# The Safety & Privacy Checklist

*Run this before you connect a service, bring in data, or turn on an automation. A minute now saves
a bad afternoon later.*

## Before you connect anything
- [ ] **I weighed value against blast radius.** I'd make this connection with my eyes open. (The *blast radius* is how far the damage would reach in the worst case.)
- [ ] **I'm connecting read-only** — Claude can see but not change or send — unless I have a clear reason not to.
- [ ] **Least privilege.** I'm granting the smallest scope that does the job — one folder or account, not everything. The scope I grant is the blast radius I accept.
- [ ] **The data isn't too sensitive for this.** I'm comfortable with a careful assistant that occasionally errs being able to read what this connection can reach.
- [ ] **I know what "approve on first use" will look like**, and I'll read the request before I approve it.

## Where secrets go (never in a prompt)
- [ ] **Any key, password, or credential goes in the connection's settings/config** — never typed into a prompt or a chat message. A secret is a house key; I don't tape it to the door.
- [ ] **I told Claude to keep secrets in config** and never to ask me to paste one into the conversation. If a step asks me to paste a secret, I stop.
- [ ] **I know how to revoke** this connection's access if I ever need to.

## Before data leaves my computer
- [ ] **I know what data is now leaving my computer**, and I've checked that nothing more sensitive than the job needs is going where it shouldn't.
- [ ] **The connection reaches only what the task requires** — not my whole account or drive.

## Before an automation acts on the world
- [ ] **There's a human gate.** Anything that sends, posts, changes, or moves stops for my approval first.
- [ ] **I sent it to myself first**, or ran a dry run, and watched it behave before letting it act for real.
- [ ] **Approvals are on.** I'm in a mode that asks before acting, not one that acts without me.

## Cost and usage awareness
- [ ] **I've accounted for cost.** I know automated runs draw from my usage pool. (Prices and limits change — I check claude.com/pricing.)
- [ ] **The interval is as long as the job allows**, so it's not running more often than it needs to.

## Before I let it run on its own
- [ ] **There's a kill switch** I know how to reach, to stop everything fast.
- [ ] **I can read the log** to see what it did while I wasn't watching.
- [ ] **A person stays in the loop** for anything with real consequences.

## What never to connect (or automate) casually
- [ ] Nothing that **moves money** runs without me.
- [ ] Nothing **sends irreversibly on my behalf at scale** without a human gate.
- [ ] My **most confidential records** aren't reached for a thin convenience.
- [ ] Nothing gets a **larger, orchestrated, unsupervised** setup — that's Book 4, *Advanced Systems*, not something to bolt on here.

## If a connection or automation misbehaves
- [ ] I **pause the automation first**, before anything else goes out.
- [ ] I ask "**what changed?**" and hand the exact symptom to Claude to diagnose.
- [ ] If a fix gets tangled, I **restore from a save point** — no drama.

> The golden rule: **you keep your hand on the switch.** Nothing connects without your say-so;
> nothing acts on the world without a checkpoint where you get to approve it. When in doubt, stay
> read-only, keep secrets in config, and slow down.