Claude Code Operating System

An assistant that
works for you,
not from scratch.

A complete, working Claude Code setup, packaged so you can run it yourself. Seven layers that turn a forgetful contractor into a team member who knows your business, follows your rules, and runs jobs while you sleep.

Plain text. Markdown, YAML, JSON, shell. Scrubbed. Zero private data. ~30 min to stand up the core.

The shift

Vanilla Claude Code

  • Re-brief it on your world every session
  • Corrections forgotten by tomorrow
  • "Write like an account manager would"
  • Dangerous commands only discouraged
  • Stops working when you close the laptop
  • Does what you ask, even if it's busywork

This setup

  • Starts already knowing your orgs, people, brands
  • Corrections become permanent memory
  • "/am draft the renewal note" - the role is pre-loaded
  • Whole classes of disaster mechanically blocked
  • Scheduled jobs run overnight, you wake to finished work
  • Flags stalled goals and drift from what matters

How it runs - seven layers

Each layer is simple. They interlock so a single request draws on all of them at once.

1CLAUDE.mdThe constitution. Loaded every session: identity, routing, hard rules. Short on purpose.
2Rules + PermissionsWhat's allowed, denied, or always-asks. Send-tools always prompt; .env and keys hard-blocked.
3HooksShell guardrails that run every time, outside the model. Make mistakes impossible, not just unlikely.
4Skills + CommandsNamed personas and one-command workflows. Delegate a role, not a paragraph of instructions.
5MemoryOne fact per file, recalled on relevance. The engine that kills the amnesia.
6Goals + TasksWhat matters now. The model surfaces overdue work and pushes back on drift.
7Scheduled TasksRecurring autonomous jobs on a cron. The layer that works while you're away.

What you actually get

1

It already knows your world

No re-briefing tax. Constitution plus memory mean every session starts informed.

2

Your corrections stick

Feedback becomes a memory file or a hook. It gets more aligned to you the longer you use it.

3

Disaster is impossible

The firewall blocks rm -rf /, force-push to main, DROP TABLE, curl-pipe-sh, reading .env - deterministically.

4

Roles, not paragraphs

Skills carry their own reporting line, decision rights, and templates. You delegate a team member.

5

It works while you sleep

Scheduled jobs pull from your tools and deliver finished summaries and fixes overnight.

6

It keeps you on-target

Reads your goals every session. Honest about your time, not just obedient.

What's in the box

read
docs/ + README.mdFive guides: advantages, architecture, setup, customize, inventory.
install
templates/CLAUDE.mdThe global constitution. Fill in the brackets.
install
templates/hooks/4 guardrails: bash firewall, auto-format, no-emdash, notify.
install
templates/settings.jsonSanitized permissions, hook wiring, plugins.
install
templates/memory/The anti-amnesia system: index + fact-file template.
install
templates/skills/Two patterns: role-agent and workflow.
install
goals.yaml + my-tasks.yamlWhat the model holds you accountable to.
run
personalize-setupThe interactive interview that builds your config for you.
35+ skill patterns 4 hooks 80-file memory model 8 scheduled jobs goals-driven no lock-in

It installs itself by interviewing you

  1. Run bash bootstrap.sh - backs up any existing config, installs the hooks, rules, and safety core.
  2. Open Claude Code and run /personalize-setup.
  3. Answer a short series of questions: your name, your orgs and their keywords, your stack, how much autonomy you want, your house-style rules, your top goals, the role you delegate most.
  4. Claude generates your CLAUDE.md, goals.yaml, memory seeds, and first skill - showing you each file before it writes it.
Nothing personal is written until you confirm. Prefer to do it by hand? Full manual walkthrough in docs/3-setup.md. Read security-notes.md first.