The Claude PM Operating System

The problem

Everything else I built with AI runs on top of one system: how Claude Code should be configured to work as a long-term collaborator across many sessions, not a fresh assistant every time I open a terminal. That meant designing a memory architecture (not just "keep a big note file"), a way for confirmed patterns to get promoted into defaults without endless re-litigation, and, because I was increasingly trusting this system with real credentials and real customer data, the operational hardening to actually justify that trust.

What I built

• A tiered context-loading system: a small always-loaded core plus load-on-demand domain files, so the assistant starts every session with the right frame without paying the cost of loading everything, every time.
• A typed memory architecture: distinct categories for facts about how I like to work, confirmed project state, and pointers to external systems, instead of one undifferentiated running log that degrades into noise.
• A hypothesis-to-rule promotion pipeline: an assumption has to be independently confirmed multiple times before it becomes a default, and a confirmed rule is demoted the moment new evidence contradicts it.
• Daily automated backup with a two-pass secrets gate: one scan before anything is committed, a second scan of any not-yet-pushed history before it is pushed.

The audit

Ran a full self-audit of the system itself: five review passes producing 25 ranked recommendations, grouped into 10 work packages. Executed 8 of them to date, including a security lockdown, a full purge of sensitive data from history, backup and disaster-recovery hardening, cleanup of dead automation, retirement of an unused local infrastructure layer, and a restructure of working-file conventions.

Stack

Claude Code (CLAUDE.md hierarchy, memory files, hooks), launchd scheduled automation, git and GitHub, Python audit scripts.