FAQ & Safety Guide | The Modern Agent AI Workshop

Real questions real agents ask at every workshop — plain English, guardrails, permissions, and which tool is which. No engineering degree required, we promise.

The 60-Second Version

Claude doesn't train on your chats unless you say so. Default is a hard no.
You give the green light before anything real happens. Every single time.
Deleting a file always makes Claude ask first. No setting turns that off, ever.
Claude can't buy anything, open an account, trade stocks, or touch a full card or ID number. That door's just locked. Permanently.
The Chrome extension can see whatever's on your screen — so don't run it while you're logged into your bank or QuickBooks. Grab a separate browser profile for that stuff.
See a pop-up with options? Pick "allow once." You can always say yes again later — you can't un-ring the "always allow" bell.

Claude's Toolbox, Decoded

Agents mix these up constantly — like, constantly. So here's the short version: think of Claude like an assistant you just hired. Everything below? Just different ways of working with that same assistant.

Where do I even find Claude?

Claude lives in four different spots — like four doors into the same office. Same Claude, same memory, different door.

Where	What it is	Best for	Cowork?	Memory?
claude.ai (in your browser)	Claude on a regular webpage. Your home base.	Deep work — writing, planning, longer chats.	Yes	Yes
Claude in Chrome (the extension)	A small Claude panel riding along on whatever web page you're already on.	"I'm looking at this right now — help me with it."	Yes — opens straight into it	Same account, same memory
Claude Desktop (the app you install)	Claude installed on your computer like a real program, not a website.	Anything touching your actual files or apps.	Yes — the only place with the full toolbox	Yes
Claude Mobile (the phone app)	Claude in your pocket.	Quick, in-between moments.	Yes	Yes

"Cowork" and memory aren't separate for each one — it's the same Claude everywhere, just a different door in. What changes is what that door can reach: your files and computer only fully open up through the Desktop app.

What's an "AI agent"? Everyone keeps saying it.

You've probably heard this word all over social media lately, usually said like it's some mysterious new thing. It's simpler than it sounds: a regular chat is like asking a smart friend a question and getting an answer back. An "agent" is like handing that friend your actual to-do list and letting them go do the work — multiple steps, on their own — checking in with you along the way instead of just talking. Every tool in this list — Skills, Scheduled tasks, Cowork — is really just a different way of turning Claude from a chat into that kind of agent. It's also where this workshop gets its name: a real estate agent and an AI agent, both getting work done on your behalf.

What's a "chat"?

A chat is one thread of back-and-forth with Claude — what people mean when they say "I asked Claude." Thanks to memory (on by default), small useful details can carry over into your next chat automatically. But for anything you want Claude to reliably remember every time — your brand voice, your market, your standing instructions — use a Project instead of counting on memory to catch it.

What's a Project?

A Project is a folder you set up once, with your business info and instructions already inside it. Every chat you start inside that folder already knows all of that — you're not re-explaining yourself every time. Set one up with the "business brain" prompt from Lesson 2 and use it as your home base.

Learn more: What are projects?

What's a Skill?

A Skill is a task Claude has already learned, step by step, so you don't teach it again. Ask for it by name and it just runs — like a recipe card your assistant has memorized. Anyone can write a simple one in plain English, no coding required.

Learn more: What are skills?

What's a Scheduled task?

It's a Skill set to run on its own, on a schedule, without you asking. "Check my sphere every Monday morning" is a Skill — turn it into a Scheduled task and it just happens every Monday, even if your laptop's closed.

Learn more: Schedule recurring tasks in Claude Cowork

What's an Artifact?

Normally Claude just answers you inside the chat, like a spoken answer. An Artifact is when Claude hands you an actual finished thing instead — a document, a one-page website, a chart — in its own window you can open, edit, and find again later.

Learn more: What are artifacts and how do I use them?

What's Claude Design? Should I use it yet?

It's a blank canvas where Claude builds you a visual mockup — a poster, a landing page — that you click around and adjust by hand instead of describing in words. It's the most advanced tool in this whole list. Get comfortable with regular chats and a couple of Skills first before you try it.

What's a "credit"?

