Next to-dos

Prototype the fridge-photo → ingredient list step (make-or-break)
Decide: detect quantities or just presence? (lean presence for v1)
Write the "minimize extra ingredients" recipe prompt and test it
Dietary constraints — up front or post-filter?
Test whether a whole-fridge photo detects reliably vs shelf-at-a-time

Recent activity

Storage re-keyed — 01KZ02HV4SN8YQHD1ZWQ0TVRNB · 1 week ago
Shared project — Editor access granted to Codex Mobile Audit for ongoing mobile compati... · 1 week ago
Created project · 1 month ago
To-do added — Prototype the fridge-photo → ingredient list step (make-or-break) · 1 month ago
To-do added — Decide: detect quantities or just presence? (lean presence for v1) · 1 month ago
To-do added — Write the "minimize extra ingredients" recipe prompt and test it · 1 month ago
To-do added — Dietary constraints — up front or post-filter? · 1 month ago
To-do added — Test whether a whole-fridge photo detects reliably vs shelf-at-a-time · 1 month ago

Design doc

Pantry Chef — design doc

What it is: Point your phone at an open fridge (or pantry shelf), and get 3 dinner ideas you can make right now with what you have — ranked by how few extra ingredients they need.

The problem it solves

"What's for dinner?" with a full fridge and zero ideas. Recipe apps assume you'll shop first. This starts from what you already own and minimizes the trip to the store.

How it would work

Photo of the fridge → a vision model lists detected ingredients (editable — it will miss the ketchup behind the milk).
You confirm/correct the list.
An LLM proposes 3 recipes, each tagged with "you have everything" or "missing: X, Y".
Tap a recipe → full steps + a tiny shopping list for the gaps.

Approach

Vision: a hosted vision model for ingredient detection; let the user edit (don't trust the model).
Recipes: LLM with a "use what's listed, minimize additions, be realistic about a weeknight" prompt.
No accounts for v1 — it's a single-screen utility.

Open questions

Detect quantities, or just presence? (Presence is probably enough for v1.)
Handle dietary constraints up front, or as a filter after?
Is a fridge photo too cluttered for reliable detection — do we need a "shelf at a time" mode?

Status

Pure idea so far. The fridge-photo demo is the make-or-break: if detection is junk, the whole thing is.