Side project · 2026 · 4 min read

My AI operating system.

Obsidian is the memory, Hermes is the hands, and Claude Code builds the tools. I designed it like a product with one user: me.

Role
Solo: design, setup, writing
Timeline
Ongoing, 2026
Surfaces
iPhone, Mac, Telegram
Tools
Obsidian, Hermes Agent, Claude Code, Shortcuts
Status
In daily use

The problem

My reminders went off at 9 am, on the train. By the time I reached the place they were about, I had forgotten them.

Notes sat in three apps. Money moved through five banks and cards, each with its own alerts. Every AI chat started from zero, so I explained my life again each time.

The system

Each piece does one job. The phone notices, the Mac decides, and Telegram is the only place I look.

The daily flow runs left to right. The vault and Claude Code sit behind the Mac.

Memory

Obsidian vault

Notes, plans and a money ledger. Hermes reads it before it answers.

iPhone

Shortcuts

Alarm off, joining office Wi-Fi, arriving home, paying with Apple Pay.

iCloud Drive

A small text file

Each trigger writes one line: what happened and when.

Mac

Scripts and Hermes

A watcher picks up the file. Cron jobs run the rest on a schedule.

Telegram

One chat

Every message lands here. I reply in the same place.

Workshop

Claude Code and skills

Writes and fixes the scripts, and does the design and build work.

The memory

Every note links to others. Hermes follows those links, so its answers draw on my own notes.

Daily notes, plans, money and research live in one Obsidian vault. Before Hermes replies, it reads the notes around the question.

~390
notes in one vault
1,200+
links between them
20
folders, one memory

A day in Telegram

Five triggers on a normal day. The grey labels show them: an alarm, a Wi-Fi network, a payment, the clock, a location.

The parcel reminder waits until I walk into the lobby, the one place I can act on it. A reminder at 9 am would only get swiped away.

Messages and amounts are mock examples.

Design principles

One inbox

Everything arrives in one Telegram chat. I don’t open five apps to find out what happened.

Quiet by default

Silent unless it matters. A price watch that finds nothing sends nothing.

Cheap by default

Plain scripts first, with no AI where a rule will do. A small model for routine sorting, the big model only for judgement.

Human in the loop

It never sends, pays or deletes without asking me. Passwords stay in a vault it can’t read.

Context over prompts

The vault is the memory. Answers start from my own notes, so I stop repeating myself.

What I designed

My part

  • Which moments deserve a message, and which stay silent
  • The message format: symbol, amount, label, one line
  • The triggers, tied to where I can act
  • What needs my yes before it happens
  • How the vault is filed, so it can be read

The tools’ part

  • Hermes runs the jobs and talks to me in Telegram
  • Claude Code writes, tests and fixes the scripts
  • Shortcuts and iCloud carry signals from phone to Mac
  • Models read, sort and summarise

What didn’t work

The wake brief arrived late.

iCloud can take minutes to sync, and the Mac was reading its own older copy of the wake file. Now it waits for the phone’s copy to land before sending.

One stray space broke a trigger.

A Shortcut saved its file as “place.txt .txt”, so the watcher never saw it. Now I check the exact file the Mac receives, not the Shortcut’s settings.

Marketplaces block bots.

Price watches hit bot walls on the big shopping sites. I chose not to work around them. Hermes watches what it’s allowed to, and I check the rest myself.

What changed

Reminders reach me where I can act on them, and I ignore far fewer.

Spending is logged when I pay, so the monthly check-in starts from one ledger. New AI chats start from my notes instead of a blank page. And when I want a new job, I describe it to Claude Code and review what it builds.

What’s next

A one-minute screen recording of a day in Telegram, with mock data. Then a short guide, so other designers can copy the parts they want. And moving more routine jobs from models to plain scripts.

Like my projects? Email me jordychew93@gmail.com or LinkedIn↗