Header image

AI That Actually Makes Dinner Easier: What Building ThisWeekEats Taught Me

Most AI demos look amazing in a tweet and die in real life.

They summarize PDFs you didn’t need summarized.
They write emails you didn’t need written.
They “brainstorm content” you’ll never post.

Meanwhile, the stupid question that has sucked time and energy out of families for decades is still there:

“What do you want for dinner?”

ThisWeekEats is my attempt to kill that question for my family.

Not with some “optimize your macros and gut health and longevity” fantasy.
Just with real meals we’ll actually cook, that fit our constraints, with a shopping list that doesn’t make us want to scream.

This is what building that taught me about AI that actually helps.

1. The Real Problem Is Not Recipes — It’s Decision Fatigue

Every family says the same thing:

“We’re tired.”
“We don’t want to think about it.”
“We’ll just DoorDash something.”

It’s not that the world is short on recipes.

It’s that every single week you have to juggle:

What’s in the fridge
What people like
What they’ll actually eat
Dietary goals (weight loss, maintenance, heart health, whatever)
Time constraints (Tuesday soccer, Thursday late meetings)
Cost

And you have to do it over and over and over.

The pain isn’t “finding a fun recipe.”
It’s having to re-solve the “What’s for dinner this week?” puzzle endlessly.

So ThisWeekEats started from one goal:

“Make good-enough weekly meal plans and grocery lists in minutes, not hours.”

Not perfect. Not Instagrammable.
Just good enough, fast, and repeatable.

2. Constraints First, Creativity Second

The biggest mistake most “AI meal” ideas make:

They treat the problem as “generate interesting meals.”
They assume you want to discover yourself through food every week.

Real families don’t live like that.

They live inside a box of hard constraints.

In our case, stuff like:

Hard no foods (family “never” list: salmon, cod, chicken, etc.)
Cuisines we actually like (Italian, American, Mediterranean, some Mexican/Caribbean)
Time caps on weeknights
Dietary goals:
- lower carb,
- heart-friendly,
- balanced proteins/fats/carbs,
- nothing insane

So the system had to:

Respect hard no’s 100% of the time
Stay mostly within preferred cuisines
Hit macro / calorie targets across the week
Respect time boundaries (no 90-minute Wednesday experiments)

Only inside that box is AI allowed to be “creative.”

If the model tries to get cute outside those constraints, it’s wrong, no matter how good the recipe sounds.

3. The Job: Five Minutes a Week, Not a Life Makeover

ThisWeekEats has a very small job:

Once a week:
- Generate a realistic meal plan for the family
- Generate a clean shopping list
- Let us tweak a few things without blowing up the whole plan

That’s it.

The success metric is stupid simple:

“How fast can we go from ‘we need meals for the week’
to ‘we have a plan and a grocery list we trust’?”

Not:

“Did you discover Lebanese-Korean fusion?”
“Did your diet become clinically optimized?”
“Did you maximize phytonutrient diversity?”

Concrete constraints beat wellness porn.

If it saves you hours of back-and-forth and wandering grocery aisles, it’s doing its job.

4. How AI Actually Fits In (And What It Doesn’t Do)

Here’s what AI is allowed to do in ThisWeekEats:

Propose meals that:
- obey all hard constraints
- hit target ranges for calories/macros over the week
- fit within prep time bands (quick nights vs long cook nights)
Balance:
- repetition (some recurring hits are good)
- variety (not the same 3 meals forever)
Translate:
- meals → ingredient lists
- ingredient lists → a consolidated shopping list

Here’s what AI does not get to do:

Override hard no’s (“but salmon is healthy!” → still no)
Invent recipes with impossible ingredients or US-grocery-store-unobtainium
Randomly throw in “fun experiments” on nights we marked as “we’re slammed”

The model is basically a constrained generator + planner:

It proposes candidate meals.
Something (code + rules) scores and filters them.
It builds a weekly plan that obeys:
- family constraints
- health goals
- time reality

AI is a component. The rules and constraints are the spine.

Constraints box

5. Feedback From the Only Users That Matter: My Family

You don’t get to hide behind “well, it’s technically correct” when your testers are:

Your spouse
Your kid(s)
And the reality of 6 pm on a Tuesday

Immediate feedback you get from that:

“This recipe is not happening on a weeknight, what were you thinking?”
“Stop trying to make [ingredient X] happen.”
“We liked this one, please keep it in the rotation.”

Practically, that turned into:

An explicit “favorites” concept:
- meals that worked well get more weight in future plans
A “do not show again” kill switch:
- if something bombs, it disappears
More nuance in time categories:
- 15–20 min,
- 30–40 min,
- “weekend only”

AI alone will happily generate “interesting” meals that wreck your actual night.

Reality forced the system to center energy level and cook’s patience, not just “requirements satisfied.”

6. Boring Engineering Matters More Than Model Cleverness

Some of the least glamorous parts of ThisWeekEats are the most important:

Ingredient normalization:
- “red onion” vs “onion (red)” vs “1 small red onion” should not produce 3 separate shopping items.
Unit handling:
- cups, grams, ounces, “1 can,” “1 bunch”
Shopping list dedup:
- combine all the “2 cloves garlic” into something that looks sane
Pantry awareness:
- don’t keep buying staples you already have

None of that is sexy AI.
All of that is where “this is actually usable” lives.

