AI-Powered School Lunch Tracker integrated into Home Assistant

The Problem

My two kids go to different schools. Every month, both schools send a PDF with the lunch menu for the entire month. Simple enough — except I kept running into the same annoying problem: I'd plan dinner, start cooking, and then realise one of the kids had already eaten that exact thing for lunch. Omelette again. Fish again.

The solution should have been easy: just check the menus before cooking. But in practice, pulling up two different PDFs, finding the right day in each, and cross-referencing them every evening was just enough friction that I kept forgetting. I needed something that would surface the right information at a glance, without any effort on my part.

That's when I started thinking: could I build a small AI web app to fix this? And could I surface the result directly on my Home Assistant dashboard, so it's just there every evening when I'm thinking about dinner?

The Idea

The concept was straightforward:

• Upload both PDF menus once a month
• Use Claude's API to extract the meal data from each PDF
• Display a combined calendar showing both kids' lunches side by side
• Highlight days where both kids had the same main ingredient — dinner warnings
• Push today's meals to Home Assistant so they show up on the dashboard automatically

I built the whole thing in collaboration with Claude — using it both as the AI backbone of the app (parsing the PDFs) and as my brainstorming and coding partner throughout the project.

The Web App

The app is a single HTML file — no framework, no build step, no server. It lives in Home Assistant's static file folder and runs entirely in the browser. You open it, upload the two PDFs, and hit a button.

PDF Parsing with Claude

The trickiest part of the project was reliably extracting meal data from the school PDFs. Each school has a different format, and the layouts are inconsistent — tables with day numbers at the bottom of cells, bilingual content in Spanish and English, special event days with no standard format.

The final approach uses Claude's vision capabilities to read each PDF directly. The prompt instructs Claude to:

• Read the table cell by cell
• Extract only the Spanish meal names
• Verify that no entries fall on weekends — a reliable signal that the day numbering has shifted by one

The Calendar View

The result is a month-view calendar with colour-coded chips for each kid per day. Days where both kids had a similar main ingredient are highlighted in amber. There's also a list view for scanning linearly, and inline editing for correcting any days Claude got wrong. Everything is saved to localStorage so the calendar loads instantly on subsequent visits without re-processing. The two kids' calendars are independent, allowing for the odd days when the youngest goes to day care but his older brother is home.

Settings Panel

All configuration — Anthropic API key, kid names, PDF upload, Home Assistant URL and token, and the optional free-text field for extra instructions — lives behind a collapsible settings panel. The default view is just the calendar. The app feels like a dashboard first and a tool second.

The Home Assistant Integration

Once the calendar is processed, hitting "Send to Home Assistant" pushes the full month's data to HA. From there, everything is automatic.

How the Data Flows

The web app encodes the JSON data as base64 and calls a shell_command defined in HA's configuration.yaml. That command runs a one-line Python snippet that decodes the data and writes it to a file called lunch_data.json in HA's www folder.

Four command_line sensors — lunch_kid1_today, lunch_kid2_today, lunch_kid1_tomorrow, and lunch_kid2_tomorrow — read from that file daily, look up the relevant date, and expose the meal as a standard HA sensor state. These are plain sensors that any card, automation, or notification can read from.

command_line:
  - sensor:
      name: lunch_kid1_today
      scan_interval: 86400
      command: "python3 -c \"import json,datetime; data=json.load(open('/config/www/lunch_data.json')); today=datetime.date.today().isoformat(); meals=[m['meal'] for m in data['kid1']['meals'] if m['date']==today]; print(meals[0] if meals else 'Sin menú hoy')\""

The Midnight Automation

A simple HA automation runs at 00:01 every night and calls homeassistant.update_entity on all four sensors. This ensures they always reflect the correct day — today's lunch flips to the new day at midnight, not at some arbitrary time based on when HA last restarted.

alias: Actualizar menú escolar
trigger:
  - platform: time
    at: "00:01:00"
action:
  - action: homeassistant.update_entity
    data:
      entity_id:
        - sensor.lunch_kid1_today
        - sensor.lunch_kid2_today
        - sensor.lunch_kid1_tomorrow
        - sensor.lunch_kid2_tomorrow

The Dashboard Cards

Two Lovelace cards sit on the main dashboard — one for today, one for tomorrow. Each uses vertical-stack-in-card with mushroom-template-card components styled to look like a single unified card. The header shows the date in Spanish, with the cutlery icon turning grey on non-school days. Each kid gets their own row with name and meal.

Tapping either card opens the full web app, so manual edits and reprocessing are always one tap away.

What I Learned

• PDF parsing is harder than it looks. The most persistent bug was Claude associating meal content with the wrong day number, because the number appears at the bottom of each cell rather than the top. The fix was a weekend verification step — if a meal lands on a Saturday or Sunday, the whole sequence has shifted and needs correcting.
• Prompt engineering is iterative. Each version of the extraction prompt was a direct response to a specific failure mode. The final prompt is more specific than I would have written upfront, but it was shaped by real failures.
• Home Assistant's configuration.yaml still matters. Some things — shell commands, command_line sensors — can't be done through the UI and require editing config files directly. That's not a complaint, just something to be aware of when planning an integration.
• The 255-character limit on input_text helpers is a real constraint. We went through a few approaches before landing on the shell_command + file approach, which is cleaner anyway.
• Vibe coding works best as a collaboration. The best results came from describing the problem clearly, reviewing what was built, and being specific about what was wrong. Claude Code was useful for architectural decisions, while Claude.ai was better for the iterative UI and prompt work.

The Result

Once a month, I spend about two minutes uploading PDFs and hitting a button. From that point, the dashboard shows today's and tomorrow's lunches for both kids automatically, every day, on every device in the house.

The dinner warnings have already saved me from serving pasta twice in one day more than once. Small problem, satisfying solution.

The next step is probably a dinner planner that actively suggests what to cook based on what to avoid — but that's a project for another month.

Tech Stack