AI Menu Description Generator for Restaurants: A Fact-to-Publish Guide

What the generator should—and should not—do

A useful generator converts a structured dish brief into alternative wording. It can reorder supplied details, make a sentence more concise, align wording with a defined voice and produce separate QR and print drafts. It can also expose gaps by returning an item to review when the brief is incomplete.

It should not fill those gaps with plausible ingredients, infer a cooking method, declare an item free from an allergen, or turn a preference such as “fresh-sounding” into a factual freshness claim. Recipe records, current supplier labels and qualified restaurant staff control those decisions. Treat generated text as unapproved copy until it passes the same operational review as any manually written menu text.

This distinction matters because dish writing combines two kinds of information:

AI may help with the second list. The restaurant owns the first.

• **Locked facts:** dish name, actual ingredients, preparation, included sides and approved claims.
• **Editorial choices:** order of details, sentence rhythm, length, tone and which verified flavour cues to emphasise.

The LOCK–DRAFT–GATE framework

Use one repeatable framework for every item rather than relying on increasingly elaborate prompts.

1. LOCK the source record.

Create a compact record from the current recipe, plating specification and supplier information. Mark uncertain fields as “pending”; never convert a blank into a guess. Record prohibited claims too. For example, if the team has not checked every component for a dietary label, write “no dietary claim permitted”, not “probably suitable”.

2. DRAFT within controls.

Give the generator only the locked record plus editorial controls: audience, voice, words to avoid and a length ceiling for each format. Ask for variants, but instruct it not to add facts. A variant is useful only if the differences are editorial, not factual.

3. GATE before publication.

Compare the preferred draft with the locked record line by line. A chef, manager or other authorised reviewer should resolve ingredient, preparation and service questions. Complete the separate allergen and dietary review required by the operation. Then check the rendered QR and print versions; copy that fits a document may wrap awkwardly on a phone or crowd a printed category.

The gate ends with a decision: **approve**, **edit and recheck**, or **hold**. “Hold” is the correct choice whenever a safety-sensitive or factual field remains unresolved.

Reusable Dish Description Input-and-QA Matrix

Copy this matrix into a spreadsheet or content system. Keep one row per dish and preserve the approved source record when generating shorter versions.

| Field | What to enter | QA question | Hold publication when… | |---|---|---|---| | Dish identity | Approved name and menu category | Does it match the service and till naming? | The name could refer to a different recipe or size | | Verified composition | Ingredients and components that may appear in copy | Is every named component present in the current recipe? | A component is assumed, seasonal or unconfirmed | | Preparation | Verified cooking and finishing methods | Does “roasted”, “smoked” or similar wording describe the actual process? | The method was inferred for appeal | | Included items | Sides, sauces, garnish and portion cues that are actually included | Could a guest mistake an optional extra for an included item? | Inclusion or portion is unclear | | Flavour cues | Kitchen-approved cues such as bright, smoky or rich | Can staff connect each cue to the served dish? | The language overpromises or lacks a factual basis | | Dietary/allergen review | Status, reviewer and record reference used by the restaurant | Has the operation completed its required review? | Status is pending or based only on AI output | | Unsupported/prohibited claims | Terms the draft must not use | Has the generator introduced provenance, certification, health or superlative claims? | Any unsupported claim remains | | Brand voice | Two or three rules plus words to avoid | Would staff naturally use this language with a guest? | The copy sounds generic or conflicts with house style | | Format limits | QR and print length ceilings set for the actual layout | Is the essential distinction visible without expansion? | Trimming removes a material detail or creates ambiguity | | Approval | Reviewer, date and decision | Does the rendered version match the approved copy? | No accountable reviewer has approved it |

The matrix is deliberately stricter than a prompt template. It creates an audit trail between what the kitchen knows, what the generator receives and what the guest reads.

Example: from a locked brief to QR and print copy

This is an **illustrative, fictional example**, not a real restaurant dish or a dietary assessment.

Locked input.

Controlled generation request.

> Using only the locked input, write two one-sentence descriptions: a concise QR version and a slightly fuller print version. Do not add ingredients, methods, origin, dietary status, allergen status or quality claims. Preserve the dish name separately from the description.

Draft variants.

**QR draft:** Charred courgette, whipped feta, mint and lemon oil on flatbread.

**Print draft:** Flatbread with smoky charred courgette, salty whipped feta, mint and bright lemon oil.

Both drafts stay inside the record. The QR version prioritises rapid identification. The print version uses only the approved flavour cues and gives them clear referents.

