A skill that guides you, step by step, as you build your own agentic OS—a file system that understands your goal and works for you. You don’t need to know how to code; the skill handles the technical work, and you make the decisions.
You tell it where you want to go, it asks two or three questions in plain English, then builds the actual files and folders for you, explaining what it did and why. Session after session, it remembers exactly where you left off.
Trains you in Identity, Substrate, Rules, Skills, Tools, and Agents in the order that actually works—never all at once.
Speak like a person. When a technical term is unavoidable, it comes with a short sentence and an everyday analogy. Assume you've never opened a terminal.
If you mark a field as private, it never writes that value to the folder—instead, it stores a reference that doesn’t identify you.
Identity is the foundation, the Substrate (the behind-the-scenes setup) is the largest layer, and Agents sit at the top — only after everything beneath them is solid. The core idea: an agentic OS is mostly plumbing. The AI model is the cheap 20%.
| # | Layer | What it is | Lives in |
|---|---|---|---|
| 1 | Identity | The soul. Who this OS is, who it serves, and what it refuses to do. | CLAUDE.md |
| 2 | Substrate / Context | The inner workings. The distilled knowledge the OS runs on. The largest layer. | substrate/ |
| 3 | Rules & Hooks | The guardrails. Black-and-white restrictions and automatic reflexes. | rules/ |
| 4 | Skills | The verbs you’ve mastered. Repeatable tasks packaged to run the same way every time. | skills/ |
| 5 | Tools / Connections | The connections outward. Read-only, well-scoped connections to real data sources. | tools.md |
| 6 | Agents | Roles with judgment that orchestrate skills. | agents/ |
No dependencies, no API keys, no build step.
The coach runs as a skill inside Claude Code.
# install the skill in any folder cp -r skills/os-agentes ~/.claude/skills/
Any empty (or existing) folder you want to turn into an agentic OS.
It doesn't need to be polished. The skill refines it with you through the first few questions.
Real commands, in the order you would normally use them.
Copy the skill folder to Claude Code's skills directory.
cp -r skills/os-agentes ~/.claude/skills/ # available from any folder
Inside the folder you want to turn into an OS, state your goal in plain language.
/os-agentes start Quero nunca mais perder o prazo de um cliente
With each answer you give, it builds the actual files and explains what it did.
/os-agentes next # move on to the next unfinished layer and train it
When you already know what you want to train first.
/os-agentes layer identity|substrate|rules|skills|tools|agents
The progress map comes straight from memory — nothing gets lost between sessions.
/os-agentes status
Scores all six layers against your goal and gives you the three highest-leverage next steps, written in OS-AUDIT.md.
/os-agentes audit
Explains what the skill does and shows all available commands.
/os-agentes help
/os-agentes start Eu vivo perdendo prazos de cliente e quero nunca mais atrasar uma entrega — the skill asks three short questions (who it’s for, the question they’d like to ask, one always and one never), writes the Identity, and in the next step builds a deadline tracker that responds right away: "Couple B, the full gallery expires in 3 days." One audit identifies your three next priorities—including a security one: write rules/never.md before a public backup exposes client data.
The Agent OS doesn't carry out tasks for you day to day — it structures the foundation (Identity, Substrate, Rules, Skills, Tools, Agents) that a reliable agent can run on. An execution platform such as OpenClaw takes that ready-made foundation and puts it to work: messages, automations, tools, dashboards.
