People treat food preferences as a static list of likes and dislikes, when they're actually a context-dependent, evolving system of rules, exceptions, and contradictions that only emerge through real-time capture.
Michael Tiffany
Our family can't eat most of what's available in supermarkets, social gatherings, and restaurants. We are severely gluten intolerant, we avoid carbohydrates and lectins, and we've spent years building a home food system around local sourcing, modular meal design, and a weekly cooking rhythm that the kids can execute. I mention all of this because when I first tried to teach an AI agent about our food, I did what everyone does: I typed up a summary. "We're gluten-free, keto-adjacent, we buy our meat from a local farm, we cook from scratch." The agent responded with perfectly reasonable suggestions that I would never in a million years actually cook.
It recommended cauliflower rice stir-fry. We eat cauliflower, sure, but we have never once riced it and stir-fried it. It suggested almond flour pizza. We use almond flour, but for a very specific breading on chicken nuggets that our kids will eat, and the idea of stretching it into a pizza crust is a project I'd file under "food experiments for fun," not "what's for dinner on a Wednesday." The agent knew the constraints but had no idea about the practice: the actual dishes we rotate, the brands we trust, the textures our kids accept, the cleanup time we're willing to tolerate, the specific rhythm of a week where Saturday is steak with rutabaga and Friday is lamb kofta with salad.
That gap between constraints and practice is where all the interesting teaching happens, and it can't be done with a profile. It requires a feedback loop.
This article is about building that loop: how to structure what your agent needs to know in the right priority order, how to capture the messy, contradictory, context-dependent reality of what you actually eat, and how to test whether the agent has crossed the line from retrieval to something that starts to look like comprehension.
Before you build the loop, you need to give your agent a foundation, and that foundation has an order of priority that matters a great deal. I think of it as three tiers, ordered by severity of failure.
The first tier is what can send you to the emergency room. True IgE-mediated food allergies are binary: trace exposure to the wrong protein can cause anaphylaxis. A 2024 CDC survey puts diagnosed food allergy prevalence at about 6.7% of U.S. adults, but the number understates the operational complexity, because a single peanut allergy in one family member restructures every meal, every restaurant choice, and every school lunch for years. Tell your agent about these allergies the way you'd brief an anesthesiologist: specific allergen, mechanism, severity, and what happens on exposure. Then verify by asking for a dozen meal suggestions and auditing every ingredient. If it slips, correct and retest. Your agent is a planning aid, not a medical device; you still carry the epinephrine and read the labels yourself. But an agent that doesn't know your allergies is worse than useless, because it will confidently suggest meals that could hurt you.
The second tier is what makes you feel terrible for days. Intolerances operate through different mechanisms and, critically, are dose-dependent. Cleveland Clinic estimates roughly 20% of the U.S. population deals with some form of food intolerance. I know from personal experience that a gluten exposure doesn't just mean stomach pain; it means weeks of mood disruption, nightmares, anxiety, and a rash that won't quit. Celiac disease sits in a strange middle ground here: technically autoimmune, not IgE-mediated, but operationally as zero-tolerance as any allergy. Your agent needs to understand where your particular intolerances fall on the spectrum from "a little lactose in my coffee is fine" to "even labeled gluten-free oats are not safe for us." That's information you can state up front, but the boundaries will sharpen over time through the feedback loop.
The third tier is what you actually enjoy. This is the vast, shifting, contradictory territory that no profile can capture. You eat raw oysters at restaurants but would never buy them for home. You cook with fish sauce constantly but find the smell of it in other people's cooking unbearable. Your kids will eat breaded chicken nuggets but reject the same chicken if it's visibly herbed. These aren't contradictions, they're context-dependent preferences, and they only become legible through observation.
The mechanism is low-tech and the habit is small: after you eat, you talk about it.
Record a thirty seconds to a minute long voice memo, while the meal is still vivid. You're not composing a review, you're leaving a message for the agent that will plan your meals next week. Say what you ate, what worked, what didn't, and who else was at the table. The reason voice works better than typing is the same reason engineers prefer incident post-mortems right after the outage rather than a week later: you capture details you'd never bother to write down, because talking is cheap and the specifics are still in working memory.
Here's what a useful memo sounds like: "Wednesday dinner. Bolognese with polenta. Used the Hazan-adapted recipe, 2.5x batch with the Pomi strained tomato and macadamia cream. Added dried porcini and fish sauce for umami. Polenta was the low-carb almond meal version. Everyone ate it, which is notable because the regular polenta has been getting rejected lately; I think the almond meal gives it a texture that works better. Nutmeg was too strong this time, dial it back to a quarter teaspoon. Plenty of leftovers for tomorrow's lunch. Cleanup was one pot plus the Vitamix, maybe twelve minutes."
