Skip to content
All work

Product build · 2026

Food Trip

A weekly meal planner that takes the deciding out of dinner: a week built from your filters, one grocery list sorted by aisle, a cook mode for the phone on the counter, and four ChatGPT features that only ever work from Food Trip’s own recipe library.

Role
Design, front-end and back-end
Stack
Next.js 15React 19TypeScriptTailwind CSS 4SupabaseOpenAI
Live
food-trip-ai.vercel.app

The repository is private — happy to walk through the code on request.

The Food Trip home page

01What the build involved

01

A week built from filters, not from scratch

A plan is seven days of breakfast, lunch and dinner, but you never fill it slot by slot. Household size, the diets that apply, a cook-time ceiling and the ingredients you refuse to eat are stored once on the preferences page, and the planner assembles a week from the library using only meals that survive those filters. Autofill, per-slot reassignment, clearing a slot and copying last week sit on top of that as small tools rather than as the main event. Leftovers are a first-class answer: marking a meal as leftovers reuses the previous night instead of sending you shopping again.

02

One list, merged and sorted by aisle

The grocery page is derived, never typed. It walks the whole week, merges the same ingredient across recipes, scales every quantity to the household size and subtracts what the pantry page says is already in the cupboard. Rows are grouped in aisle order — produce, dairy, dry goods — because a list you shop in the order of the shop is a list you finish, and ticks persist so a mid-trip refresh does not lose your place. Metric and imperial are converted at render time from one stored quantity, so switching units never edits the plan.

03

Four AI features, each with a fence around it

ChatGPT does four jobs behind four routes: build a week from the filters, propose three alternatives for one slot, write a recipe from what is in the pantry, and turn a recipe URL into a library entry. Each prompt carries part of Food Trip’s own catalogue, so the model is choosing from real recipes with real ingredient lists instead of inventing dishes that would not survive the grocery page. The guardrails matter more than the prompts: per-account daily call and token limits, a per-feature quota, a twelve-hour response cache, a fallback model when the first is busy, and honest errors — 503 when the key is missing, 429 when a limit is hit. Without a key every non-AI feature still works.

04

Cook mode is a screen, not a printout

Cook mode shows one step at a time with the quantities scaled to the table, reads timers out of the step text, keeps the screen awake while you work, lets you tick ingredients off as they go in and moves with the keyboard. A service worker caches the pages you have opened plus a dedicated offline page, and the manifest makes the whole thing installable, which is the difference between a recipe site and something you actually use with wet hands. The recipe page itself is the unit everything else is built from — ingredients, numbered method and an estimated nutrition panel per serving.

05

Nothing leaves the browser until you sign in

Without an account the plan, the pantry and the templates live in localStorage and are never sent anywhere. Signing in is a Supabase magic link, and plans then sync to your own row under row-level security, which is what the paid tier actually buys. A shared week is encoded into the URL, so a link can be read by anyone without an account. The Sunday email is a cron route guarded by a bearer secret and sent through Resend, and Stripe checkout with a webhook is wired but optional: if billing is not configured the upgrade button says so plainly instead of failing.

02What it left behind

Outcome

Live at food-trip-ai.vercel.app, with the whole non-AI half usable without an account or a key: preferences, the planner, the derived grocery list and cook mode. The AI routes are quota-limited per account rather than open-ended, so the demo can stay up without an unpredictable bill.

Another pass

I would build the recipe record and the grocery derivation before the planner screens. Every AI feature leans on ingredient data being clean, and I ended up reshaping the schema after the scale-by-household maths landed. I would also show the pantry subtraction per ingredient line — “why is this not on my list” was the question I kept asking while testing.

03Inside the product

The home page states the one promise the app keeps — a week of dinners planned from your filters — and describes the four AI features next to the plan they feed.
Every recipe is a real record: ingredients that scale by serving, a numbered method, aisle data and estimated nutrition. Adding it to a week puts each ingredient on the list automatically.

Click any screenshot to view it full size

Next case study

Everly Wedding Planner

A wedding planner that keeps the whole wedding in one place — budget, guest replies, seating, suppliers, checklist and the wedding-day run sheet — with an assistant that reads the plan before it answers.