Personal OS

I let Claude maintain my second brain.

Every knowledge system I built before this one died the same way. I captured notes, linked them, reviewed them, and then a busy week hit and the graph rotted. That is the standard failure mode of personal knowledge management: the human is the maintenance worker, so the system is only as good as the human's worst week.

So I inverted the roles. The vault is plain markdown in a git repo, and agents do the maintenance. They capture, link, enrich, and review. I design the work, make the judgment calls, and consume the output. The graph stays dense because keeping it dense is no longer my job.

Personal OS is that system, genericized and open sourced: the schema, the operating manual, the routines, and the guardrails that make handing over the keys safe.

What makes it work

Three design commitments do the heavy lifting. They are what separate this from pointing an LLM at a folder of notes.

• Typed, schema-backed markdown. Every note declares what it is: Person, Organization, Event, Project, Plan. Relationships are typed, directional links in frontmatter, not a generic bag of related items. Agents traverse and update the graph without parsing prose.
• Explicit section ownership. Every file type marks its sections human-only or agent-managed. Your personal notes on a person are sacred. The overview and history are agent-owned. Clear write contracts are what make autonomy safe.
• Git as undo. Every agent action is a commit. A bad run is one revert away. That is what lets you hand over the keys without flinching.

The daily loop

Overnight, a routine preps notes for the day's meetings, enriches whatever changed yesterday, and writes a morning brief. During the day, meetings land as notes linked to the people, organizations, and projects they touch, and those links are the fuel: enrichment works from what actually changed, not from a queue I have to maintain. In the evening, a close-out recaps the day and files loose pages. On Sundays, a weekly review closes the week against quarterly and annual targets.

Four moving parts. I touch none of the plumbing.

What's in the repo

• The operating manual. The model, the structure, and the date discipline every agent session reads before it does anything.
• Templates and saved views. One template per record type, plus Obsidian Bases views over the graph.
• Routines and skills. The overnight run, the Sunday review, the evening close-out, and meeting prep on demand.
• Agent docs and logs. Documentation written for the machine, and behavior tracking in files that outlive apps.
• Ten chapters of documentation, starting with the philosophy and ending with what to change when you make it yours.

Take it

The template is public: github.com/prounds-4/personal-os. Clone it into a private repo, because your vault is your life and the template is the only part that belongs in the open. Fill in the placeholders, open the folder in Obsidian and in Claude Code, and register the routines.

Two guardrails worth stealing even if you take nothing else. The agent never edits its own configuration: if a routine thinks it should change, it says so and you decide. And human-only sections are never touched, no matter what else the agent is doing.

The same idea, pointed at a different problem: Fitness OS. The rest of what I build is at Projects, and the career record is at Work.