Transcribe and paste that into your agent. A memo might contain: a specific recipe with quantities, a brand preference (Pomi), a substitution that worked (macadamia cream), a flavor note worth remembering (too much nutmeg), a textural observation, a logistical fact (leftovers cover lunch), and a cleanup assessment.
After a week or two of this, ask your agent to plan three dinners as a test. Compare what it suggests to what you'd actually cook. When it suggests something wrong, correct it out loud, and feed that correction back in: "You suggested a stir-fry, but we almost never stir-fry at home because the oil splatter is a nightmare to clean and the smoke sets off our alarm. We sauté with a lid on, which gets a good sear while containing the mess." That correction teaches the agent something a profile never could, because you'd never think to write "we don't stir-fry" in a settings panel.
Your food preferences don't exist in isolation if you're cooking for a household. "What's for dinner" is a constraint satisfaction problem: your macros, your partner's intolerances, your kids' ever-shifting texture aversions, the contents of your fridge, your energy level after work, and how much cleanup you can face before bedtime. One voice memo after a family dinner captures preference data for everyone at the table.
We'll cover teaching your agent about the people in your life in a later article, but you can start now by narrating multi-person reactions: "Made the kofta. I had mine wrapped in lettuce with tahini. My partner added rice. The kids ate the meatballs plain, which is fine, because the recipe hides organ meat and collagen in there anyway. One of them asked for seconds, which almost never happens with lamb." That single memo maps different serving configurations to the same base protein, captures a hidden-nutrition strategy, and notes an unusual signal of approval from a picky eater.
Here's a protocol if you want to measure this systematically.
Day 0. Give your agent a conventional food profile: dietary restrictions, allergies with severity levels, cuisine preferences, a few dislikes. Ask it to plan three days of dinners and take a screenshot of the output. This is your baseline, and I predict it will be adequate but uninspired.
Days 1 through 14. Capture at least one voice memo per day after a meal. Aim for variety: home-cooked, restaurant, takeout, the sad desk lunch, the snack you ate standing in front of the open fridge. When something triggers an intolerance reaction, note the specific food, the approximate quantity, and when symptoms appeared. Transcribe everything and feed it to the same agent.
Day 15. Use the same prompt as Day 0 and ask your agent to plan three days of dinners again. Compare the two outputs and ask yourself: Would I actually cook any of these? Did it remember the hard constraints without exception? Did it pick up on the specific brands, stores, and preparation methods I mentioned? Did it learn which dishes the whole household will eat versus which ones are just for me? Did it suggest anything I hadn't explicitly mentioned but would genuinely enjoy?
That last question is the real test. Parroting your words back at you is retrieval but predicting an unstated preference from the pattern of your reactions is comprehension. If the agent knows that you like the jasmine rice from the Asian grocery but not the Mahatma from the supermarket, that's retrieval. If it infers that you probably care about rice provenance in general and suggests trying a local farm's short-grain, that's the beginning of something useful.
If it's still generic after two weeks, your reactions might be too vague. "Dinner was good" teaches nothing. "The salmon was overcooked because I left it in the oven too long, but the asparagus was perfect; I steamed it until bright green and hit it with lemon, olive oil, and a little flaky salt" teaches five things. If the suggestions are too narrow, you may be over-indexing on complaints. Make sure to capture what you loved, not just what went wrong.
Do I need to record every meal? Once a day during the experiment, and after that, only when something notable happens: a new recipe you tried, a restaurant discovery, a dish that bombed, a food that triggered a reaction. You're building a representative sample, not a food diary.
Should I trust my AI with allergy information? You should tell it, because an agent that doesn't know can't help you avoid exposure. But never outsource vigilance; the agent screens recipes and menus, you read labels and carry epinephrine and talk to the people cooking your food.
Does this work with any AI tool? Any tool that maintains memory or long-running context: ChatGPT, Claude, Gemini, a local model with persistent state. The key requirement is continuity; your agent needs to accumulate your reactions, not start fresh every conversation.
Is it safe to share food and health data with an AI? Allergy and intolerance information is health data. Review the privacy policy of your tool, understand where your data lives and whether it's used for training, and decide based on your own risk tolerance. Voice memos can be transcribed on-device so the raw audio never has to leave your phone.
How long until my agent actually knows my food preferences? Food knowledge is a long game. Your preferences shift with the seasons, your intolerances change as you age, your household composition evolves, and the agent's model of you should evolve with all of it. The feedback loop doesn't end, it just gets less frequent as the agent's accuracy improves.
Pick one meal today and leave yourself a 30-second voice memo afterward. Say what you ate, what you thought, and what you'd change. Transcribe it and give it to your agent. If you have known allergies or intolerances, brief the agent on those first, with the specificity you'd give a doctor.
