There are endless ways to set up Claude Cowork. I made a ton of mistakes with mine.
So here are my top recommendations for getting started right, so you do not have to make the mistakes I made. Five settings, in the order you will meet them. For each one, I also tell you when the opposite is the right call.
1. Put your work in a Project (this is the most important one)
A Project is a home for work you do more than once. It holds your files, your instructions, and Claude's memory of what you did last time. Anthropic's documentation is clear on the key part: a Project's memory is scoped to that Project, so what Claude learns in one Project does not carry over into another.
Here is why that matters. Work inside a Project, and Claude shows up already knowing your business, your files, and what happened last session, and that knowledge stays pointed at that one body of work. Your client work does not bleed into your personal admin. Your Q3 numbers do not wander into a different client's report.
One thing here has changed, and it is worth knowing so you are not working from an old mental model. A plain task outside a Project is no longer a blank slate. When a Cowork session runs in the cloud, Claude starts from what it already remembers from your chats, and what comes up in the task carries back to chat. Sessions that run locally on your computer do not use that memory. So the argument for a Project is not that nothing else remembers anything. It is that a Project's memory is aimed, and it stays where you aimed it.
The rule: anything you will do more than once goes in a Project. Your monthly report. Your client work. Your content. A true one-off, like converting a file or asking a quick question, is fine as a plain task.
Setting one up is quick: you give the Project a name and you pick the folder it works in. That folder choice matters more than people realize, and number 3 below is about exactly that. (The rest of Project setup is its own topic, and we will get there. For now: have one, and use it.)
And one hard-won tip, because this is the mistake I kept making: check that you are actually in the Project before you start working. It is very easy to open a plain chat by accident, do an hour of great work, and only later realize none of it landed in your Project, which loses the whole point. Make it a habit: start every session from inside the Project itself, and if you do not see your Project's name on the session, close it and start again from the Project's page.
Where this is wrong. A genuine one-off is faster outside a Project, and keeping one-time clutter out keeps your Project's memory clean.
2. Pick the middle approval setting
When Claude works for you, it takes real actions: opening files, editing documents, visiting websites. You get to choose how much it checks with you first, and there are three choices:
| The setting | What Claude does | What it costs you | When to pick it |
|---|---|---|---|
| Ask me first | Stops before every action and waits for your OK. Safest, and slowest, because you become the bottleneck. | In between. Cheap on your usage limit, but it costs the most of your own time. | Work that is hard to undo, like sending emails as you or spending money |
| Check its own work (the middle setting) ← MY RECOMMENDATION | Keeps working without interrupting you, but reviews each of its own actions and blocks anything that looks unsafe, like a hidden instruction in a file or webpage trying to trick it, or something trying to pull your data out. | Most expensive on your usage limit, because the checking is extra work Claude does. | Most everyday work |
| No checks at all | Just goes. Nothing reviews its actions automatically. | Least expensive, on usage and on your time. The catch is that nothing is watching. | Only if you understand exactly what you are giving up |
My recommendation for most work: the middle setting. You get the speed without giving up the safety net. The trap is that the middle and the last one both feel like "I trust it, let it run," and only one of them has anything watching. A lot of people are on the no-checks one without realizing it, because for a long time it was the only way to stop the constant approval prompts.
Switch to ask-me-first for anything that would be hard to undo, like sending emails as you or spending money. And whichever setting you pick, you are never locked out while Claude works: you can stop a task mid-run and redirect it whenever you want.
Stepping away from your desk and want the work to keep going? Two things. First, put that task on the middle setting, so Claude is not sitting frozen waiting for your OK. Second, and this is the part people miss, tell it up front: "I will be away, do not wait for me. Make reasonable choices, and leave me a note about anything you decided or could not finish." Claude stalls when it hits a question only you can answer, so hand it the answers, the files, and the permission to decide before you walk away. Deleting files is the one thing that always waits for you, in every mode.
Three things worth knowing:
The middle setting uses more of your usage limit, because the checking itself is work Claude does. In Anthropic's words, "Because Claude does this extra checking for you, auto mode consumes more of your usage limit than the other modes." Not a reason to turn it off, just the answer to "why am I hitting limits faster."
When it blocks something, it gets more careful. Claude looks for a safer way or asks you directly, and if it keeps hitting blocks it goes back to asking permission for each step.
Deleting files always asks. In every mode, including no-checks, Claude asks before permanently deleting your files.
Full honesty about my own setup: for much of my work, I work with the checks off. I am an unusually trusting early adopter, and I do not think most people should copy me. But I would rather tell you what I actually do than hand you a clean recommendation I do not follow. And here is the part that makes it work: in every mode, once the work is done, I ask Claude to check it. Usually a fact check, and often a review by a panel of experts I have Claude assemble.
Where this is wrong. If a task only reads and summarizes, and never sends, publishes, or overwrites anything, sitting through approvals costs you time and protects you from very little.
