agentic-os

This skill teaches how to treat Claude Code as a persistent runtime or operating system rather than a one-off chat session. It solves the problem of context and automation that disappear when a session ends by codifying an architecture where a kernel config routes tasks to specialist agents, memory lives in files, and automation runs on a schedule. State is stored in a JSON and markdown filesystem rather than an external database, making the whole system inspectable and editable without debugging code.

The architecture has four layers, each a directory in the project root: the kernel (CLAUDE.md) which holds identity, routing rules, and an agent registry; agents/ containing specialist markdown prompts with scoped memory and tools; .claude/commands/ holding user-facing slash commands like /daily-sync, /outreach, and /research; scripts/ for daemon tasks; and data/ for append-only logs, project state, and decision records. The kernel acts as an orchestrator that parses request intent, matches keywords to the agent registry, loads the matching agent file, hands off execution, and synthesizes results. Slash commands are auto-discovered markdown files, and multi-agent tasks can run sequentially or in parallel via Claude Code's background task capability or shell scripts. Memory follows conventions like daily logs, ADR-format decisions, and an inbox for triage, plus an auto-reflection pattern that appends session retrospectives to create a feedback loop.

This skill is aimed at users building multi-agent workflows, a personal or agentic OS, or long-running projects where context must survive across sessions. A key stated design principle is that the kernel should be small and declarative, with routing logic in plain markdown tables rather than code. A notable limitation and requirement: scheduled automation should use external cron such as a macOS LaunchAgent, because Claude Code's built-in cron dies when the session ends.