Owners only.

The advisor never sets foot in your house. The contractor does real work in it - opens drawers (reads files), rebuilds rooms (edits), uses power tools (runs commands). That power is fenced by the permission system below, so nothing irreversible happens without you.

Claude Code runs in the desktop app (Mac and Windows), in a terminal, on the web, and in IDEs. The kit assumes the desktop app's Code tab.

Markdown (.md) files don't run like programs - they're readable notes that tell Claude how your business works. CLAUDE.md is the operating manual it reads first, context/ holds your priorities and team, .claude/skills/ holds the recipes.

Because it's all plain text in a folder you own, there's no lock-in: you can open, edit, or move any of it yourself.

Two kinds: native connectors you authorize with a click inside Claude (no keys), and API-key plumbing where a one-time key goes into ~/.claude/credentials/, a private folder outside your kit. Either way, credentials never sit in the chat or in your repo.

Plugins and settings install at a level too. When a plugin acts weird, ask Claude which level it's installed at.

This is also the privacy model: what Claude reads, it can send to the AI service to think with. What it never opens - like your credentials folder - never leaves your machine. You control what's reachable.

Every skill is a plain-text file you can read: .claude/skills/<name>/SKILL.md. Nothing is hidden.

Caught yourself doing the same thing twice? Just say it out loud: "Claude, make me a skill that does X." Claude writes the recipe, you save it, use it forever. That's how the system grows around your business.

Run /ceo weekly (or whenever you feel scattered) and it:

• Reads the strategic state and what changed since last week
• Surfaces what's on fire, what's drifting, what's working
• Picks this week's top 3 priorities - and names what you are explicitly NOT doing, which is half the value
• Briefs each Director in writing, so every department knows its focus
• Hands you the open decisions that genuinely need your call, with options

The rule underneath: the CEO surfaces and recommends, you decide. It never makes strategic calls or takes revenue-affecting action on its own.

Things to say to it: "Where are we against this quarter's target?" · "What should I focus on this week?" · "Brief the Marketing Director on the launch push."

A second computer. Your setting travels with your kit folder. The rules behind it do not, until the install step is run once on that machine. Until then your dashboard drops to the most careful setting and tells you why. That is on purpose, not a fault. Run /dashboard install there once and it settles.

A new app. When you connect a new app later, message me and I re-run the install step so the new app is held the same way as the ones we set up together. That is a support request, not a chore you own.

One kind of read we cannot block from a terminal. On your dashboard your passwords and logins are off limits to it. If you type a command yourself in a terminal window that reads one of those files, the engine falls back to its own asking rather than our block. We chose to tell you rather than round the claim up.

Sign-off steps are instructions, not walls. The steps each Director follows, keeping content in draft and running health copy through the compliance check, are instructions the OS follows rather than switches that stop it. Read what it hands you before you approve it.

Your dashboard has no skip-everything switch. There is nothing there for you to turn off. In a terminal window the engine still offers one, called Bypass permissions. Leave it alone: it is the one place your kit's own limits stop applying.

Exact prompt wording varies by action, but the pattern is constant: approve once, approve always for this, or reject.

The default is right for 90% of days. Change brains only when a task is unusually big, unusually small, or unusually stuck:

Effort is how hard the model thinks before answering (low to max). Raising effort beats switching models more often than you'd expect - /effort, or the slider in the /model picker.

Fast mode (/fast) keeps the same Opus brain but streams up to 2.5x faster. Note: it runs on usage credits rather than your plan's included allowance, so treat it as a paid turbo button for bursts, not an all-day setting.

Each session is its own worker with its own context. The one rule for many windows on one kit: tell each Claude what the others are doing if they touch the same files, and let /cleanup reconcile at the end. If two windows ever clash, tell Claude to clean it up.

/cleanup writes what happened - decisions made, files changed, where you left off - into the kit's memory and a handoff file, then commits it all. The next session starts by reading the handoff and picks up exactly where you stopped.

Think shift-change notes at a hospital: the incoming shift reads the chart, nobody re-diagnoses the patient. Context in a thread evaporates; documentation compounds. That habit is what makes the system smarter every week.

Mid-session tools exist too: /clear starts a fresh thread, and Claude can summarize a long one to carry it forward.