a working setup, written down
An always-on Mac runs Claude agents that read my calendar and to-dos, ingest whatever I capture by voice, and text me on my phone — and it plans the day the way I would, because my own frameworks are baked into how it thinks.
The honest frame: strip this to what's load-bearing and it's only a handful of pieces. You can get most of the way with the first two layers and skip the rest until you want them. Start with the spine.
Not "here's a tool" — it's the architecture, and the few highest-leverage pieces that make everything else work. Four layers, each doing exactly one job.
Most "productivity systems" are scripts you run when you remember. The difference here is that there's a box that's awake and reachable 24/7, with a brain wired into the real services I already live in — calendar, to-dos, voice capture, reading — and a layer of rules that make it plan like me without my restating anything. Then it reaches me where I actually am: my phone.
Each layer does one job. Layers 1–2 get you most of the way. The magic — the "how I do it" — lives in layer 3.
A Mac that never sleeps and stays reachable around the clock. I run a Mac mini, but any capable Mac works — an M-series laptop is plenty. This is the actual differentiator. Everything else is "scripts you run when you remember" until there's a box that's awake whether you're at the desk or not.
Keeping a laptop genuinely always-on takes a little trickery — keep it on power, turn on "wake for network access," and keep the lid open or on an external display (closing the lid forces sleep). None of it's hard; defaults just won't hold.
Tools that let the agent actually do things, not just chat: calendar, a to-do system, a voice-memo → transcription pipeline for capture, reading. These hook in as MCP servers and small CLIs, so Claude can read and write real data.
The part that matters: calendar + one to-do system + voice capture. That trio alone gets you a morning brief grounded in real data instead of vibes.
Two parts: always-on rules and invokable skills. The rules are the secret sauce — they make it plan like me without my ever restating anything:
The skills are the workflows. The big one is a connected review loop — morning prep → evening review → weekly → monthly — where the day's commitments flow down and get scored back up. Claude runs the loop; it doesn't just answer prompts.
State is a daily note — one markdown file per day that the whole thing reads and writes. That's the rolling memory; nothing fancier is needed to remember across days.
Delivery is really two jobs: a session runner that keeps a Claude process alive and lets things send it prompts, and a message bridge that carries those messages to my phone. I use a tmux-based session manager and a small Telegram bridge — but both are swappable. The roles matter; the products don't. A 15-minute "heartbeat" fires reviews when the day looks started, not on a rigid clock.
A fresh agent spins up each morning with full access to the calendar, tasks, and voice memos. Here's a day, top to bottom.
A fresh agent session and a new chat thread come up automatically. Resetting daily keeps context from rotting — every day starts clean but with yesterday's note as memory.
It reads my real calendar and to-dos, estimates the day in Pomodoros against my capacity, flags over-commitment, and points at the highest-leverage move first. The pre-9:30 creative block is treated as already spoken for — it never pads it with admin.
Whatever I talk into my phone gets pulled in on a schedule, transcribed on-device with Whisper, classified, titled, and dropped into the day's note — no manual typing. Drafts and meeting notes flow in the same way.
When I start or finish a focus timer, my phone pings a tiny webhook on the Mac, which writes it straight into the day's Pomodoro log and texts back a ✅. Effort gets tracked without me thinking about it.
It processes the day's captures, scores the day against the week's commitments, and writes a scorecard plus a reflection scaffold — then proposes (never auto-applies) which unfinished tasks roll to tomorrow. It also seeds tomorrow's "first move" with the highest-leverage creative action.
The weekly review sets the next week's commitments (which then flow down into every morning prep), and the monthly aggregates four weeks of commitment-vs-reality into a higher-altitude view.
The thing that makes it feel alive: I just reply from my phone, and a Claude session with full calendar + task + vault context answers. Aubrey called it right — "a little personal assistant that gives you a human message about where it's all at."
Not the tool list — the design choices that make it feel less like automation and more like a collaborator who already knows how you work.
The rules encode my frameworks once, so every plan it writes already sizes in Pomodoros and filters by leverage. I never restate "this is $10 busywork."
Commitments flow down from the weekly review, get scored each evening, and reconcile back up. It's a closed feedback system, not four disconnected prompts.
A soft "heartbeat" triggers reviews based on what the day's note actually contains — not a rigid clock. Idempotent by content, never double-fires.
Rolling tasks to tomorrow, reflection prompts — it surfaces and suggests, then waits. I stay the editor; it stays the assistant.
Talk into my phone, and it's transcribed on-device, classified, and filed by evening. The daily note becomes a near-effortless record of the day.
Session runner, message bridge, state surface — every piece is a role you can swap a product into. Nothing's locked to a specific app.
Everything past these is real but additive. This is the spine.
The pieces split cleanly into two kinds. Worth knowing which is which before you start.
The rules (estimation unit, leverage filter, creative-block protection, week-start) and the skills (the morning → evening → weekly → monthly loop, the to-do skill). Frameworks, not plumbing. They drop into anyone's setup mostly unchanged — the part I can literally hand you as a folder.
The always-on machine, the connections to your calendar / to-dos / capture (your accounts, your auth), and your one delivery channel. Nobody can hand you these — but it's a short setup, not a project.
So the handoff is: copy the text core, follow a brief guide to stand up the machine + one channel, and let it quiz you to fill in your frameworks (your estimation unit, your leverage filter, your protected ritual) and your tool targets. You get to a working spine fast, then add the additive layers when you want them.