Every message uses a small amount of your plan's usage, which resets every 5 hours. Go over it on a paid plan, and credits let you keep working at standard rates instead of getting stopped. To stretch what you have: start a fresh chat for a new topic instead of one giant thread, keep your standing business info in a Project instead of retyping it, and save your heaviest work for right after a reset.

Learn more: Manage usage credits for paid Claude plans

How do I bring over what ChatGPT already knows about me?

You don't have to start from zero. In ChatGPT, paste this exact message: "List every memory you have stored about me, as well as any context you've learned about me from past conversations. Output everything in a single code block." Copy everything it gives you. Then in Claude, go to Settings > Memory > Start import, paste it in, and click "Add to memory." It's still an experimental feature, so take a minute afterward to glance over what got saved.

Learn more: Import and export your memory from Claude

Common Questions

What can Claude never do, no matter what?

Some rules aren't a setting you can flip — they're baked into Claude itself, the same way every car ships with brakes no matter who's behind the wheel. Claude's trained to spot a manipulation attempt even when it's dressed up official or urgent — so if a phishing email says "wire money right now," Claude treats it exactly like the red flag it is. It also scans documents, emails, and web pages before acting on anything in them, because scammers love hiding sneaky instructions inside a file, hoping Claude just follows along. And a few doors just stay bolted shut, period: no purchases, no opening accounts, no buying or selling stock, no handling a full credit card or ID number.

Learn more: Use Claude Cowork safely

What are Manual, Auto, and Skip mode?

Think of it like training a new hire — you decide how much they check in before acting. Manual: Claude asks before every single step. Slowest, safest, and exactly where you want to be for anything touching money, contracts, or a client's private info. Auto: Claude keeps moving but still double-checks itself and stops the second something looks off — good once you've built up some trust. Skip: Claude just goes, no asking (deleting a file still always asks, no matter what) — save this for the low-stakes stuff, never money or client info. Start on Manual. Level up to Auto when you've earned it.

Learn more: Get started with Claude Cowork

What do the permission pop-up buttons actually mean?

Allow once: Claude does the one thing, one time. Your safest bet, always. Allow for this task: Claude's cleared for the rest of what you just asked, then it resets next time. Always allow: a standing yes from here on out — fine for something repetitive and low-risk, never for money or client info. Decline: does what it says on the tin. A few things always ask no matter what you've picked before — deleting a file, typing something sensitive into a website, sending money, or messing with your own permission settings. Simple rule: client info, your money, or anything you can't undo? Allow once. Every single time.

Does Anthropic use my chats to train Claude?

Nope, not unless you say yes. Default is off — your chats hang around for 30 days, then they're gone. Opt in, and they can stick around up to 5 years to help train future Claude. Flip the switch anytime in Privacy Settings.

Learn more: Anthropic's consumer privacy update

Is it safe to paste my client's info into Claude?

Probably fine — just don't get sloppy about it. Black out account numbers and ID numbers before you paste, purely out of habit, even when nobody's making you. And remember: your brokerage or MLS might have its own rules stacked on top of whatever Claude allows. Those still apply to you.

What happens when I connect Gmail, Drive, or Calendar?

Connecting your Google account isn't handing over your password — it's more like a hotel key card. It opens specific doors, and you can cut it off anytime. Gmail: Claude can search, read, and draft right away, but sending or forwarding still asks first by default. Drive: reading and uploading, sure — sharing, moving, or trashing a file asks first. Calendar: a bit more freedom to create, change, and delete events unless you flip on manual approval yourself. Want to see or cut a connection? Customize > Connectors inside Claude, or myaccount.google.com/connections.

Learn more: Use Google Workspace connectors

Can I run Claude in Chrome while I'm logged into my bank or QuickBooks?

This is the one everyone asks about, so let's just be blunt: no. Claude in Chrome works by literally screenshotting the page to understand it — so if your bank, QuickBooks, or a health portal is open while it's running, Claude can technically see whatever's on screen. Same as someone reading over your shoulder. Banking, investment, and crypto sites get blocked by default or Claude stops and asks first — and it's never, ever allowed to trade stock or type sensitive numbers into a site, no matter what you've approved. Solution: a separate Chrome profile, no bank or accounting logins, and keep Claude over there instead. Same goes for QuickBooks Online — not technically a "bank," but it's your real numbers, so treat it the same. In healthcare? Claude in Chrome currently isn't offered to HIPAA-covered organizations at all.

