SHELF / SYSTEM II
A vault written by agents, for agents.
The thesis
I don't read my notes. I talk to an agent that reads and writes them for me. Every formatting decision in the vault optimizes for agent retrieval, never for a human reading linearly — that single inversion reshapes everything downstream.
The rules with names
CAPTURE-FIRST
Capturing costs seconds and never asks where something should be filed. Routing is the nightly routine's job, not mine. The vault that made me decide-at-capture-time is the vault that died.
HUMAN VIEW IS A VERB
Briefings and reports are generated on demand and thrown away. A stored summary is a copy, and copies rot.
LIVE STATE LIVES IN ONE PLACE
Each front of life has exactly one status file: where it stopped, next step, recent decisions, blockers. Lesson paid for in duplicated-state cleanup.
It runs itself at night
An always-on Claude Code instance lives on a small server: a systemd unit keeps a respawn loop alive, remote control lets me capture from my phone, and a nightly cron drains the inbox — routing captures into the wiki, linting the graph, logging what it did, committing. I wake up to a consolidated vault and an audit trail.
The work twin
Work gets its own vault, PARA-structured, driven by a three-skill cycle: a morning radar that collects everything in parallel and diffs it against yesterday's snapshot; a context assembler that builds briefings for a unit of work — or hands them off as notes any agent can consume, with an expiry date renewed by touch; and a journal that routes what happened and tags evidence of impact for 1:1s and review cycles. Its governing principle: the agent is a guide, not an executor — it says what I can do next, it never offers to do it for me.
Evolution, honestly
Vault one: a flat folder of notes. Abandoned. Vault two: learned from that — structure, but state still got duplicated, and duplicated state rots. Vault three is the current one, LLM-first from day zero. The system also improves itself: friction gets logged silently, and periodically a skill reads its own feedback and proposes edits to its own spec — a human reviews and commits.