How an AI recipe generator actually writes a cookable recipe

Photo, sentence, or a pile of ingredients

Open Capture. Photograph a plate, a fridge shelf, a menu, or an ingredient pile. You can also type a sentence such as "weeknight chicken thighs, 30 minutes, one skillet" or share an image from another app. The model first identifies the dish or the constraint set, then asks for missing pieces: serving count, oven versus stovetop, whether you own a food processor. Identification is a classification job. Generation is a synthesis job under many simultaneous constraints. Those are different failure modes, which is why the app keeps them as separate steps you can correct.

A photo of birria tacos is not a recipe. It is evidence of a braise, a consume, a tortilla, and a garnish. The generator has to invent a method that a home kitchen can run. If your Taste Profile says no dairy, the consume cannot quietly include crema. If you marked wheat as an allergy, flour tortillas get swapped, and the swap has to be named in the ingredient list, not hidden in a tip. That is the difference between a chatbot paragraph and a contract-validated recipe: title, cuisine, difficulty, servings, prep and cook times, equipment, ingredients with quantity and unit, ordered steps with action and temperature, nutrition per serving, allergens, and storage.

Text input is often cleaner than a photo when you already know the dish. "Dal tadka, two burners, 40 minutes" gives the model a cuisine and a time box. A blurry restaurant photo of the same dal still works, though you should confirm the identification before you spend credits on a full generate. If the dish name is wrong, tap to correct it. Generation from a wrong title wastes a credit and produces a recipe you will not cook.

Taste Profile is the constraint file, not a vibe

Onboarding stores diet, allergies, dislikes, preferred cuisines, skill, equipment, default servings, time budget, spice tolerance, and unit system. Every generate call reads that file. If you own a 12-inch skillet and a 400 °F oven, the method should not call for a 6-quart Dutch oven you do not have. If your time box is 25 minutes, the recipe cannot start with a 90-minute braise and a cheerful "or use leftovers." The time field is a hard cap on total active plus passive cook time as written.

Allergies are name-matched against a synonym table, case-insensitive and locale-insensitive. Peanuts, tree nuts, milk, eggs, fish, shellfish, soy, wheat, sesame, plus any custom string you add. A generate that would place an allergen in the ingredient list is blocked before you see the card. That gate does not make the recipe medically safe. Soy sauce, oyster sauce, and many bottled marinades hide wheat. Mustard and celery hide in spice blends. The model can miss those. Read the bottle. If the label lists wheat, barley, rye, or malt, do not cook the step as written. Swap the bottle or stop.

Skill changes verb choice. A beginner recipe names the fond (the browned bits stuck to the pan) and tells you to deglaze with a measured splash of water or stock. An intermediate recipe can say "deglaze" and assume you know the sound of a wet pan hitting hot metal. Spice tolerance changes chili form. A 1 to 5 slider in the caption is too coarse. Kashmiri chili, gochugaru, and cayenne are not interchangeable by heat number. The generator should pick one and give a gram amount.

Cook Mode is the recipe with wet hands

Once the card exists, Cook Mode keeps the phone awake, shows one step at a time, and runs timers that survive a lock screen. You can scale servings, toggle grams and cups, and check off mise en place before the first flame. The scaler is arithmetic on the ingredient list, then a rewrite of any step that names a pan size or a bake time that does not scale linearly. Doubling a stir-fry does not double the wok time. The fond still needs a minute. The oil still needs to shimmer.

Finish Cook Mode and the app offers four next actions: photograph what you cooked, check in the cook streak, log the plate to the diary, and rate the recipe. The photo goes to the Cook Journal. The log uses the recipe's per-serving nutrition as a starting point, then lets you say how much of the pot you ate. If you cooked four servings and ate one and a half, the diary should show 1.5 times the per-serving numbers, not a round 1. That sounds obvious. Many calorie apps round you into a fiction.

Timers matter more than the prose. Pasta water at a rolling boil, a 3-minute rest on a chicken thigh, a 12-minute air-fryer cycle. Cook EZ attaches those durations to the step, not to a footer. If a step says 165 °F in the thickest part of poultry, that number comes from USDA leftover and poultry guidance, not from a lifestyle blog. The USDA Keep Food Safe page is the source for the cold holding, hot holding, and two-hour rules you will also see in storage notes.

Nutrition on the card, then the diary

Per-serving calories, protein, carbohydrate, and fat are estimates assembled from USDA FoodData Central entries and, for packaged items, Open Food Facts. FoodData Central is the U.S. nutrient database at fdc.nal.usda.gov. Open Food Facts is the crowdsourced product catalog at world.openfoodfacts.org. Neither is a lab assay of your pan. Oil absorption, trim waste, and the difference between 80 percent and 93 percent lean ground beef will move the number. Treat the card as a planning figure. Treat the diary as a log you can edit.

