toasted, so it holds up in the box.
Every added phrase answers a question an app user can't ask: how spicy, how big, what's included, and whether it will arrive soggy.
Tasting menus go the other way. An assistant's first attempt at a scallop course tends to read "Hand-dived scallop, silky cauliflower purée, crispy capers, golden raisins", which is four claims and three adjectives for a line the server will talk through anyway. Trimmed to the format: "Scallop, cauliflower, capers, raisin". If the scallops really are hand-dived, that's the server's story to tell at the table, where it lands better.
A worked evening: rewriting a 20-dish menu
Say a 40-cover neighbourhood grill (an illustration) wants to refresh its whole card before a seasonal change. A realistic evening:
- 6:00-6:30: record the voice note, pick the ten best current descriptions, write the "never" list.
- 6:30-7:00: have the AI draft the voice card; correct it by hand.
- 7:00-7:45: fill in the facts sheet for 20 dishes. This is the slowest part and the most valuable, because it forces decisions about what's actually true.
- 7:45-8:15: run the prompt in batches of five dishes, answering the questions it raises.
- 8:15-9:15: the edit: read aloud, trace claims, scan for banned words, pick one option per format.
- Next day: the blind test with front-of-house staff before service, then any fixes.
About three hours for a full menu in three formats, against the day or two it takes many owners to write one from scratch. Most of the saving comes from not staring at a blank page; the facts and the final judgement still take your time, as they should.
Stage 5: the edit that puts you back in
Never paste AI output straight onto a menu. Run each description through four checks:
- Read it aloud. If you wouldn't say it to a customer, change it. This single test catches most AI tone.
- Trace every claim. Each ingredient, method and origin must be in the facts row. Words like "homemade", "fresh" or "local" can carry legal weight in some places and invite complaints when they aren't strictly true, so keep them only when they are.
- Scan for the "never" list. Use Find in your document for each banned word. They creep back in on the fourth or fifth batch.
- Check the kitchen can deliver it. "Grilled to order" on a dish that's batch-cooked at 11am will be noticed by the first regular who watches the pass.
Then do a blind test with your team: shuffle five AI-assisted descriptions with five of your originals and ask front-of-house staff which are new. If they can't tell reliably, you've kept your voice. If they spot every AI one, look at what gave them away (usually rhythm or one pet phrase) and add it to the voice card.
A typical giveaway (illustrative): front-of-house at the grill picked out all five AI descriptions in under a minute. Nothing in them was wrong. Every one simply opened the same way, ingredient, comma, cooking method: "Lamb shoulder, roasted overnight…", "Hake, pan-roasted…", "Beetroot, salt-baked…". The owner's own descriptions started all sorts of ways ("The one regulars ask about", "Chargrilled, then…", "Our take on…"). One line added to the voice card, "vary how descriptions open; no more than a third start with the main ingredient", fixed it on the next batch, and the second blind test was close to a coin toss.
Before-and-after examples of edited AI copy shows the kinds of change that make the difference.
Where AI menu copy goes wrong in practice
- Sourcing it made up. "Line-caught", "free-range", "from the coast this morning". Diners do ask, and so do inspectors.
- Dietary labels. Asked to "make the vegan dishes sound appealing", an assistant may label a dish vegan that has honey or a butter finish. Labels come from your records only.
- Everything sounding equally special. If every dish is "our signature", none is. Tell the AI which two or three dishes deserve the extra words; the rest get plain descriptions. If you don't know which dishes those should be, menu engineering with AI tells you which ones are worth promoting.
- Online and printed menus drifting apart. Update both from the same facts row, or customers get a dish online that doesn't match what arrives.
- Words diners don't know. A "gochujang glaze" or "nduja crumb" may need three extra words of explanation online, where nobody can ask. Ask the AI to flag unfamiliar terms and suggest a short gloss: "nduja crumb (spicy, spreadable pork sausage, fried until crisp)". Check the gloss is true of your version; if you make the crumb with a vegetarian substitute, the AI's standard gloss is now wrong.
