| name | agentic-os-builder |
| description | Agentic OS — three-layer architecture for turning any project into a structured, AI-navigable operating system. Use when scaffolding project domains, memory layers, or observability dashboards. Triggered by /cks:agentic-os, or when the user wants to organize work into domains + tasks, set up a Raw/Wiki/Output memory structure, or create a dashboard for non-technical stakeholders.
|
| allowed-tools | Read, Write, Edit, Bash, AskUserQuestion, Glob, Grep |
Agentic OS — Domain Knowledge
This skill provides domain expertise for the agentic-os-builder agent. Read templates/ files
when generating scaffolded output.
The Three Layers
Layer 1 — Architecture (The Backbone)
Codifies the project's recurring work into navigable, repeatable units.
- Domains: Top-level areas of work (e.g. Research, Proposals, Client Reports, Sales)
- Tasks: Discrete repeatable activities within a domain (e.g. "Write competitive analysis")
- Skills: Standardized instructions for Claude to execute a task consistently every time
Good domains are things the user does repeatedly — not one-time project phases. If a domain only
has one task, it's probably not a real domain.
Layer 2 — Memory (The Knowledge Store)
Gives Claude a navigable, token-efficient knowledge base across sessions.
memory/
raw/ Staging: meeting notes, research dumps, raw context — process and move to wiki/
wiki/ Codified: crystallized reports, decisions, reference docs — stable, referenceable
output/ Deliverables: exports, proposals, client-ready docs — final products
index.md Navigation map — Claude reads this first to find the right file without scanning all of memory/
The index.md is critical. Without it, Claude spends tokens listing and reading memory files to
find context. With it, one read gives a map of what exists and where.
Layer 3 — Observability (The Dashboard)
Makes the OS visible and accessible to non-technical users.
- Terminal dashboard (
/cks:agentic-os status): Real-time view of domains, memory state, skill shortcuts
- HTML dashboard (
dashboard/index.html): Static file with pre-filled claude CLI buttons for each domain's tasks
The HTML dashboard requires no backend — buttons contain pre-filled terminal commands that
stakeholders copy-paste (or technical users can script to run headlessly).
Domain Naming Guidelines
- Use nouns, not verbs: "Research" not "Researching"
- Match the user's vocabulary — use the words they actually say
- 3–7 domains is the sweet spot — fewer is too coarse, more is hard to navigate
- Each domain should have 3–7 tasks — if it has more, split the domain
Templates
templates/domains.md — .agentic-os/domains.md structuretemplates/domain-skill.md — per-domain skill stubtemplates/memory-index.md — memory/index.md startertemplates/dashboard.html — static HTML dashboardtemplates/claude-md-injection.md — CLAUDE.md blocks to inject
Common Rationalizations
| Rationalization | Reality |
|---|
| "I'll fill in the skill stubs later" | Stubs are useless unfilled — ask for at least 2 tasks per domain during init |
| "memory/index.md isn't important" | Without it Claude scans all files wasting tokens — it's the most important file |
| "The dashboard is overkill" | It's a static HTML file; the user explicitly asked for it |
| "One domain covers everything" | Monolithic domains defeat the purpose — push back and split |
