01

I treat Claude like an operating system, not a chatbot

Everything is anchored on three principles: a single vault as source of truth, predictable session boot from canon files, and modular composability between voice files, skills, and commands.

Personal stack as of July 2026.

02

The vault — single source of truth, PARA-organised

One iCloud-synced folder, opened in Obsidian. Markdown-only. Every Claude surface (Cowork, Claude Code, Desktop) reads it.

Path: ~/Library/Mobile Documents/com~apple~CloudDocs/AI/Claude 2/.

03

Three surfaces, same brain

All three read the same vault. Boot sequence is identical.

04

Every session boots from the same canon files

Boot sequence applies to all three surfaces. system-overview.md (added 9 May 2026) is the new canonical orientation file for any new Claude agent.

05

Voice files — six profiles by format and language

All six live at 10 System/. Only one is loaded per session — the one matching the task.

06

Skills, plugins, and commands

Skills handle craft. Vault commands orchestrate workflows. All composable. (The scheduled-task layer was decommissioned in 2026 — see below.)

Active inventory at 10 System/Cowork Skills Inventory.md.

07

Every project has the same shape

Post-DB-methodology adoption (May 2026). Every active project under 20 Projects/<Name>/ has four artefacts, each with its own discipline:

Phase 1 + 2a + 2b complete across the implementation queue (Phrygia, Phrygia Tax, SAFT Waiver — Orobit, OroBit, Aristeia Press, Nemo) as of 8 May 2026.

08

ADRs — decisions get frozen

Every substantive forward-binding decision becomes an Architecture Decision Record. Five required sections, immutable once status: accepted:

---
id: ADR-NNNN
title: ...
status: proposed | accepted | superseded
date: YYYY-MM-DD
---

## Context
What forced this decision. What was true at the time.

## Decision
The choice made.

## Alternatives considered
Each alternative with a one-line rejection reason.

## Consequences
Forward-binding implications. What this commits us to.

## References
Wikilinks to LOG entries, other ADRs, source documents.

System-level ADRs live at 10 System/decisions/ (e.g., ADR-0001 adopting system-overview.md). Project-level ADRs at 20 Projects/<Name>/decisions/.

Template: 11 Templates/ADR.md. Backfilled 31 ADRs during the 8 May Phase 2b pass.

09

Pattern-candidates — rules earn their place at N=3

The umbral discipline. New rules don't get added to my-rules.md in advance. They sit in 10 System/pattern-candidates.md until they recur three times in real work.

Stops my-rules.md from accumulating preventive scaffolding I'd interpret too literally. The protocol grows by use, not by design — and several candidates have since reached N=3 and been promoted to canon.

10

cowork-authorities.md — the closed delegation list

A short, written, closed list of operations Cowork can perform without asking. Anything outside the list = ask first. Expanding the list requires a new system ADR.

This is the inverse of "the AI uses judgment." It's "the AI uses the list."

Source: 10 System/cowork-authorities.md. Modelled after DB's [chore-trivial] protocol.

11

The session-close flow — three tracks

Closing a session isn't one button. 10 System/Commands/session-close.md defines three explicit tracks. Each item routes by track.

Memory mirror (step 7) copies Cowork's auto-memory to 10 System/Memory/ — one-way, vault accumulates dormant evidence.

12

Project lifecycle across sessions — sample case

Sample case: Project Compass — Q2 board strategy review for a portfolio company.

Session log filed at 20 Projects/Compass/Session Logs/<YYYY-MM-DD-HHmm>.md.

13

Two daily routines I built — now mothballed

The scheduled call loop is retired; transcript processing still runs on demand. Detto replaced Tome as the meeting-notes source in June 2026.

14

A self-improving research wiki

The Karpathy-method wiki at 60 Resources/. Sources, topics, companies, and people pages with structured frontmatter and back-references.

Each ingest follows a fixed schema: what the source claims, the evidence, the open questions, and where it came from. Schema enforced by 10 System/wiki-rules.md. Opus-only ingest rule — reasoning quality matters more than throughput.

Source-type taxonomy now live: factual research, type: doctrine (opinion / manifesto), how-to / instructional — each with a different ingest schema. The doctrine page schema and the publisher-stub check (wiki ingest step 2c) are both in force. Andreessen's Techno-Optimist Manifesto was the first doctrine-type ingest; a second has since been added.

Library now holds 71 sources read in full, 62 cross-linked company pages, 14 topic pages, and 2 doctrine pages — including the 63-source Burry / Cassandra corpus ingested June 2026. Counter resets on every-10-readings audit.

15

Notes vs working files — two homes, mirrored names

Rule of thumb: if Claude reads it to understand a project, it lives in vault. If Claude produces it, it lives in Claude Assets.

16

MCPs — connectors into the rest of the stack

MCPs are scoped to read or draft. The one connector that can send (Multi-Gmail) is gated behind a recipient allowlist and a push-notification approval — and any actual send still follows the red lines.

17

Hard rules and red lines

Canon: 10 System/my-rules.md. Read every session via boot. Security audit completed 17 May 2026: 11 critical findings closed, ~16 credentials rotated.

18

What I've built with this system

Deployment tracker is the single source of truth for what's live: 30 Areas/Deployments/deployment-tracker.md.

19

Four day-to-day workflows — the actual wiring

The system was built for board memos and investment decks. It now runs four very different domains using the same project shape.

Same project shape — state + journal + decisions + outputs — applied to four very different domains. The shape generalises.

20

The principle — modular, reusable, controlled, emergent

Every choice in the stack passes through these four words: modular, reusable, controlled, emergent.