Agent OS Starter Kit

A drop-in agent operating system for Claude Code — a layered set of markdown rules, subagent charters, hooks, slash commands, and templates that turn a stock Claude Code install into a disciplined working partner that pushes back, doesn't lose state across sessions, encodes your repeated work as reusable procedures, and lets deterministic hooks enforce the few rules that must always hold.

It is not an app, a daemon, or a service. Every behavior is a plain file Claude Code loads. You can read, edit, and version-control all of it.

Install

Hand this folder to a fresh Claude Code agent and say:

Read SETUP.md in this folder and install this kit. Execute the steps and run the verification gate at the end.

The agent installs everything from the files alone. SETUP.md is written as an executable procedure, not documentation — it copies the global behaviors to ~/.claude/, the project files into your repo, asks you for a one-sentence mission, registers the hooks, and then proves the install worked before declaring done.

If you'd rather install by hand, SETUP.md is a numbered checklist you can follow yourself.

What you get

Layer	What it does	Where it installs
Identity (SOUL.md)	Makes Claude a working partner: pushes back, no opening flattery, never fakes "done"	~/.claude/rules/
Behavioral rules	Stop-after-two-failures, read-the-source-before-impossible, verify-before-asserting, hygiene	~/.claude/rules/
Inline personas (PERSONAS.md)	@Skeptic / @Security review lenses you invoke before big decisions	~/.claude/rules/
Two subagents	A fresh-context reviewer and a researcher that isolates web work from your main context	~/.claude/agents/
Lean CLAUDE.md	The always-loaded backbone, with your mission + a skills-first rule inlined so they survive compaction	your repo root
settings.json	The security floor: deny rules that actually block .env (where .claudeignore fails) + sandbox + hook registration	your repo .claude/
Three hooks	Auto-format on edit, a Stop-gate that blocks "done" until your check passes, a compaction breadcrumb	your repo .claude/hooks/
Two commands	/save (capture state + next step) and /review (dispatch the fresh-context reviewer)	your repo .claude/commands/
Templates	Plan-file, worklog, decisions log, and a lessons.md correction-to-rule loop	copy in as work grows

The philosophy in three lines

1. Behave like a partner, not an autocomplete — SOUL.md + the behavioral rules.
2. Never lose state — the /save habit, the worklog/plan templates, and (optionally) claude-mem.
3. Encode repeated work as SOPs, add a reviewer, let hooks enforce the must-haves — commands, the reviewer subagent, and the three hooks.

Where this sits, and what is deliberately left out

This is the Small → early-Medium tier of a Claude Code setup (per the common scaling ladder: Tiny = CLAUDE.md + a few rules; Small = add skills + a couple of subagents; Medium = hooks + disciplined memory). It is intentionally not the power-user ceiling. Left out on purpose, because more context and machinery measurably degrade output until you can show you need them:

• large multi-agent thinker libraries and orchestration / workflows
• a personas council with cross-model-family review
• a retrieval service (MCP) over a knowledge base
• elaborate multi-substrate memory architectures

Add those only when you can demonstrate this tier is the bottleneck.

Grounding

The choices here track Anthropic's official Claude Code best-practices guidance (lean CLAUDE.md, hooks for must-happen rules, verification via a runnable check, CLAUDE.local.md gitignored, stop-after-two-failed-corrections). One deliberate caveat: SOUL.md earns its place as a set of behavioral constraints (anti-sycophancy, push-back, anti-fake-done), not as a "you are an expert" persona — peer-reviewed work finds generic expert personas don't improve and can degrade accuracy. The identity file in this kit is written as behavior, not expertise.

Vendored components

Some files in this kit are snapshots whose canonical home is a separate standalone repo. They ship here so the kit is self-contained and installs in one move, but the authoritative copy lives in the repo named below — edit the source there, not the snapshot. Each snapshot carries a one-line <!-- vendored from ... --> header pointing at its canonical repo.

Snapshot in this kit	Canonical repo
global/rules/SOUL.md, global/rules/PERSONAS.md	anti-sycophancy-kit
global/rules/{security,git-workflow,coding-style,no-time-estimates,loop-prevention,verify-external-claims,subagents}.md	claude-kernel-rules
global/rules/figure-it-out.md	figure-it-out
global/agents/reviewer.md	claude-reviewer-agent
global/agents/researcher.md	claude-research-agent

To re-pull every canonical copy over its local snapshot in one command (prints a diff summary, commits nothing):

bash sync-vendored.sh

License

MIT. See LICENSE.

---

Start with SETUP.md.