Check current, 2026-08-21. On the current build the three options read "Manually approve," "Automatically approve," and "Skip all approvals." If your screen still says "Ask before acting" and "Act without asking," you are on an older build with only two. Those first two labels are Anthropic's own renaming of the older pair. Labels move; the behaviors above are the durable part, which is why the rest of this section describes what each mode does rather than what it is called.
3. Make your working folder a Google Drive folder
Remember the folder you picked when you set up your Project? Claude can only read and write files in folders you have connected to it, so that choice matters more than it looks. Here is what I do, and it changed how much use I actually get out of my setup: my working folder is a Google Drive folder, through Google Drive for Desktop, so on the computer it behaves like any normal folder.
Why this earns a top-five slot. Everything Claude produces lands in Drive the moment it is saved, which means I can open the finished report on my phone from the car, check on a deliverable from my laptop, and hand a link to someone without hunting for the file. Nothing is trapped on one machine, and everything is backed up without me thinking about it. When the work lives where you already are, you actually look at it.
Where this is wrong. This one is not for everyone. If you work on one computer and never check anything from your phone, a plain local folder is simpler and works fine. If you use Dropbox or OneDrive instead of Google Drive, the same idea works with those. And two things should never go in any synced folder: passwords, keys, or anything secret (secrets stay local), and big technical folders like a software project's installed dependencies, which will grind the sync to a halt.
4. Ask for related things in one prompt, not one at a time
This one is not a setting, it is a habit. When several deliverables come out of the same work, ask for them together.
Ask for the research plus the doc plus the spreadsheet plus the deck in one go, and you get four things built on one set of facts. Ask for them as separate tasks, and each one does its own research and builds its own slightly different version of reality. The mismatch does not show up as an error. It shows up three weeks later, when the number in the deck does not match the number in the spreadsheet and nobody can tell which one was right.
Where this is wrong. When the jobs genuinely do not depend on each other, parallel is faster and you should run them side by side. The test is not whether the outputs look related. It is whether they need to agree.
Three more habits that pair with this one, learned the slow way:
- Keep sessions a reasonable size. When you switch topics, or a session has gone on for hours, start a fresh one. Long, sprawling sessions get muddier, and inside a Project you lose nothing by starting clean, because the Project carries the memory.
- Put a cost line in your Project instructions. Mine says, in effect: use the cheapest model that can do each step, and save the big model for the genuinely hard judgment calls. Formatting, lookups, and file wrangling do not need the expensive brain.
- Tell Claude it can run several agents at once. When work splits naturally, like researching three competitors or drafting four posts, give it standing permission to fan the work out in parallel instead of doing everything single file. It is dramatically faster and you approve the results the same way.
5. Describe the outcome and let Claude choose its own tools
Naming the tool in every prompt is a habit people bring from software, where being explicit is almost always correct. Here it mostly gets in the way, because you are specifying the method to something that can see the shape of the job better than you can from the outside.
Learn what the tools do anyway. Not so you can dictate, but so you can read back what Claude did and judge whether it chose well. That is a different and more useful skill, and it is the one that scales.
The other half of this one: tell Claude to ask you questions. Add a standing line to your instructions: if you need access to something, or you think I might already have a tool or system you do not know about, ask me before working around it. If I ask for better inventory tracking, the first thing I want back is "what system do you use to capture orders today?", not a solution built on a guess. Claude will happily build on a guess if you let it, and telling it not to is one sentence.
Where overriding pays. Say "use bash" when the job is dozens of files at once or real computation, and say "read the file directly" when you want one careful edit to a single document you care about. Those two overrides are worth keeping because they map to a real split in how the work gets done, not to a preference.
The bottom line
Here is the sneaky part about all five of these. When you get them wrong, nothing breaks. Everything still seems to work, so you never realize anything is off. You just keep wasting a little time every week and you cannot figure out why.
So grab twenty minutes today and set these five up. You will get that time back every week from here on out. I wish someone had told me this stuff on day one.
One topic I left out on purpose: connecting Claude to your private systems, like your email or an internal tool. It matters, and it deserves its own plain-language post rather than three confusing paragraphs here. That one is coming next.
Which of these five are you setting up first? And if you have a tip I missed, I want to hear it.
Sources
Verified 2026-08-21 against Anthropic's primary documentation.
- Get started with Claude Cowork, Anthropic Help Center. Project memory scope, the three approval modes and the renaming of the earlier two, parallel workstreams, the desktop-only path for plugins bundling local MCP servers, and the connected-folder file boundary.
- Use Claude Cowork safely, Anthropic Help Center. Screening behavior in the middle mode, degradation to asking after repeated blocks, absolute deletion protection, and the guidance to switch to manual approval for sensitive or hard-to-undo work.
The usage-limit sentence quoted in section 2 is Anthropic's own wording on auto mode's extra checking. The connected-folder file boundary in section 3 is from the Get Started documentation above.