Design·2026·Self-initiated
Chef4Me: Telegram Kitchen Assistant

01 · The problem
Meal planning runs on two frictions most people never solve: knowing what's actually in the kitchen before it goes off, and knowing what to cook with it. Ingredients expire unused, and "what should I make" becomes a decision made from scratch every single day.
02 · The approach
Same starting point as Biscuit AI: meet people where they already are. Telegram is the interface, not a separate app to install. Ingredients can be added one at a time through a guided flow, or in bulk as a freeform list the AI parses into names, quantities, and categories on its own.
- Google Gemini turns whatever's on hand into ranked suggestions across 40+ cuisines: match percentage, difficulty, and cook time included, not just a recipe name
- A scheduler checks expiry dates in the background and proactively flags what's about to go bad, before it's wasted
- Cooking history and 1-5 ratings quietly shape future suggestions, so the bot stops repeating cuisines the user rates low
- An optional Notion sync gives anyone who wants a visual dashboard one, without forcing it on everyone else
03 · AI workflow
A smaller build, and a deliberately fast one. The question was how much working product could come out of a weekend when the scaffolding is generated rather than written.
- Claude Code
- Command handlers, the SQLite schema, and the Docker setup. Generated from a written spec, then corrected where it guessed at intent.
- Google Gemini
- The suggestion engine. Prompt work went into holding it to ingredients actually on hand rather than inventing a shopping list.
- RCTF prompting
- Kept the model briefed on role and constraints, which is most of the difference between a usable suggestion and a generic recipe.
Built in a weekend. The constraint that mattered was not speed but scope: generation makes features cheap, which makes them tempting, and most of the work was deciding what not to add.
04 · The outcome
A full command-driven kitchen assistant: inventory, expiry alerts, cuisine- and diet-filtered AI suggestions, detailed recipes, cook history, ratings, and an AI-generated shopping list. Built and deployed the same way as Biscuit AI: solo, end to end, no handoff.
Next project
BrandLeadway Pensure: Brand & Comms →