Five AI agents that handle lessons, logistics, groceries, paperwork, and screen time for your household. Zero code. Build the first one tonight.
A mom of seven went viral for running five AI agents on Mac Minis in her house. A photo of a textbook page becomes tomorrow's full lesson plan. A grocery cart builds itself: $127, untouched by human hands. A custom TV app with no ads and no algorithm, shipped in four days by someone who had never opened a terminal. This playbook is that exact stack, written as steps you can execute today.
Plus the foundation: one always-on machine, one house rules file every agent reads, and safety rails on anything that spends money or touches your kids.
You need one computer that stays on. A Mac Mini is the classic pick: silent, sips power, hides behind a monitor. Any always-on Mac or PC works. Do not buy five machines. Start with one and split later if you want the symmetry.
Install an agent runner, the tool that reads plain English and does the work. Claude Code is the pick:
npm install -g @anthropic-ai/claude-code
Make the home base folder:
mkdir -p ~/family-stack/{inbox,lessons,calendar,paperwork,tv-app}
cd ~/family-stack
Now write STACK.md, the house rules every agent reads before acting. This one file is your whole household philosophy, machine-readable:
# STACK.md: House Rules
Kids: ages 4, 7, 9, 11, 13.
Allergies: none. Dinners: 30 minutes max.
Grocery budget: $150/week. Never check out without my approval.
Approved channels: Mark Rober, SmarterEveryDay, Cosmic Kids
Yoga, Art for Kids Hub, PBS Kids.
Screen time: TV app only, 60 minutes on school nights.
Never message the kids. Kid-facing output comes to me first.
Every agent below inherits this file. Keep it updated and the whole stack behaves.
The workflow: snap a photo of a textbook page, drop it in inbox/ (sync that folder to your phone with iCloud or Dropbox so this takes ten seconds), and let the nightly run do the rest. Save this as a standing instruction:
Every night at 9pm, check inbox/ for new photos.
For each textbook page: identify the subject and chapter,
then write tomorrow's lesson plan to lessons/YYYY-MM-DD.md.
Include a morning block (45 min), an afternoon block (45 min),
three discussion questions, and one hands-on activity using
normal household supplies. Then build a printable worksheet
as a clean HTML file I can print from the browser.
Move finished photos to inbox/done/.
Then tell the agent: "set this up to run every night at 9pm." It wires the schedule itself. That is the zero-code part: you describe the job, it builds the job.
Two standing instructions:
Sundays 4pm: read next week's family calendar, write a
7-day dinner plan that fits the budget in STACK.md, and
save it to calendar/meal-plan.md.
Mondays 9am: turn the meal plan into a grocery list, add
every item to my Instacart cart, and stop at the review
screen. Send me the total. Do not check out.
Honest note: the grocery step needs a connection, either the store's app, an integration, or an agent that can drive a browser. Start with the meal plan and the list. Add the auto-cart when you trust it. The mom from the reel let hers order $127 of groceries solo. You can cap yours at the review screen and still get the hour back.
Every day at 6am, send me one message: today's calendar,
the weather, a link to today's lesson plan, tonight's
dinner, and anything due in the next 72 hours. Plain text,
under 120 words. Deliver it where I actually read: email,
iMessage, Slack, or Telegram.
One message, whole day. This is the agent you will feel first.
When a photo of a school form lands in inbox/paperwork/:
read it, pull out every date and deadline, add them to the
family calendar, write a two-line summary to paperwork/,
and set a reminder 48 hours before anything is due.
Permission slips stop living in backpacks. Field trip money stops being a 7am surprise.
The headline build, and it is one prompt:
Build a kids TV web app in tv-app/. Hard requirements:
only videos from the approved channels in STACK.md, no
recommendations, no autoplay chains, no comments, no
shorts, no sidebar. Big tap targets, works in a TV
browser. Use the YouTube Data API with a channel
whitelist. Run it on a local server and give me the URL.
Then parent-test it and iterate in English:
It needs a free YouTube API key, and the agent will walk you through getting one. No API key? Point it at a folder of downloaded, pre-approved videos instead. Same result, fully offline. Four days from prompt to TV is realistic. The reel was not exaggerating.
family-stack folder created, STACK.md house rules written50+ free playbooks on AI agents, automation, and running a one-person empire. No pitch, no pricing pages. Just the goods.
Get all 50+ free playbooks