Keeping it consistent when the menu changes
Once the voice card and facts sheet exist, specials become a two-minute job: add a row, run the prompt, edit, done. The discipline that stops printed and online menus drifting apart is a short change log at the bottom of the facts sheet, filled in whenever a row changes (illustrative):
| Date | Dish | Change | Printed | Online | Board |
|---|---|---|---|---|---|
| 12 Oct | Lamb flatbread | Pomegranate swapped for pickled cherries | Done | Done | n/a |
| 12 Oct | Pork chop | Relish now bought in; "our own" removed from option B | Done | Done | n/a |
| 19 Oct | Hake | Off the menu; cod on the specials board instead | Next reprint | Done | Done |
The second row is the one that matters. A relish that changes from made-in-house to bought-in is a small kitchen decision and a real menu claim, and without the log the old description quietly stays on the delivery app for months. Every few months, add your newest favourite descriptions to the card's examples so the voice keeps developing rather than freezing at the version you started with. The same card works for social posts about new dishes too; writing Instagram captions that sound like you uses the same approach for captions.
Menu writing questions
Should I tell diners that AI helped write the menu?
There's no general expectation to label a menu that you've edited and checked yourself, any more than you'd credit a friend who helped with wording. What matters is that every claim is true and every allergen label comes from your own records. If a regular asks, an honest answer such as 'I use it for first drafts, then rewrite them' is fine.
Can AI translate my menu for tourists or a mixed team?
It can produce a good first translation, especially if you give it the dish facts rather than your English prose. Food terms are where it slips: regional dish names, cuts of meat and cooking methods often translate literally and wrongly. Have a fluent speaker check it, and keep allergen information in a format taken directly from your allergen records.
Which AI tool is best for menu writing?
Any of the major assistants will do the job; the method matters more than the brand. The useful feature is a project or workspace where you store your voice card and dish-facts sheet as standing instructions, so you don't paste them in every time. ChatGPT and Claude both offer projects with instructions and files.
Further reads
- How Caterers Use AI to Quote Events and Plan Menus — Menus for events and quotes, rather than the everyday card.
- Your First 30 Days of AI in a Restaurant, Week by Week — Where menu writing fits in a restaurant's first month of AI.
- How to Turn Customer Reviews Into Marketing Copy With AI — Borrowing diners' own words for your menu and posts.
- Brand Voice Tools Compared: Jasper, Writer and Custom Instructions — Whether a paid brand-voice tool beats a simple voice card.
- How Much Does AI Cost a Small Restaurant Each Month? — A line-by-line monthly AI budget for a small restaurant, three priced set-ups, and the staff hours that cost as much as the subscriptions.
- How to Build a Kitchen Training Manual With AI — Capture how your kitchen really works, let AI shape it into station cards and quizzes, and keep every temperature and allergen under the head chef's control.
- AI Prompts for Bakers: Pricing, Captions and Customer Replies — Twelve copy-ready prompts for bakers covering costing and prices, social captions, and customer replies, each with what to check before you use the answer.
- How to Write B&B Room Descriptions and Guest Guides With AI — Room fact cards, prompts for each listing, the words that lead to refund disputes, and a quick way to turn a walk round the house into a guest guide.
- How to Write a Salon Service Menu and Price List With AI — Price each service from your chair-hour cost, then use three AI prompts to structure, describe and stress-test the menu before it goes live.
- How Food Trucks Can Use AI to Plan Stock, Pitches and Posts — A trading log, a weekly planning routine and copyable prompts that help a food truck plan prep, choose pitches and post its schedule with AI.
- What AI Can and Cannot Do for an Independent Café — The desk jobs AI does well in a café, the ones it can't touch, and a 30-minute test on your own till data before you pay for anything.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI help pages on Projects in ChatGPT; Anthropic help pages on Claude projects and personalisation. Checked September 2026.