Edit log and gate.

| Check | Finding | Action | |---|---|---| | Ingredient match | All named components appear in the locked input | Keep | | Method match | “Charred” applies only to courgette | Keep | | Flavour support | Smoky, salty and bright are approved cues | Keep for print | | Unsupported claims | None introduced | Keep | | Dietary/allergen wording | No label appears, but status is still pending | Do not use this exercise as dietary/allergen approval | | Format | QR is leaner; print carries more sensory detail | Preview both in their real layouts |

**Editorial selection:** use the QR draft for the mobile placement and the print draft where the layout allows it, subject to the restaurant completing its separate operational approval. If the final recipe or supplier information changes, reopen the locked record and review every derived version.

• **Dish name/category:** Charred Courgette Flatbread / small plate
• **Verified composition for this exercise:** flatbread, charred courgette, whipped feta, mint and lemon oil
• **Verified preparation:** courgette is charred; other methods are not specified
• **Included items:** all five listed components are included
• **Approved flavour cues:** bright, salty, smoky
• **Dietary/allergen status:** pending; publish no dietary or allergen label from this exercise
• **Prohibited claims:** vegan, gluten-free, house-made, local, healthy, wood-fired, award-winning
• **Voice:** direct, relaxed, ingredient-led; no “mouth-watering”, “perfect” or “best”
• **Format controls:** QR draft shorter than print; one sentence; add no facts

Implementation steps for a full menu

1. **Assign ownership.** Name the person who maintains recipe facts and the person authorised to approve menu copy. They may be the same person, but the responsibility should be explicit. 2. **Build source records.** Complete the matrix for each dish before generating text. Resolve obvious conflicts between recipe, prep and service notes first. 3. **Set a voice card.** Define a few observable rules: sentence style, permitted flavour vocabulary, treatment of unfamiliar terms and banned hype. “Friendly” alone is too vague. 4. **Set format ceilings in the actual design.** Decide how much copy fits the QR card and print layout. Use those limits in the request instead of asking for “short” copy. 5. **Generate controlled variants.** Work from one dish record at a time. Instruct the tool to use only supplied facts and to identify missing information rather than completing it. 6. **Reject factual drift.** Compare every noun, modifier and method with the record. Delete or verify additions; do not keep them merely because they sound plausible. 7. **Run the approval gate.** Complete ingredient, dietary, allergen and operational checks through the restaurant’s established process. Record reviewer and date. 8. **Preview in context.** Check line breaks, hierarchy, price separation and scanning on a real phone-sized view and the intended printed page. 9. **Publish from the approved master.** Keep QR and print variants tied to the same source record so a recipe change triggers review in both places.

If you want to place approved descriptions into an editable menu with structured sections, a hosted QR link and print-ready output, [start with MenuCrafters](https://menucrafters.com/). The writing and safety review described here still remain the restaurant’s responsibility.

Decision checklist: generate, revise or hold?

Use this short gate after every draft:

For batch work, review adjacent items together after individual approval. This catches category-level problems such as every dish beginning with the same verb, inconsistent naming of the same sauce, or a tone shift halfway through the menu.

• **Generate** only when the dish record contains verified composition, preparation where mentioned, included items, voice rules and format limits.
• **Revise** when the facts are complete but the copy is vague, overlong, repetitive, off-brand or poorly adapted to the layout.
• **Hold** when a named ingredient, method, provenance statement, certification, dietary label or allergen statement is unverified.
• **Hold** when the generator adds a claim that is absent from the source record.
• **Hold** when QR and print versions disagree about what the guest receives.
• **Approve** only after an authorised person checks the copy and its rendered placement.

Limitations and trade-offs

AI cannot inspect the dish, know whether a substitution occurred, read a supplier label that was not provided, or take responsibility for a customer-facing claim. It may produce fluent wording that is factually wrong. Fluency should therefore increase scrutiny, not reduce it.

A strict locked-input process also has trade-offs. Preparing source records takes time. Tight controls may produce restrained copy, and separate QR and print variants add version-management work. For a very small, stable menu, careful manual writing may be simpler. For safety-sensitive claims, uncertain recipes, frequent undocumented substitutions or disputed supplier information, generation is not appropriate until the underlying records are resolved.

Length adaptation has limits too. A shorter QR description is not automatically clearer: removing an included side or a distinguishing preparation can mislead. Conversely, a large print area is not a reason to add unsupported story, provenance or sensory claims. The approved facts—not available space—set the boundary.

Finally, this framework is an editorial workflow, not legal, medical or food-safety advice. Restaurants must follow the rules and review processes that apply to their operation and location.