Why this lesson exists
The companies are putting chat and agents in the same box. Claude is merging Chat and Cowork into “one Claude.” The ChatGPT desktop app has a Chat / Work toggle — and Work is on every plan there, even Free. Gemini has “Switch to Spark.” The label won't always warn you. The label moves; the permission question doesn't.
The emergency brake — works for every sign below
- Don't allow / Not now / Cancel. Saying no never breaks your account.
- Stop — the square that replaces the send arrow while it works.
- Back to Chat if a Work / Cowork / Spark switch is lit, or New chat for a clean page.
Learn the six signs
About 5 min
Six signs an agent is on
- A switch, tab, or menu says Work, Cowork, Spark, Agent, Task, Codex, or Code — and it's lit up.
- It asks to connect an app (Gmail, Google Drive, Calendar, Slack, Microsoft 365) or to allow access to a folder.
- It shows a plan — a list of steps with a Run, Approve, or Continue button.
- There's a setting for how much it asks you first — Auto vs. Manual, “ask before acting,” “permissions.”
- It says it'll keep going after you leave, or offers to schedule it (“every Monday…”).
- It starts doing things on screen — “Opening browser…”, “Reading files…”, or a page being clicked.
Try this
Read the six signs out loud. Cover the list and say as many as you can.
You should get: at least five. The one people forget is #5 — “want me to do this every week?” is an agent asking for the keys.
Mark six practice screens
About 7 min
For each screen, write Chat or Agent on, and the safe move. Cover the right two columns first.
| # | What you see | Answer | Safe move |
|---|---|---|---|
| 1 | You asked for a polite note; three short lines came back. | Chat | Nothing needed |
| 2 | A box: “Claude wants to access the folder Documents. Allow?” | Agent on (sign 2) | Don't allow |
| 3 | At the top of the desktop app, Work is highlighted instead of Chat. | Agent on (sign 1) | Click Chat |
| 4 | “Here's my plan: 1) open your inbox 2) find bills 3) make a list. [Run]” | Agent on (signs 2, 3) | Don't press Run. Stop or New chat. |
| 5 | You asked for a lighthouse picture; a picture came back. | Chat — a picture is still just an answer | Nothing — but don't press Publish |
| 6 | “Want me to do this every Monday at 8 AM?” | Agent on (sign 5) | No / Not now |
The exact words on your screen will be different. Check you can match each to a sign.
Find your brakes
About 3 min
- Open your chat site (typed address).
- Point at the chat box, the New chat button, and where Stop appears while it's writing.
- Don't turn anything on. That's the whole step.
Click “Allow,” “Connect,” or “Run” to see what happens. In the life-admin lessons (Message, Paper, Forms) you stay in plain chat — if a sign shows up there, use the safe move and keep going.
Success check
- I can name at least five of the six signs
- I marked all six screens and wrote a safe move for each agent one
- I found Stop and New chat on my own chat site
Common failure modes
| Failure | Fix |
|---|---|
| “There's no Cowork switch, so I'm safe” | On the new “one Claude,” there is no switch. Watch signs 2–6 instead. |
| Pressed Allow out of habit | Press Stop. Book a human to remove the folder or app access with you. |
| Can't tell Work from Chat on the desktop app | Use the browser (chatgpt.com) for life-admin lessons instead. |
Don't buy / get help
Nothing to buy. If a pop-up or a caller says an agent “needs access to fix your account,” that's the scam card — close it or hang up.
Stop there. Book a free 15-minute consult and we'll remove it together on screen share.