as-skill: Instructions file for Codex

AGENTS.md

This file is the entry point for any AI agent working with the skills in this repository. Read it in full before your first action.

What this repository is

A multi-domain collection of rules, skills, and agents for Claude Code. Each domain is a self-contained vertical (Obsidian PKB, Git workflow, DevOps/Kubernetes, content, a code dev-loop harness). The layout is uniform; the working contract is shared across domains.

Always-on skill

Always run with the caveman core skill (domains/core/skills/caveman/) active — ultra-compressed output, full technical accuracy kept. Don't wait for the user to ask for brevity; it's the default in this repo.

Architecture

as-skill/
└── domains/
    ├── <domain-name>/
    │   ├── manifest.yaml      ← name, description, requires_env, targets
    │   ├── rules/             ← durable domain rules (optional — not every domain has one)
    │   ├── skills/<name>/     ← atomic operations: SKILL.md + README.md
    │   ├── agents/            ← orchestrators
    │   └── hooks/             ← optional; only `code` has one today
    ├── code/
    ├── content/
    ├── core/
    ├── devops/
    ├── git/
    └── obsidian/

Each domains/<name>/manifest.yaml declares:

• name, description — what the domain is
• requires_env — environment variables the domain won't deploy without (e.g. OBSIDIAN_VAULT for obsidian)
• targets — where rules/skills/agents land after install

Dependencies run top-down: an agent knows about a skill, a skill knows about a rule. Not the reverse — except in code (see below).

Exception — code: the /code-plan, /code-build, /code-review, /code-debug skills are orchestrators that spawn subagents (code-planner, code-skeptic, code-reviewer, code-debugger) and merge their output — skill → agent, inverted from every other domain. code also has no rules/ directory; instead its skills/agents read .memory-bank/index.md and swarm-report/<slug>-*.md inside whichever project they're running in.