Horizon — an AI daily operations manager

Most task apps are a place you write things down. Horizon is the layer above that: it takes your goals, your standing constraints ("I work 9–5, gym M/W/F, deep work mornings"), and your calendar, and turns them into a specific plan for today — then delivers that plan into the tools you'll actually look at.

Live: https://horizon-nu-orcin.vercel.app (demo mode — no sign-in needed)

Board → task completed → copilot → goals across horizons. Captured in demo mode, which runs on sample data with built-in copilot answers — the live copilot and the AI planner aren't shown here.

The core loop

Goals + routine/context  →  AI plan  →  structured tasks  →  you execute  →  review

Concretely: you set goals across four horizons (today/week/month/year) and write a standing briefing once. Each morning an AI planner reads that board, checks your calendar, and composes a realistic day — front-loading deep work, generating concrete next actions when you're under-scheduled, honoring your briefing strictly. The plan ships two ways: a Discord message you'll actually read, and structured JSON an iPhone Shortcut turns into real Reminders. In the evening, a second pass reviews what got done. You can also ask the in-app copilot a free-text question about your own board — it goes to Claude with your real data as context. (Its quick-reply buttons are built-in shortcuts, not AI.)

That's the whole product. Everything else exists to support that loop or to demonstrate it working.

Product vs. implementation vs. lesson — worth keeping separate:

• Product: an AI daily operations manager. Goals + constraints + calendar in, an actionable plan out.
• Implementation detail: Claude scheduled tasks, Supabase, Discord/Shortcuts for delivery. This layer is swappable — it's not the promise, it's one way to keep it.
• Lesson: external automation needs monitoring and a verified fallback, not just one that exists on paper. See below.

Why this is interesting as a case study

The planner is a real scheduler, not a template. It reserves protected deep-work blocks around your actual calendar, estimates how much open deep-work time is left in the week, and only generates filler tasks when you're genuinely under-scheduled — derived from your week/month/year goals, in that priority order.

The copilot is scoped to your real data, with no service-role key involved. It's the one function called directly from the browser, so it authenticates the caller's own Supabase session and runs every query through a client scoped to that JWT — Postgres row-level security does the authorization, not application code. Ask it something the data can't answer and it says so instead of guessing.

It has already taught me something by failing. Horizon's morning/evening planning ran as an external scheduled Claude Code task — separate infrastructure from the web app, with a pg_cron fallback in the database specifically built to catch the scheduler going down. Both the primary and the fallback went silent at the same time. As of this writing they're still paused, deliberately — restarting them was cheap, but doing that before deciding whether this layer belongs in the finished product would just be sinking more effort into the exact infrastructure that's in question. Not hidden here — see docs/internal/ALWAYS-ON.md for how the fallback was designed to work and where the monitoring gap actually was.

Architecture

Piece	Tech	Notes
Web app	Single-file static HTML/JS + supabase-js	index.html, deployed on Vercel
Database	Supabase Postgres	goals, profiles, habits, plans — RLS owner-only
Auth	Supabase email + password	single user; public sign-ups off — visitors use demo mode
board fn	Deno edge function	snapshot for planners; rolls day-tasks forward
plan fn	Deno edge function	stores/serves structured daily plan JSON
send-msg fn	Deno edge function	Discord webhook relay, chunked ≤3 messages
copilot fn	Deno edge function	real AI (Claude Haiku), scoped to the caller's own data via RLS
Morning planner	scheduled Claude Code task (7:05am)	reads board + calendar, composes and delivers the day — currently paused, see below
Evening check-in	scheduled Claude Code task (9:07pm)	review + streak nudge — currently paused, see below
iPhone Reminders	Shortcuts automation (7:15am)	pulls plan, creates Reminders

Server-to-server edge functions (board/plan/send-msg) authenticate via an X-Horizon-Key header checked against a secret with a constant-time compare. The browser-facing copilot function instead authenticates the caller's own Supabase session JWT and never touches a service-role key. Calendar is read-only input to planning — the assistant never writes time blocks.

Also in the app (built and working, just not the headline)

Habit tracking, a streak that only advances when every task clears, an insights panel, weekly/monthly planning prompts, and a design pass audited against Emil Kowalski's and Apple's animation/motion standards. Kept out of the main pitch because they're elaborations on the loop above, not the loop itself.

A two-way Discord bot (edge-discord-bot.ts) was also built — /board, /done, /add, /skip — but was never armed (no DISCORD_PUBLIC_KEY set) and has never run. Kept as source, not claimed as a feature.

Repo map

• index.html — the app
• edge-board.ts / edge-plan.ts / edge-send-msg.ts / edge-copilot.ts — deployed function sources
• edge-discord-bot.ts — built, never activated (see above)
• prompt-7am-discord.md / prompt-9pm-checkin.md — the two scheduled-task prompts that run the loop
• prompt-sunday-planning.md / prompt-monthly-planning.md — secondary planning cadences
• SHORTCUT-SETUP.md / DISCORD-BOT-SETUP.md — setup docs
• migrations-*.sql — applied database migrations
• server.js — tiny local dev server (node server.js → :8731)
• docs/internal/ — build history and internal planning notes, kept for reference

Built collaboratively with Claude.