Learn more: Use Claude in Chrome safely

Should I use ChatGPT to build my prompts for Claude, since it knows me better?

You really don't need to. Claude can write and sharpen its own prompts, right there in the same chat — that's literally what "Meet your Claude tutor" and "Your business brain" already do for you in the Skills Library. Already dumped a ton of info into ChatGPT and don't want to retype it all? Fine, ask ChatGPT to summarize what it knows about you in one paragraph and paste that in. But bouncing between two tools every single time just adds a step without adding much upside.

Where do I go for photo or image generation, since Claude isn't as advanced there yet?

True today: Claude can't generate images on its own — it can analyze photos, whip up simple graphics like diagrams, and write genuinely great image prompts. For real estate content, most agents lean on Canva (Magic Media, since you're probably already living in Canva anyway), ChatGPT or Google Gemini for quick photorealistic shots, or a dedicated virtual-staging tool. The real move: let Claude write the detailed image prompt, then hand that prompt to whichever image tool you're using. It's still doing the thinking — you're just handing off the last step to a specialist.

Can I just paste what Claude writes straight into an MLS listing?

Give it a compliance read first, always. Claude doesn't know Fair Housing law, and language that sounds totally harmless — "family-friendly neighborhood," "perfect for young professionals" — can accidentally cross into steering, which is illegal no matter how innocent you meant it. Treat every AI-drafted listing description, neighborhood blurb, or CMA talking point like a rough draft, not a final answer: read it the way your broker would before it goes anywhere public. Some MLS boards have their own rules about disclosing AI-assisted content too, so a quick check with yours is worth it. Quick disclaimer: this isn't legal advice — when it's unclear, ask your broker or compliance officer.

What's a sane way to set all this up when I'm just getting started?

Manual mode for anything touching money, contracts, or client info. Leave data training off until you decide on purpose to flip it on. Grab a separate Chrome profile — no bank, no QuickBooks, no medical logins — for anything you do with Claude in Chrome. Connect Gmail and Drive read-only first, and leave sending, sharing, and deleting on "ask me" until you trust the pattern. Check your connections every few months and cut loose anything you're not using. And whenever you're not sure? Allow once.

Words You'll See

AI agent: AI that does multi-step work for you instead of just answering a question — carrying out an actual task, checking in for your OK along the way. Cowork, Skills, and Scheduled tasks are all ways of using Claude as an agent instead of just a chat.
Guardrail: A safety rule built into Claude that can't be turned off, no matter what settings you pick.
Connector: The tool that lets Claude talk to an outside app, like Gmail or Google Drive, after you approve it.
OAuth: The technology behind "connecting" an account without giving away your password. Like a hotel key card instead of your house key.
Permission prompt: The pop-up box that asks for your OK before Claude does something new.
Manual / Auto / Skip mode: How much Claude checks with you before acting, from "ask every time" to "don't ask at all" (except deleting, which always asks).
Training data: Conversations a company uses to make their AI smarter over time. Anthropic only uses yours if you opt in.
Project: A folder-style workspace where you store your business info and standing instructions once, so every chat inside it already has that context loaded.
Skill: A repeatable task written down once, in plain English — you ask for it by name instead of re-explaining the steps every time.
Scheduled task: A Skill (or any task) set to run automatically on a repeating basis, without you asking each time.
Artifact: A finished, standalone thing Claude hands you in its own window — a document, mini-website, or chart you can open and reuse — instead of just text buried in the chat.
Claude Design: A visual canvas for freehanding a full mockup or page design by clicking around, instead of describing it in words. The most advanced tool here — worth waiting on until you're comfortable with the basics.
Usage credit: Extra usage you can buy at standard rates once you've used up your plan's included allowance for that 5-hour window.

Want the ongoing version?

Live questions answered, real workflows, and the "wait, is this safe" check before you try something new — that's what Tech Lab / AI Advantage is for, inside the Rebel Agent Posse.

Join the Rebel Agent Posse & Tech Lab