If the AI component is brilliant but the list is chaos, people won’t use it.

The unsexy data cleaning and list logic is half the value.

7. Same Pattern, Different Domain

ThisWeekEats isn’t actually a new pattern for me.

Conductor did the same thing for credentialing and state contracts:
- Took a brutal coordination problem,
- Removed the repetitive admin grind,
- Let humans focus on judgment and relationships.
The 588-person dashboard did it for schedulers:
- Software handled assignment logic and status,
- Humans handled edge cases and real conversations.

ThisWeekEats just turns that lens on my own house:

“Take a recurring, boring decision loop (weekly meals), encode the constraints, and let a system handle 80% of the thinking so the humans can live their lives.”

Different stakes, same idea:

Remove coordination grind.
Respect constraints.
Let humans show up with energy for the parts that actually matter.

8. What This Taught Me About “Good” AI Projects

A few principles that came out of this:

Narrow beats grandiose
- “Make weekly dinner planning take 5 minutes” beats “revolutionize nutrition.”
Constraints are a feature, not a bug
- The more clearly you encode constraints, the better AI performs.
- “Do whatever you want” is the worst instruction you can give a model or a person.
Integrating into a real life loop is the hard part
- The model is not the product.
- The product is:
  - weekly rhythm
  - trust
  - low friction
  - a shopping list that doesn’t suck
Families care more about consistency than novelty
- Some repeat meals are comforting.
- Nobody wants a “brand new experimental dish” three times a week.
The best AI disappears
- The ideal state is:
  - “We barely think about it. We just get our plan and list, and move on with our lives.”

If your AI project demands constant attention, it’s not helping. It’s just a new toy.

9. If You Want to Do Something Similar (Without Building an App)

You don’t need to recreate ThisWeekEats to steal the core ideas.

You can:

Define your constraints clearly:
- foods you won’t eat,
- time bands,
- goals.
Use any LLM to:
- generate meal ideas inside those constraints
- then refine to a week plan
Build a tiny spreadsheet or template:
- to normalize ingredients into a list
- to track favorites vs “never again”

The point isn’t “use my tool.”

The point is:

“Stop letting ‘What’s for dinner?’ burn CPU cycles in your head every single week if AI can do 80% of that work cheaply.”

If AI isn’t making your life easier somewhere as mundane and constant as dinner,
what exactly is the point?

Sometimes the most "advanced" thing you can do with AI is not another dashboard or copilot.

It's making sure you don't have to have the same pointless argument about dinner for the 500th time.

Context → Decision → Outcome → Metric

Context: Weekly meal planning consumed 2-3 hours of decision-making and grocery list chaos. DoorDash spending was climbing. Same "what's for dinner?" conversation every night.
Decision: Built ThisWeekEats as a constrained AI system: hard rules for dietary preferences, time limits, favorites tracking, shopping list normalization. AI proposes within constraints; humans approve and tweak.
Outcome: Weekly meal planning now takes under 10 minutes. Grocery lists are consolidated and usable. Family actually eats the planned meals instead of defaulting to takeout.
Metric: Planning time dropped from 2-3 hours to under 10 minutes. DoorDash spending down 60%. Family meal plan adherence up from ~40% to ~85%.

Anecdote: The Salmon Incident That Drove the "Hard No" Feature

Week two of testing, the AI suggested a beautiful-looking salmon dish. Looked healthy. Fit the macros. My wife looked at the plan and said: "I've told you a thousand times. No salmon. Ever."

She had told me. I'd forgotten to encode it. The AI, of course, had no idea.

That incident created the "hard no" feature. Now the system has an explicit never-list that the AI cannot override, no matter how nutritionally optimal the suggestion is. "But salmon is healthy!" doesn't matter. The constraint is: "We will not eat this. Period."

Sounds trivial. But that's exactly the kind of thing that separates "AI that works in demos" from "AI that works in your actual kitchen on a Tuesday."

Anecdote: The Wednesday Night Meltdown

Three weeks in, the system suggested a 45-minute Thai curry on a Wednesday. On paper, it fit the constraints: within calorie targets, used ingredients we liked, reasonable prep time.

In reality, Wednesday is soccer practice. We have 20 minutes between getting home and needing to leave again. A 45-minute recipe might as well be a 3-hour recipe.

That's when I added time-band categories: quick (15-20 min), standard (30-40 min), and weekend-only. The AI now knows which nights are slammed. It doesn't just check "can this be cooked in under an hour?" It checks "can this be cooked by a tired parent with one kid asking about homework?"

Context matters. The model doesn't know your context unless you encode it.

Mini Checklist: Building AI for Real-Life Problems

[ ] Identified a recurring, boring problem that burns mental energy weekly
[ ] Defined hard constraints (foods, time, costs) as explicit rules, not suggestions
[ ] Built feedback loops: favorites, "never again," and explicit overrides
[ ] Handled the unsexy parts: ingredient normalization, unit conversion, list deduplication
[ ] Tested with actual users in actual tired-evening conditions
[ ] Measured success by time saved and real adoption, not "demo looks cool"
[ ] AI stays in the "suggest" lane; humans make final calls
[ ] System designed to disappear—goal is "we don't think about it," not "we engage with it"