Dan Bennett
← Back to portfolio

Meal Planning Pipeline

Live

A 3-minute conversation produces a fully-deployed app — shopping lists organized by my grocery store's layout, cook-mode walkthroughs with built-in timers, and everything published to a family docs server I use on my phone every week. Built in under 2 hours because the Skills were already composable.

Claude Code Skills Human-in-the-Loop Python

The Problem

Some weeks I own meals and groceries for my family of four in Haarlem, Netherlands. Meal planning is deceptively heavy cognitive work: figuring out what to cook based on what's worked before, building a shopping list, navigating a Dutch supermarket efficiently, and then actually cooking everything with two small kids in the background.

I also needed a general-purpose way to get AI-generated outputs onto my phone without requiring a terminal — not just meal plans, but daily briefings and task summaries too.

My Role

Solo builder and primary user. Designed the system, built the infrastructure, and use it weekly. Built in 1–2 hours because the meal planning skill and the documentation skill were both already working — they snapped together.

The Approach

The key insight was that meal planning has a natural conversation-to-output structure. I know my family's preferences. The AI knows recipes. The interesting part isn't a fancy UI — it's the pipeline from a short planning conversation to a fully-deployed, mobile-ready app.

The workflow uses two Claude Code Skills in composition:

  1. Meal Planning Skill: I start a Claude session. The AI recommends recipes based on past family successes — what the kids actually ate, what was easy enough for a weeknight, what worked well as leftovers. I approve or reject until we land on 3 meals that will each produce ~8 servings (two dinners from one cook).
  2. Documentation Skill: The AI automatically generates a shopping list organized by how my specific grocery store (Albert Heijn) is laid out, a cook-mode walkthrough with step tracking and timers, and publishes everything to the family docs server.

Rather than building a dedicated meal planning app, I built a lightweight documentation server that could serve any AI-generated HTML — then pointed my meal planning workflow at it. One infrastructure, many content types.

What I Built

  • Human-in-the-loop recipe selection — conversational approve/reject loop until 3 recipes are locked in, informed by family preferences (not too spicy for the kids, good for leftovers, balanced)
  • Store-layout-aware shopping lists — items organized by supermarket section (Groente & Fruit, Vlees & Vis, Diepvries, etc.), ordered by physical store layout so you walk the aisles once, not zigzagging. Dutch product names with English translations.
  • Color-coded recipe tags — each shopping item tagged to its recipe so you know which items go with which meal
  • Cook Mode — step-by-step instructions with countdown timers (Web Audio API), screen wake lock for cooking with dirty hands, per-recipe step tracking via localStorage
  • Single-file HTML apps — each meal plan is self-contained with no dependencies, works offline, dark mode, progress tracking
  • Shared docs server — Python HTTP server serving meal plans, daily briefings, and project docs over Tailscale to any device
  • CLAUDE-INSTRUCTIONS.md — a meta-artifact that teaches future Claude sessions how to produce and publish new meal plans without re-explaining the system

Why This Matters

This project demonstrates Skills composition. The meal planning skill and the documentation skill were both built for other purposes — but because they follow Claude Code's skill interface, they snap together. The 1–2 hour build time was possible because I wasn't building from scratch; I was composing existing capabilities.

The pattern — conversational input, automatic multi-step output, deployed to a real user's phone — is exactly the "AI agent that delivers value to end users" workflow. The ratio of human input to delivered output is extremely favorable: 3 minutes of conversation produces a complete, deployed application.

The CLAUDE-INSTRUCTIONS.md file is worth noting separately. It's a 21-line document that teaches AI how to produce future meal plans and publish them correctly — a micro-example of the "AI building AI tools" pattern.

The Result

Before grocery runs, I open the docs server on my phone. The shopping list is organized by store section, so I walk Albert Heijn once from produce through freezer, checking items off as I go. When I get home, I switch to Cook Mode and follow the steps. The timers ping when something needs attention. The screen stays on while I've got flour on my hands.

The entire process from "what should we eat this week" to "shopping list on my phone" takes about 3 minutes of conversation. The AI handles the rest — recipe details, ingredient quantities, store-section organization, HTML generation, server deployment.

The system has been in weekly use since April 2026. The same docs server also serves daily overnight briefings and project documentation — one infrastructure, many content types.

Tech Stack

  • AI: Claude Code (Opus) with custom Skills
  • Frontend: Vanilla HTML/CSS/JS, single-file apps, localStorage, Web Audio API, Wake Lock API
  • Server: Python HTTP server with API endpoints, Tailscale for remote access
  • Orchestration: MARVIN (Claude Code-based AI chief of staff)
  • Publishing: Auto-discovery index generator, CLAUDE-INSTRUCTIONS.md for reproducible workflows