Barcode logging is free. Scan a yogurt, a frozen vegetable mix, a loaf. Cook EZ looks up the GTIN in a cache first, then Open Food Facts if needed, and writes the serving to breakfast, lunch, dinner, or snack. You can also search the local food index or quick-add a gram amount. Photo logging of a mixed plate is a separate, credit-using path: the model segments items, you correct portions with a slider, and the macros recompute live. Use the barcode when the package exists. Use the photo when the plate is homemade or restaurant food with no GTIN.

The FDA Nutrition Facts label page explains serving size, dual-column labels, and the percent Daily Value. If you are comparing a generated recipe to a boxed meal, read the label's serving size in grams, not the marketing "bowl." A recipe that claims 520 kcal per serving and a box that claims 320 kcal are often describing different grams of food. Convert both to per-100-gram before you argue with the diary.

Maillard, fond, and carryover cooking still apply

An AI recipe that skips browning because browning takes two extra minutes will taste like steamed protein in sauce. Maillard reactions (the brown crust on meat, the toasted flavor on bread, the color on onions past translucent) need a dry-enough surface and heat typically above about 140 °C / 284 °F at the interface. Wet crowding in a cold pan makes steam. The generator should say "pat dry," "do not crowd," and "wait for the fond." If it does not, add those lines yourself before you cook, or regenerate with a note that you want a hard sear.

Carryover cooking is residual heat. A 2-inch pork chop pulled at 140 °F can climb several degrees on the board. USDA FSIS still wants leftovers reheated to 165 °F. Those two facts live in the same kitchen. The recipe should tell you when to pull, how long to rest, and whether the rest is under foil. A 3-minute rest on a thin cutlet is mostly juices redistributing. A 10-minute rest on a roast is carryover plus juice. Mixing those up dries the meat or leaves the center under the safe temperature for the audience you are cooking for.

Storage notes should follow FSIS leftover guidance: refrigerate within 2 hours (1 hour if the room is above 90 °F), use cooked leftovers within 3 to 4 days, reheat to 165 °F. The FSIS leftovers page is the primary source. A generator that says "keeps a week on the counter" is wrong, and you should discard that card. Cook EZ's storage field is meant to quote those windows, not invent new microbiology.

What the model still misses

Wheat in sauces is the most common miss. Tamari is often gluten-free. Many "all-purpose" soy sauces are not. Hoisin, oyster sauce, malt vinegar, soy-based marinades, and imitation crab can carry gluten. The allergen banner on a generated recipe is a name-match against the ingredient list the model wrote. If the model wrote "2 tbsp soy sauce" and did not specify a gluten-free bottle, a wheat-allergic cook has to choose the bottle. Verify the label every time. Do not outsource that to a language model.

Equipment lies are the second miss. "Blend until smooth" on a 400-watt stick blender versus a 2-horsepower Vitamix is a different recipe. "Air-fry at 400 °F for 12 minutes" on a basket that holds two chicken thighs is not the same as a toaster-oven air-fryer with a sheet pan of six. If your Taste Profile lists a 5.8-quart basket air fryer, the batch size should fit. If the card still says "in a single layer" for 2 pounds of wings, split the batch. The model is optimistic about pan real estate.

Cuisine authenticity is a third miss, and it is a taste issue rather than a safety issue. A generated "mapo tofu" that uses sweet Italian sausage and heavy cream is a different dish. You can still cook it. Call it what it is. If you wanted Sichuan mapo, regenerate with "doubanjiang, no dairy, medium heat." The Taste Profile cuisine chips bias the generator. They do not certify a regional standard.

When to generate, when to cook from memory

Generate when the constraint set is messy. Allergies plus a 25-minute cap plus a missing ingredient plus a toddler who will not eat cilantro. Memory recipes fail that matrix. Generate also when you are staring at a photo of a dish you cannot name. Identification plus a cookable method is the product. Do not generate a scrambled-egg recipe you have made a thousand times unless you want a scaled grocery list or a diary-ready nutrition card.

The Dietary Guidelines for Americans are the federal nutrition policy document, not a recipe book. They talk in food groups and patterns. An AI recipe generator can aim a week of dinners at those patterns (more beans, more seafood, less added sugar) if you say so in the prompt. It cannot replace a clinician. If you are managing diabetes, CKD, or a diagnosed allergy, the generate is a draft. Your clinician and your labels are the authority.

Credits are spent on identification and on full generation. Correcting a dish name before generate is cheaper than regenerating after a wrong cuisine. If the first card is close, edit a single step (swap the chili, cut the salt, change the starch) rather than rolling the whole recipe. Edited steps stay on the card you already paid for.

Questions

Sources

• USDA, Keep Food Safe
• USDA FSIS, Leftovers and Food Safety
• USDA FoodData Central
• FDA, Nutrition Facts Label
• Open Food Facts
• Dietary Guidelines for Americans