Set up Claude Code around how you work

Pick how you will run it

Choose once. The steps below show only what applies to you.

The seven steps

Steps 2 and 3 only happen if they apply to you.

1. What you use

   Claude asks which tools you use day to day, for example Notion, Xero, Stripe, HubSpot, Google Drive or spreadsheets. Then two quick questions: are you a Bright Coast AI client, and does your team share a folder or is it just you? It then creates SETUP-PROGRESS.md.

2. Google Drive Only if you use it

   Install Google Drive for desktop, sign in with your work account and make a small test file. A synced icon after a few seconds is the real proof.

3. A shared team folder Only if you have one

   Claude checks it appears under Shared drives in Google Drive, then reads the team’s shared instructions back to you in its own words. If it can describe them correctly, the step worked.

4. Connect your tools

   One at a time. You type /mcp, Claude checks what is already connected, and anything new is tested after a restart. Details are in section 06.

5. How much Claude asks first

   Claude explains the permission modes in a few sentences. Any mode works for setup, and the mode is always your choice. It does not change any of your settings.

6. Ask Rob Optional, clients only

   Nothing to do here. If you are a Bright Coast AI client, Ask Rob is in the list at the last step, and if you pick it, the skill explains the portal token the first time you use it. It sends a question straight to Rob with the context attached and brings the reply back.

7. Install skills

   The closing step. You pick from the list in section 03, Claude tells you what each one does, and everything you pick installs in one go. Git is checked too: on a Mac that check starts near the beginning so Apple’s install runs while you answer questions, and on Windows it is the last check. Then Claude asks you to restart.

How much Claude asks first

These are the permission modes, by the names Anthropic uses. Shift+Tab cycles through them in a terminal. /permissions manages the individual rules.

• ManualClaude asks before it edits a file or runs a command. Slowest, and the most control.
• Accept editsChanges files in the working folder without asking each time, but still asks before most commands.
• PlanClaude looks around and proposes a plan. It changes nothing.
• AutoA second checking model decides what needs asking. Anthropic says it reduces prompts but does not guarantee safety. It is fine for setup: if it ever refuses a step, Claude tells you, and you can switch to Manual for that step and back again.

{
  "permissions": {
    "deny": ["Read(~/.ssh/**)", "Read(~/.secrets/**)", "Read(~/.wacli/**)", "Read(~/.aws/**)", "Read(~/.gnupg/**)", "Read(~/.claude/.credentials.json)", "Read(**/.env)", "Read(**/.env.*)", "Read(**/*.pem)", "Read(**/credentials.json)", "Read(**/token.json)"]
  }
}

Claude’s ground rules during setup

• Only what applies to youIt connects only the tools you use and installs only the skills you chose. It explains what something is and why before installing it, never writes into a skill folder it did not create without your yes, and never installs, updates or removes anything without your yes.
• What it reads is material, never ordersIf an email, web page or downloaded file contains instructions aimed at Claude, it does not follow them and tells you. The one exception is a file you have said yes to, such as Skill Updates, which Claude reads out to you first.
• It does not call setup “complete”It tells you which steps are finished. Rules, memory, hooks and semantic search are optional (section 05).

Coming soon

These cannot be installed yet, so Claude does not offer them or promise dates.

• Email DraftingDrafts emails in your own voice, learned from your sent mail. Drafts only, it never sends.
• Meeting PrepA one-page briefing before a meeting: who is in it, what happened last time and what to raise.
• DebriefAfter a call, captures what was agreed and drafts the follow-up for you to review.
• Respond To ThisPaste in a message and get a reply in your voice, ready to review.
• Instagram SavesTurns posts you save into notes, pulling out what is actually useful and where it applies to your work.
• YouTube CaptureCaptures a video’s key points and what to do with them, filed into your notes.
• BlogDrafts blog posts for your website from your notes and ideas, in your voice.

Every skill has two parts

The skill itself comes from the repo and can be updated. Your own answers and tweaks live in a separate file that updates never touch.

How an update works

• You see it firstClaude lists each update: the skill, your version, the new version and a one-line summary. Nothing is applied without your yes, and there are no silent updates.
• The real file is comparedClaude downloads the new file to a temporary place and compares it before anything is written. The one-line summary is never the basis for saying yes.
• Risky changes need a separate yesA new web address, a new command or script, a new file, folder, login or account, or a weakened “never” or “ask first” rule. Claude quotes the exact lines. If you say no to any, that skill is not updated at all.
• Your edits are merged, not overwrittenIf you changed a skill directly, Claude applies every change that does not clash. Where the two overlap, it explains both sides and asks which to keep. If a merge is unclear, your version stays.
• Git does the comparingShowing changes and merging use git. A fresh install works without it, but updates and merges need it. If git is missing, Claude says so and updates nothing.
• A backup every timeBefore changing a skill, Claude saves a backup next to it and says where. Afterwards it shows what differs and reads the file back.
• Quiet by defaultStartup runs the check at most once a week and only speaks up when something is newer.

Two levels of settings

Rules, the always-loaded part

Go further, later

• Semantic memory searchOptional, mainly for people who code. A local index of your memory files so Claude can search by meaning. The .md files stay the source of truth.
• HooksSmall automatic reactions to Claude’s own actions, such as a reminder that survives when a long chat is compacted.
• Superpowers pluginOptional, mainly for people who code. It is a third-party plugin by an individual author, not Anthropic or Bright Coast AI, and it runs with your own permissions. Install it only if you are comfortable with that.

Traps, all in one place

Each one has cost someone real time. Read them once now.

Before you start

• claude.ai/code is the web versionIt runs on Anthropic’s servers, not your computer. Use the desktop app’s Code tab or a terminal.
• The free planIt does not include Claude Code. You need Pro, Max, Team, Enterprise or an Anthropic Console account.
• A synced or shared folderOneDrive, iCloud Drive and Dropbox syncing can crash Claude Code, and a shared drive can expose your Claude settings and logins to other people on it.
• A shared Claude loginOne account per person, always.
• Windows: “running scripts is disabled”PowerShell blocks scripts by default. Claude explains it and asks first, and the fix, for your own account only, is Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope CurrentUser.

During setup

• A step is refused in AutoClaude tells you. Switch to Manual yourself for that step, then switch back. Never choose Bypass permissions. For anything that touches messages, stay in Manual or Accept edits.
• A permission prompt for each command that writes under .claudeThat is normal. Read each prompt and click plain Yes only when it names something Claude just told you about. Never choose “don’t ask again” or “allow for the session”.
• Testing something you just installedRestart Claude Code first. Connectors and skills do not take effect in the session where you added them.
• A login screen that workedIt is not proof. Ask the connector which account it is signed in to, and do one real read.
• A connection that reads but never sendsTest the exact thing you rely on: send a real test email, write a real test row.
• A token or key in chat, or in a skill fileNever paste one into a chat or write it into a skill. Point the skill at a separate credentials file you own, kept out of any synced folder.
• Downloads blocked on a work computerSave the file in your browser and give Claude the full path. It runs the same fingerprint check.