Playbook, planning
Plan a week of meals with real constraints
Generic meal plans die by Wednesday. We feed the model our real constraints, then make it negotiate the grocery bill.
The short version
- Constraints first: budget as a number, real cook nights, diet rules, pantry contents.
- Ask for the plan, an aisle-grouped list, an estimated total, and planned leftovers.
- Treat draft one as an opening bid and negotiate the misses.
- Get a Sunday prep hour so weeknights take 15 minutes.
- Rerun weekly with whatever is left in the fridge.
Meal planning fails for a boring reason: the plan ignores your actual week. Two nights you won't cook. A kid who won't touch fish. A budget that is a number, not a mood. We stopped asking AI for meal ideas and started handing it constraints, and the plans finally survived past Wednesday. Planning with AI is a negotiation, and the constraints are your side of it.
The order matters. Anthropic's prompting docs say to treat the model like a bright new hire: state the task, the rules, and the context explicitly, because it treats anything you leave out as a free variable. Name 90 dollars and it plans to 90 dollars. Say cheap and it guesses what cheap means to you.1
The old-school advice still holds underneath. USDA's Nutrition.gov guide to food shopping and meal planning comes down to three moves: make a list, plan around what you already have, and let sales set the menu. The model just does the assembly in seconds instead of eating your Sunday hour.2
Here's our routine. It plans a trip the same way: swap meals for days and the grocery budget for the trip budget.
- Dump every real constraint first. The weekly budget as a number, the nights you'll actually cook, diet rules, allergies, and what's already in the pantry. Two minutes of typing here saves the whole plan.
- Ask for the plan plus a shopping list grouped by aisle, with an estimated total and leftovers planned on purpose. An ungrouped list gets abandoned in the store.
- Negotiate draft one. Over budget? Demand swaps. Tuesday looks grim? Name it. The second draft is always better, because now the model knows what you actually meant.
- Ask for the Sunday prep hour: what to chop, cook, or marinate ahead so the Wednesday version of you gets a 15-minute dinner instead of a takeout order.
- Save the prompt. Next week, paste whatever is left in the fridge and rerun it. The reruns get better because the constraints stop changing.
Draft one is an opening bid, not a plan.
The negotiation step is where people quit, and it is the step that pays. We used to accept whatever draft one said, cook three of the seven dinners, and quietly order pizza on Thursday. Now we spend ninety extra seconds pushing back, and the plan reflects the family we actually are instead of the family the model imagined. The difference shows up on the receipt.
Two honest limits. The model does not know your store's prices, so treat the estimated total as a sanity check and compare it against one real receipt; feed the actual number back and the next estimate lands closer. And a plan you won't eat is worth nothing, which is why the picky-eater constraint goes in first, not last.
Use it this week
- Run this tonight with your real numbers and see whether the total survives contact with your store.Plan dinners for my week. Constraints: budget $[number] total, I will cook [n] nights, [diet rules or allergies], household of [n] including [picky eater notes]. Pantry already has: [list]. Return: a day-by-day plan, a shopping list grouped by store aisle, an estimated total, and which meals produce leftovers for lunch. If the budget does not fit, show me the two cheapest swaps instead of going over.
- After you shop, paste the receipt total back and make the model recalibrate.My actual receipt total was $[number] against your estimate of $[number]. Adjust your price assumptions, tell me which items drove the gap, and keep next week's plan within 10 percent of budget.