Best practices for working safely

Estimated time: 8 minutes

Learning objectives

By the end of this lesson you'll be able to:

Set up your workspace so the important stuff stays protected
Write prompts that don't leave room for the wrong action
Recognize the moments when "stop and think" matters more than speed

Your part in working safely with Cowork

You already know from Lesson 2 that Claude always asks before deleting, and — in the default permission mode — before sending or sharing too. That's the floor. This lesson is what you bring on top — the pre-emptive moves that keep an autonomous tool with file, app, and connector access from doing something you didn't intend.

Set up so mistakes can't reach what matters

The single highest-leverage move is the folder you point Claude at. It's the boundary for what Claude can read, write, and (with your confirmation) delete.

**Use a dedicated working folder, not a catch-all. ** Pointing Claude at Documents, Downloads, or Desktop is the equivalent of letting a new colleague rummage through every file you have. Make a folder for the work. Move (or copy) in what's needed. Point Claude there.
**Back up anything irreplaceable before you start. ** If a file matters and a fresh copy can't be regenerated — old client deliverables, contracts you can't get re-issued, anything you'd be sorry to lose — make sure a copy lives somewhere Cowork can't reach. Cloud backup, a separate folder, a drive that isn't connected. Claude won't delete without asking, but the cost of clicking through the wrong confirmation is the cost of the file.
**Test new workflows on copies first. ** For example if you're building a scheduled task that'll run every Friday, the first run goes against a copy of the data. Once you've seen it behave, point it at the live folder.

Write prompts that leave no room for the wrong action

How you ask matters as much as what folder you point at.

**Be specific about destructive verbs. ** "Cut the section" can be read as "remove from view" or _"delete from the file. "_ "Update the file" can mean "rewrite it" or _"add to it. "_ If the wrong reading would be irrecoverable, name what you mean: _"Remove the section from the draft, but keep the file. "_ _"Add a new appendix; don't rewrite the existing sections. "_
**Name the bounds in the prompt. ** _"Only the 3 most recently updated files in this folder. "_ _"Only contracts that closed in Q3. "_ _"Don't message anyone — draft only. "_ This narrows what Cowork is doing and gives you a clear line for spotting drift.
**Use scheduled tasks for drafts initially. ** Scheduled tasks run while you're not watching. Until you're confident that the task runs the way you need it to, prompt it to draft for your review rather than send on your behalf.

In the moment: the three checks that catch the rest

**Read the plan once it has been made. ** When Claude starts a task, it lays out what it's going to do in the progress tab. Skim it. Consider: does the plan make sense? are the steps in the right order? is it using the right sources? Redirect as necessary.

**Watch for unexpected patterns. ** You don't need to validate every command. But if Claude is touching files or sites you didn't mention, or scope is creeping past what you asked for, stop the task. "Something feels off" is a real signal — pay attention to this.

**Approve confirmation prompts deliberately. ** Stay in “Ask before acting” for anything that sends, posts, or shares — and when a confirmation prompt does appear, read it. Most mistakes don't happen because the safeguards failed; they happen because someone clicked through a confirmation that wasn't quite the action they intended. The dialog exists because the action matters — treat it that way.

When Cowork isn't the right tool

A short list:

**Regulated workflows that need an audit trail. ** Cowork activity isn't captured in audit logs, the Compliance API, or data exports.
**Anything you wouldn't trust a smart, quick colleague to do unsupervised. ** Sending the legal doc to a counterparty, posting the public announcement, pushing a customer-facing change. Claude can prepare; you ship.
Highly sensitive personal data outside the boundary your IT team has explicitly approved.

Go deeper

Use Claude Cowork safely is a valuable resource for learning what to do — and not to do — when working with Claude in Cowork. It covers the rest of what warrants more thought so that you understand the built-in guardrails and what you remain responsible for as Claude acts on your behalf.

Lesson reflection

Review the interactive below to get a sense for how to work safely with Cowork.

This Cowork task is about to run. **Five things about it would make a careful colleague pause. ** Find them.

Hazards found 0 of 5 Risky Safe

Connectors

Microsoft 365

Claude in Chrome can act on pages in tabs you've allowed it to access 1

Asana

Clean up last quarter's client files3 — archive the old stuff4 and post a note in each client's Slack channel that their folder's been tidied5.

Scheduled · Every Friday 5pm

~/Documents/Work Ask Sonnet 4. 6

Click anything in the window that looks risky.

All five found. Now watch it become a task you'd run unattended.

Make it safe →

Same job. Nothing here can't be undone.

Look at the connectors, the folder, the prompt wording, and what runs when. Click anything that gives you pause.

As you think about the tasks you're going to hand to Cowork:

Which folder would you point Claude at — and is there anything inside it that shouldn't be reachable, or anything irreplaceable that needs to be backed up first?
Is there a destructive verb in the prompt you'd write that you'd want to be more specific about?

What’s next

In the next lesson, you'll learn how to make sure the skills and plugins you build actually behave — using lightweight evals to check their output before you rely on them or share them with anyone else.

Feedback

As you progress through the course, we'd love to hear how you're using concepts from it in your work, plus any feedback you may have. Share your feedback here.