How to Turn a Restaurant Menu Photo Into Allergy Questions

Anyone dining out with allergies can turn a restaurant menu photo into allergy questions in 8 minutes: upload the card, force 5–7 guest-ready questions about allergens and dietary limits, and mark UNCLEAR when the print is fuzzy. Run it in ChatGPT, Claude, or Gemini vision before the model invents ingredients the card never printed.

A menu photo allergy questions pass is a guest checklist for the server, not a medical diagnosis and not a rewritten menu. If the model invents cashew cream that is not on the card, you never did the job. Printed words plus UNCLEAR beats a confident fake recipe.

This is not a shopping list from a recipe photo. A shopping list is for a kitchen. This is for a table conversation. It is not receipt photos into an expense list. Expenses are money lines. Allergy questions are safety lines. It is not a whiteboard photo into a checklist. A whiteboard checklist is tasks. This list is questions a guest can ask out loud.

What you need

One clear photo of the printed menu page you will order from. Your allergy or dietary constraints in plain nouns (peanut, shellfish, gluten, dairy, sesame, egg). ChatGPT, Claude, or Gemini with vision. Eight minutes and one cold read with a friend who did not write the prompt.

Banned before you start: inventing sauces not printed, diagnosing severity, rewriting the chef’s recipe, “safe for everyone,” guessing hidden oils, skipping UNCLEAR on blurry lines, and turning the output into a shopping list.

Turn a menu photo into allergy questions

1. Upload the photo. Paste only printed constraints

Attach the menu photo. Paste your constraints as nouns only: peanut, tree nut, shellfish, gluten, dairy, sesame, egg, or UNKNOWN if you are asking for a guest. Do not paste a medical essay.

Instruct ChatGPT, Claude, or Gemini: “Read only what is printed on this menu photo. Do not invent ingredients. Do not diagnose. Output 5–7 short questions a guest can ask the server about allergens and dietary limits I paste. If a line is blurry or unreadable, write UNCLEAR for that dish and still ask a clarifying question. Then stop.”

Good constraint paste: “peanut, shellfish, dairy.” Bad: “I might die if I eat anything weird.” Drama is not a label the model can check against ink.

2. Force five to seven guest-ready questions

Shape each line as a question the guest can say: “Does the pad thai use peanut oil or crushed peanut?” “Is the risotto finished with butter or parmesan?” Cap at seven. More than seven is a deposition, not a table ask.

Instruct: “Output exactly 5–7 questions. Each question must name a printed dish or section when possible. Ban inventing ingredients not visible. Ban medical advice. Ban rewriting the menu. If text is fuzzy, prefix that dish with UNCLEAR.”

Good: “UNCLEAR on the specials board — can you confirm whether the soup base is shellfish stock?” Bad: “The chef probably uses fish sauce, so avoid everything.” Probably is invention.

Keep questions tied to printed dishes

If the model asks about a dish that is not on the card, cut that line. Vision models love “house specials” that live only in training data. Your job is the ink in the frame.

3. Ban invention. Require UNCLEAR on fuzzy print

Tell the model: “If a word is illegible, write UNCLEAR. Do not guess. Do not fill sauces from memory of similar restaurants. Do not claim a dish is safe.”

ChatGPT will invent a cream sauce because the photo looks Italian. Claude will soften into “you may want to avoid creamy dishes.” Gemini will invent a gluten-free symbol that is not printed. Cut all three gifts.

Hard fail: any ingredient named that you cannot point to on the photo; any “this is safe”; any severity ranking; any rewritten recipe.

4. Cold peer scan before you speak to the server

Hand the question list to one friend who did not write the prompt. They must point to each dish name on the photo. If they cannot find it, delete the line.

Open your phone notes. Read the questions out loud in under sixty seconds. If you sound like a lawyer, shorten. If you sound like a food blogger inventing techniques, delete.

5. One human pass at the table, then stop

Ask the server the surviving questions. Log answers next to the dish names. Do not let the model “update the menu” after the fact with guesses.

Check traps: invented ingredients, missing UNCLEAR on blur, medical tone, more than seven questions, shopping-list format. Fix by regenerating only the broken lines: “Replace only the lines I mark. Do not invent. Keep UNCLEAR.”

One pass. If the list still invents cashew cream, you never turned a menu photo into allergy questions. Stop.

What good looks like

Trigger: You are about to ask “is this safe for my allergy” with no dish names and no printed check.

Input: Menu photo, constraint nouns, ChatGPT/Claude/Gemini vision, cold peer scan.

Output: 5–7 short guest questions tied to printed dishes, UNCLEAR on fuzzy lines, zero invented ingredients, zero medical claims.

Stop: After one cold peer scan and one table ask. No recipe rewrite. No “safe for everyone.”

Treat the question list like a script for thirty seconds of speech, not a research memo. Servers answer short asks. Long asks get shrugged. If a line needs a footnote, it is not table-ready. Keep dish names as they appear on the card, including odd spellings. Correcting the restaurant’s spelling in your question is how the server looks at the wrong line. When vision models “helpfully” translate dishes into English approximations, restore the printed spelling before you speak. The cold peer scan exists to catch invented sauces and missing UNCLEAR tags. Skip that scan and you will ask about ingredients that never existed. That wastes trust at the table.

Takeaways

- Upload the card. Paste constraint nouns. Force 5–7 guest questions only.

- Ban inventing ingredients. Require UNCLEAR when the print is fuzzy.

- Tie every question to a dish you can point to in the photo.

- One cold peer scan. Then ask the server. Stop inventing after the fact.