You ask Claude for a week of dinners: four people, a nine-year-old with a tree nut allergy, about $150 at the store, nothing over 40 minutes on a school night. The plan comes back in seconds and it's good. It uses up the half bag of spinach you mentioned and leaves the nuts out without being reminded.

Next Sunday you open a new chat. It may remember the allergy. It doesn't know what you made last week, that nobody touched the lentil soup, or that the stomach aches came back on Thursday, because none of that was ever in a chat.

How many people already do this

More than you'd guess. In a KFF poll taken in late February 2026, 32% of US adults said they had used AI for health information or advice in the past year, about the same share who turn to social media for it. Asked why, 65% of them said a major reason was wanting a quick answer.

Pew Research Center asked in June 2026 what people use chatbots for. A quarter of adults use them to work out what's behind a symptom, and 34% use them for at least one health reason. In an earlier Pew survey, people who get health information from chatbots called it highly convenient (48%) far more often than highly accurate (18%).

Those health surveys don't ask about food on its own. A smaller one does. In a March 2026 survey of 366 people by the agency MorganMyers, 41% had used ChatGPT or another AI tool to plan meals in the past month, 34% to change a recipe, and 25% to write a grocery list. OpenAI, by its own count, says more than 40 million people a day ask ChatGPT a health question.

What it brings, and what it can't know

The newest models are good at a lot of this. Three pediatric gastroenterologists rated ChatGPT 5.0's and Gemini 2.5 Pro's answers to 34 questions parents ask about celiac disease, and both averaged at least 4.3 out of 5 for accuracy, safety, completeness and how easy they were to follow (JPGN Reports, 2026). The answers got weaker where the guidelines themselves are thin.

They still miss. In May 2026 researchers put 50 everyday questions about eating with cirrhosis (scarring of the liver) to ChatGPT 5.5, Claude Opus 4.7, DeepSeek V4 Pro and Gemini 3.1 Pro. Every model gave some answers the raters flagged as potentially unsafe, between 6% and 12% of them (Nutrición Hospitalaria, 2026).

What helps most is what the model is given. When ChatGPT-5 estimated calories from photos of 195 dishes, its estimates got steadily closer as it was told more: a description of the dish, then the ingredients with amounts (Nutrients, 2025). And in an October 2026 study that hasn't been peer reviewed yet, GPT-5.4 planned meals for 300 people with rules such as allergies and foods they don't eat. Writing plans on its own, it met 87.5% of those rules. Built into a planner that checked each plan against the person's rules, the same model met all of them (arXiv, 2026).

What no model can supply is your household. The newest assistant doesn't know what Sam ate on Tuesday unless something tells it. The companies know this. In January 2026 both OpenAI and Anthropic started letting people connect health records and apps like Apple Health to their assistants, betting an assistant does better when it can see your actual history.

For meals, that history comes in two parts: the rules for who's eating, and the record of what happened. You can hand over the first part yourself tonight.

Give it the house rules once

ChatGPT and Claude both let you group chats into a project and write instructions that apply to every chat inside it (ChatGPT's help page, Claude's). Both offer projects on free plans, and Claude's free plan allows five. Make one called Dinners, or whatever you'll find again, and paste in something like this with your own people and rules:

Our household: two adults, Sam (9) and Ava (5).
Sam is allergic to tree nuts. That includes almond milk, almond flour, nut butters, pesto and most granola. Ava doesn't eat dairy right now. Butter and cheese are fine for the rest of us.
I keep my salt down on my doctor's advice. Don't salt my portion, and point out salty ingredients like soy sauce, stock cubes and cured meats.
Groceries are about $150 a week, one shop on Saturday. School-night dinners take 40 minutes or less. Sunday has time for one big batch.
When you suggest a meal, list every ingredient, including what's in the sauce, the stock and the bread. For anything packaged, remind me to check the label for nuts.
If you're not sure a food is safe for someone, say so. Don't guess.
Last week: nobody ate the lentil soup. Sam liked the chicken tacos.

Change the last line on Sunday. Everything else stays put, and every new chat in the project starts with it.

A few habits that still matter

Read the label yourself. The assistant hasn't seen the jar you buy. "May contain" and shared lines are on the package, not in the chat.
Take your numbers from your doctor or dietitian for a diet they set (for kidneys, blood pressure, or blood sugar on medication), and use AI for recipes that fit inside them. Their numbers are about you. A chatbot's are about people in general.
Get the plan out of the chat. Put it wherever your week actually lives: a calendar, the fridge door, a shared note. Next week's chat will bury this one.
Keep track of what happened, not only what you planned. What got eaten, what came back to the kitchen, how people felt after. An assistant can only point out what tends to go together in a record it can see.

Number 4 is where most of us stop. Project instructions hold the rules. They don't hold the history, and pasting a week of meals and stomach aches into a chat every Sunday is not a habit anyone keeps for long.

Your assistant, with your history

That second half is what Meal Possible is built to hold: who's eating and what each person can have, your recipes, the meal plan, the grocery list, and the food, symptom and mood logs. Through MCP (an open standard that lets an AI assistant reach an app you choose), you can connect ChatGPT, Claude, or another assistant to your account and keep doing the thinking where you already do it. You tick what it may see: recipes and meal plans are one choice, and your food log, symptoms and journal are separate ones.

Then the conversation changes. Ask Claude whether Sam can have almond milk, and it checks instead of answering from general knowledge. The answer is Allergy, because tree nuts are on his list, from the same rules the app's own "can I eat this?" page uses. Ask it to fill in Monday to Wednesday, and the meals land in your planner, each picked to fit every eater's allergies and diet stage, with the grocery list built from them. Ask what everyone ate the week the stomach aches came back, and it reads the log instead of your memory, then helps you think through what to try next.

If you connect it, add a few lines to the project instructions so the assistant knows to use it:

You're connected to Meal Possible. It has our profiles, allergies, recipes, meal plan and food log.
Before you tell me a food is safe for someone, check it in Meal Possible for that person.
When we settle on meals, put them in the Meal Possible planner instead of only listing them here.
When I ask what happened on a day or a week, read the log rather than guessing from this chat.

What it won't do: the connection doesn't make the assistant's own general advice right. When it doesn't check, it answers from what it knows, which is why those lines are worth adding. The food check covers the foods it has rules for and says "we don't have a rule for this yet" rather than guessing. Some AI apps want their own paid plan before they allow a custom connector.

Whatever your assistant reads becomes part of your chat with that company, so tick only what you'd be comfortable typing into it yourself. We don't train AI models on members' data and we don't sell it. We email you whenever a new connection is approved, and you can cut one off at any time. You don't need an outside assistant at all, either. The app has its own, and it already knows your recipes and your log.

If you do one thing tonight, make the project and paste in your version of the rules. It takes ten minutes, and every chat after that starts out knowing who's at your table.