# octopus-skill 🐙

**One brain, many arms.** A curated library of battle-tested prompts for long-horizon agent work, compiled to whatever host you run — Claude Code, grok, Cursor, Codex. The methodology is shared; each *arm* adapts it to a host's native shape (a loop, or a goal). Like an octopus, one nervous system reaching into different environments and changing color to fit each one.

[](LICENSE) [](CONTRIBUTING.md) 

English · [简体中文](README.zh-CN.md)

<img alt="the graph: an executor node and a clean-context supervisor node communicating only through durable files" src="assets/graph.png" width="100%" />

## Why this exists

A strong model authoring a tuned, opinionated prompt beats hand-driving a host ad hoc — for repeatable, long-horizon, high-stakes work. (For a quick one-off, just type the task; the library's ROI is in reuse.) The durable value isn't any one mechanism — it's the **discipline**: "done" means verified not written, no test theater, no speculative building, forced convergence against growth, and hard owner red lines. That discipline is host-agnostic. octopus is where it lives, once, and gets compiled down to each host.

## The arms

| Arm | Use when | Ships | |-----|----------|-------| | **[`loop-graph`](skills/loop-graph)** | you'll drive it with `/loop` (Claude Code, grok, Cursor, shell); multi-round work that scope-creeps or fakes "done"; multi-milestone phases; executor and supervisor split across hosts/models; owner-gated | an **executor node** (works a single-source-of-truth ledger) + a clean-context **supervisor node** that re-verifies from outside and corrects via a one-way directives file — two loops | | **[`quest`](skills/quest)** | you'll hand it to a goal comman