Agent OS

A personal operating system built on Claude Code. One chief-of-staff session that holds context across your whole operation, a roster of specialist sub-agents it dispatches to, a persistent file-based memory layer, and hooks that inject the right context at the right time.

This is a working system extracted from daily use, published as a fill-in-the-blanks template. The machinery is real; the person has been replaced with placeholders.

What's inside

• CLAUDE.md — the chief-of-staff contract: identity, bright lines, two-gate approval model, dispatch rules, memory architecture. The front door.
• agents/ — five built sub-agents: memory-agent (memory synthesis), defrag-agent (architecture audit, read-only by design), arch-implementer (applies defrag findings), agent-skill-creator (builds new skills/agents), coach-agent (weekly data-anchored coaching).
• skills/ — the slash commands the system composes: /morning (daily brief), /jot (quick capture into memory), /route-memory, /dedupe-memory, /coach, /dispatch-protocol, /build-pipeline, /build-interview.
• hooks/ — SessionStart hooks that auto-inject your identity and work-style files into every session, a keyword-driven context router (context-router.py + map), a single-page-rule enforcer for memory files, and an example bedtime lockout.
• architecture/ — the system's self-documentation: agent registry, permission tiers, hook registry, memory schema, change protocol. architecture/index.md is the map.
• personal-context/ — the memory layer scaffold. Empty by design; the system is only as good as what you put here.
• commands/brief.md — /brief, a git status dashboard across one repo or a whole directory of repos.
• settings.json — permission allowlist, hook wiring, and plugin config to adapt.

Core ideas

• Two-gate model. The agent decides nothing that touches the world (Gate 1), and even after a "go," the actual push/send/spend needs approval again (Gate 2). Between the gates sits a pre-approved autonomy zone.
• Memory vs context. Memory is what's true (persistent files in ~/personal-context/). Context is what's active (auto-injected or keyword-routed per session). /jot is the bridge.
• Single-page rule. Every memory file is ≤100 lines, one topic. Enforced by a hook, not by discipline.
• Audit-fix loop. defrag-agent (read-only) proposes findings; arch-implementer applies approved ones. The auditor can't edit; the editor can't decide.
• Docs are the system's memory. Architectural changes update the canonical docs in the same turn, following the change protocol in architecture/index.md.

Install

1. Back up your existing ~/.claude configuration.
2. Copy CLAUDE.md, agents/, skills/, hooks/, commands/, and architecture/ into ~/.claude/.
3. Copy personal-context/ to ~/personal-context/.
4. Merge what you want from settings.json into ~/.claude/settings.json — at minimum the hooks block if you want the auto-injection. Review the permission allowlist before adopting it; it reflects one operator's risk tolerance.
5. Make the hooks executable: chmod +x ~/.claude/hooks/*.sh.
6. Fill in the blanks (below), then start a Claude Code session anywhere and say "morning".

Requirements: Claude Code, jq (used by the SessionStart hooks), Python 3 (context router). Optional: a Supabase project if you want the Life OS integration the morning/coach skills query — replace YOUR_SUPABASE_PROJECT_ID where it appears, or strip those sections. The skills reference two optional plugin ecosystems (compound-engineering, superpowers); the system degrades gracefully without them, but some skill compositions will no-op.

Fill in the blanks

Search the repo for FILL IN — every site is marked. The load-bearing ones:

1. personal-context/identity.md — who you are. Injected into every session.
2. personal-context/work-style/hard-rules.md — your non-negotiables. Agents enforce what's written, not what's implied.
3. personal-context/work-style/voice-canon.md — how agents should sound.
4. CLAUDE.md § "Who I am" — your priority ranking and sacred windows.
5. hooks/context-router-map.json — your projects' keyword → context-file routes.

A note on what this is not

This is one person's system, shaped by their constraints. The agents assume a solo operator with a day job, side projects, and a habit tracker. Fork the shape, not the life.