Build an AI Nutrition Assistant That Plans Real Life

My nutrition assistant knows I'm eating out Thursday before I've thought about what that means for Thursday.

It sees the reservation on my calendar, builds the rest of that day's plan around it, and leaves space for the martini it correctly assumes I'm having. That's the difference between an AI meal planner and an AI nutrition assistant: one generates recipes, the other plans around your actual life.

Here's the full setup. (The one-sentence version: an AI nutrition assistant is a dedicated, persistent AI chat that plans your meals around your goals and your actual calendar, keeps your food log because you told it what you ate, and adjusts the rest of the day in real time.)

One note before the how: this is an organizing tool, not a medical one. It counts, logs, and plans. It doesn't diagnose, and it doesn't replace a professional, which is a point I'll come back to, because mine works with my human nutritionist, not instead of her.

What it's connected to

My Oura ring, for sleep, readiness, and workouts
My calendar, so it knows when I'm eating out
A running log of everything I eat, which it maintains itself

The wearable and calendar connections are the built version. If you're starting simpler, you can paste in your week and your workouts by hand and get most of the value; the log is the part that matters most, and that works in a plain project from day one.

The weekly rhythm

Every Sunday, the week's meal plan lands in my Telegram, built on my goals, my macros, and what it knows I actually like to eat. I've been training it for about six months, so at this point it knows the difference between what I aspire to cook and what I'll actually make on a Wednesday. I review the plan, make changes, and swap in anything new I want to try.

Because it reads my calendar, the plan already accounts for real life. A dinner reservation isn't a deviation it scolds me about later; it's in the plan from Sunday, with the day adjusted around it.

During the day, it keeps the log. It reminds me to log my meals. I can send a picture of a plate or a nutrition label, but usually I just tell it what I ate. No app, no barcode scanning, just a sentence.

And it looks forward, not just back. After lunch it might tell me I'm running light on protein and to keep an eye on that. It knows the protein shake I like and when suggesting it actually helps. If I'm headed out to eat, I send it the menu and ask what to order based on what I've already had that day.

That forward-looking part is what a tracking app can't do. An app tells you what you ate. An assistant tells you what to do about dinner.

The instructions that make it work

The build path is the same as any assistant: make the AI interview you, then have it write its own instructions. The one-file assistant guide has those prompts. What's specific to a nutrition assistant is the context you load:

Your goals and your macro targets, in numbers
What you actually like to eat, and what you'll realistically cook on a weeknight
Your training schedule, or the wearable that knows it
The non-negotiables: allergies, foods you won't touch, the things every plan must respect

To get the context out of your head, run a tailored version of the interview:

I want to build you into my nutrition assistant. Interview me, one question at a time, about: my goals and macro targets, what I actually eat on a normal week, what I refuse to eat, my allergies and non-negotiables, how I cook on weeknights versus weekends, how often I eat out, and what my training week looks like. Then write your own instructions from my answers, and tell me what you'd still want to know.

Then give it the standing behaviors:

Every Sunday, generate my meal plan for the week based on my goals, my macros, and what you know I like to eat. Check my calendar first and plan around any meals out. During the week, remind me to log my meals, keep the running log, and tell me when I'm drifting from my targets while there's still time to adjust. If I send you a menu, recommend what to order based on what I've already eaten today.

What the day-to-day actually looks like

Two exchanges, so you can see the register you're building toward. These are templates of the pattern:

You, 1:40pm: logged. lunch was the chicken bowl from the salad place,
double protein, no dressing

It: Logged. You're tracking well; protein's a little behind for a
training day, so tonight's plan already covers it. Nothing to change.

You, 5:15pm: change of plans, going to [restaurant] for dinner.
here's the menu [photo]

It: Fun. Based on today: the branzino or the strip steak both fit,
skip the pasta side for the greens, and if you're having a drink,
you have room for one. Tomorrow's plan doesn't need to move.

The register to train toward: specific, brief, zero guilt. If an exchange ever ends with you feeling scolded, fix the tone line in the instructions that same day.

Two extensions once the basics run:

The grocery list. Once the Sunday plan exists, have it generate the shopping list from it, organized by store section. The plan and the list should never be two separate chores.
The travel week. Tell it when you're traveling and ask for a lighter version: restaurant strategies instead of recipes, and a "good enough on the road" target instead of the normal one.

When it gets things wrong

It keeps planning meals you'd never make. Feed it a list of your actual last two weeks of dinners and say "this is reality, plan for this person."
It forgot your allergy. Anything that must never happen goes in the standing instructions, never just in chat. Corrections in chat evaporate; instructions are permanent.
The macro math seems off. Ask it to show the math per meal. It makes arithmetic errors confidently, and asking for the breakdown catches them.
The plans are perfect and joyless. Tell it the ratio you want: "plan 80% for my goals, 20% for the fact that food is one of my favorite things."

Where the human fits

I also work with a human nutritionist, and the assistant makes her better, not redundant.

Because it logs everything, she has a live food log she can check throughout the week instead of a recall exercise at our next session. Her expert guidance sits on top of what the AI handles day to day. An expert has limited hours; the assistant has none of her judgment. Together they cover what neither does alone.

If you work with a nutritionist, a trainer, or a doctor, build the assistant as their data layer, not their replacement. And if an AI's food suggestions ever conflict with what your professional told you, the professional wins. Every time.

Common mistakes

Treating it as a recipe generator. The value is the log and the calendar awareness. Recipes are the garnish.
Skipping the training period. The first few weeks of plans will include things you'd never eat. Correcting it's the work, and it compounds; six months in, mine barely misses.
Letting it scold. If the tone drifts into food guilt, fix the instructions immediately. You want a planner, not a judge, and tone is always yours to set.
Feeding it a goal you didn't choose. It will optimize whatever you give it. Give it targets you and your actual life agreed on, ideally with a professional in the loop.

FAQ

Is this medical advice? No, and it shouldn't pretend to be. It's logging, planning, and arithmetic. Diagnosis, conditions, and clinical guidance belong to professionals.

Do I need a wearable? No. The Oura connection makes the readiness picture automatic, and telling it "slept badly, skipped the workout" does the same job manually.

What about weight tracking? Track whatever you and your professional decide matters. The assistant handles whatever inputs you choose to give it.

Can it really read a menu? Send a photo or a link and ask what fits your day. It's one of the most genuinely useful moves in the whole setup.