Featured projects
All three are public on GitHub — including the specs and the AI prompts that produced them. Around here, the prompt is the source code.
Family Routines App Live app
A daily routine tracker for families: morning and evening check-offs, streaks, and a noon-rollover cycle, with per-person dashboards and an admin mode. The whole app is a single HTML file backed by Firebase — and it was generated, not hand-written.
The interesting part is how it was built: a full specification and a library of prompts live in the same repo — the master design spec, the generator prompt that produces the code, the build prompt, and the phase plan. Change the prompt, regenerate the app.
AIMF — AI Messaging Format Protocol
An open protocol that lets AI assistants exchange structured messages — notes, file drops, requests, status reports — over ordinary email, with sender verification, consent agreements between the humans behind them, and versioned updates that every participant has to acknowledge.
Designed, tested, and run live between two different AI assistants in October 2026: the spec carries test vectors a bot can grade itself against, a trial prompt any assistant can be onboarded with, and a public registry of agreements. The canonical home is the repo — if it isn't on GitHub, it isn't an update.
This website You're looking at it
The showcase itself is one of the builds: a single static HTML page on GitHub Pages — no framework, no build step — where every project and automation carries the spec and the AI prompt behind it, on copy buttons. It also runs GoatCounter analytics on page views and copy clicks, a feedback/contact form that never exposes an email address, and a public comments section (below) powered by GitHub Discussions.
Built the same way as everything else here: spec first, generated, pushed, and verified live. Its copyable prompt is kept in lockstep with the site itself — whenever the site changes, the prompt changes to match, so copying it rebuilds this site, not last month's version.
Automations
The unglamorous systems that do the daily work. Each one was designed around a simple rule: it should run quietly, explain itself when it acts, and never guess when it can ask. Each card's prompt is a complete, self-contained brief — copy it, paste it into an AI assistant that can see your calendar, and it can build the same system for you.
🚗 The self-maintaining calendar
Every morning, a scanner reads the next seven days of the calendar and adds private drive-time blocks before in-person events — mapped drive × 1.2 + a cushion, chained from home or the previous stop, with get-ready blocks, pickup routing for family events, and a plain-English note in each block explaining how the route was deduced. Tight gaps get a warning pin; ambiguous driving duties get batched into one confirmation list instead of a guess.
⏰ Time-to-leave alerts, with a light show
Every 15 minutes, the next trip is checked against live traffic from the current GPS position. When it's time to go: one ping — and if I'm home, all five smart lights run a prism show at full brightness, then restore themselves three minutes later. If I leave and stay away, the house turns its own lights off.
⛽ Gas tracker
I report fill-ups and the gauge needle; the system estimates miles driven from the calendar, projects the day the tank crosses a quarter, and pins a "get gas" stop onto a trip I'm already making — at the best-priced convenient station on that route, never slipping an arrival time to do it.
🎺 Band schedule monitor
A daily diff of a high-school marching band's weekly schedule document, plus coordinator emails and crew texts. Changes surface the same morning; on performance days the brief leads with the call time (not the show time) and the crew sign-up link.
🗂️ PARA + GTD filing system
Every new document lands in a PARA folder structure in Drive; every task lives in exactly one GTD list and carries links back to its project resources, so a task and its paperwork are always one search apart. A weekly review task generates itself every Friday.
📬 Priority comms watch
Every 15 minutes during waking hours, texts, email, and call logs are checked for one priority contact, so a time-sensitive family request never sits unseen. Everything else batches into a single morning list — no drips.
🧹 Inbox hygiene
School digest and notification mail is auto-labeled and swept out of the inbox daily; anything stale past five days trashes itself. The inbox stays for things that need a human.
🌅 Morning day-plan
One consolidated brief to start the day: today's drives and tight spots, tasks due, replies owed, and anything that changed overnight — assembled from the same systems above.
How these get built
Spec first, prompt as source, verify by reading back
- Write the spec — behavior, edge cases, and what "done" means — before any code.
- Keep the prompts in the repo next to the output they generate; a fix starts as a prompt edit.
- Generate, then verify by reading the result back from where it actually lives.
- Track every known issue in the open, in the repo, with a status log anyone can audit.
- Automate the boring verification too — the systems watch themselves and report changes.
Steal this site
This whole page — the inventory, the cards, the visuals, the copy-prompt buttons — was built by Muse (Meta's personal AI agent) from a conversation. Here is that conversation, distilled into one prompt. Copy it, paste it into your own AI assistant, and get your own version.
🎁 The "build my showcase" prompt
It walks your assistant through everything: inventorying your GitHub, featuring the specs and prompts behind each project, writing up your automations without leaking private details, wiring the copy buttons, share links, feedback form, and analytics — then publishing to GitHub Pages and verifying the live page before declaring victory.
Get in touch
Like something? Spot a mistake? Want to talk about building systems like these? All of it lands in the same place.
Share, feedback & contact
Share the site with a friend, send feedback or an idea, or just contact Andrew directly — the form has a "Contact Andrew" option, and you can leave your email if you'd like a reply. Prefer plain email? hello@andrewdhorowitz.com reaches the same inbox.
Community
Built something with one of these prompts? Questions, ideas, corrections — this is the public square. Comments run on GitHub Discussions via Giscus — sign in with GitHub to join in.