What we’ll cover, by section.

Claude Cowork: contents.

Cowork reads and writes your real files.

Point Cowork at a folder on your computer and it works inside it — reading the files the task needs, then creating and editing files right there. Chat keeps its work inside the conversation; Cowork works on the real files in your folder, on your machine.

Set the approval mode per task: Manual asks before each action, Auto pauses only if something looks unsafe.

Scheduled tasks: say it once, it runs on its own.

Chat and Cowork both keep a workspace. Here's the one difference.

Both a Chat Project and a Cowork folder are persistent workspaces that hold your files and context — so what's actually different? It comes down to what the work is: the conversation, or the files.

The tell — at the end of the session, what should have changed: the answer, or the files?

Where to find Cowork.

Cowork only opens folders inside three approved paths.

One of them is the folder where synced SharePoint and Teams libraries land.

On a Mac
OneDrive - PwC / Claude
Home folder / Claude
Library / CloudStorage / OneDrive-SharedLibraries-PwC
  (where synced libraries land)

On Windows
OneDrive - PwC \ Claude
C:\Users\<you>\Claude
C:\Users\<you>\PwC
  (where synced libraries land)

Sync a SharePoint folder, then open it in Cowork.

An old OneDrive shortcut blocks Sync: two ways to fix it.

Clicked Add shortcut to OneDrive (or Add shortcut to My files) on a library before? Sync then refuses with “You're already syncing a shortcut to a folder from this shared library.” Either fix works. Both start in OneDrive on the web › My files, at the folder with the small link icon.

Three common synced-folder issues, and how to fix each one.

Cowork grayed out in the Claude app? Reinstall Claude from Company Portal.

If you’re seeing this

1

···

2

• 1Cowork is grayed out next to Chat, so it can’t be picked.
• 2“Your current seat doesn’t include access to Code” when you open Code.

From a practitioner’s laptop. The Code message on its own is expected if Cowork works: Chat, Cowork and Claude Code are separate licenses at PwC.

Try these steps

1. 1Did you get Claude from the Claude website? Uninstall it. Not sure where it came from? Uninstall it anyway. WindowsSettings › Apps › Installed apps, then Uninstall next to Claude (or Control Panel › Programs). Deleting the shortcut isn’t enough. MacQuit Claude (Claude menu › Quit Claude), then drag Claude from Applications to the Trash.
2. 2Restart the laptop before you reinstall. Skipping this is why most retries fail.
3. 3Install Claude from Company Portal (the Start menu on Windows, Applications on a Mac): search for Claude, choose Claude Desktop and click Install. If it still shows as installed, choose Reinstall anyway.
4. 4Open Claude and check the org. The lower-left corner should show the PwC US org. Then click Cowork in the message box.

Seeing one of these messages? Reinstalling won’t fix it: submit a support ticket.

1When you sign in

Shows after you sign in, even in the Company Portal version of the app. The Claude enablement team fixes access on their side.

2When you pick a folder in Cowork

Shows for every folder, including the three approved ones, so it isn’t the same as “Folder not allowed”. It’s an administrator setting on the account or the laptop that you can’t change yourself.

AI ways of working: contents.

Before you type anything, know what folder you're in.

Every file in the folder you pick is fair game. Read the folder name under the message box before you type, every session.

Work out of a shared folder.

What goes in the shared folder: one example.

PwC/  ← synced SharePoint libraries
  US-INT-ADVS-Project Aurora - Documents/  ← launch Cowork here
    CLAUDE.md  ← read automatically, by everyone
    project-context.md
    stakeholders.md
    01 Diligence/
    02 TSA Costing/
    03 Day 1 Readiness/
    04 Synergy Tracking/
    05 Meeting Transcripts/  ← save Teams transcripts here
    06 Weekly Digest/  ← a scheduled task writes here
    skills/  ← the team’s skill files; add them in Customize
    working/<your-name>/  ← optional: drafts

What a bare-bones CLAUDE.md looks like.

# Project Aurora — carve-out support

# Context
For anything about the deal, read project-context.md.
For who's who and what they want, read stakeholders.md.
For the workstream plan, read plan.md.

# Where things go
Put drafts in working/<my-name>/.
Finished deliverables go in deliverables/.

# Who am I
I'm Glenn. Prefix files I create with "gt-".

One folder per project team. One chat per task.

Claude has access to everything in the folder you pick, and everyone with access to it sees what Claude saves there. Give each project team its own folder, so each team sees only its own documents.

↪ Before you start fresh, capture what matters to a .md in the folder. The folder is the memory that carries across chats; the chat is just the working space.

Every session starts fresh. What carries over is what you wrote down.

The model was picked before you typed a word. Read the name once, the same way you read the folder name.

Brief Claude the way you'd brief a teammate taking over the work.

If you don't tell Claude, it doesn't know.

Draft the steerco deck for Friday.
Use this week's notes.
Make it look good.

OK, I need the steerco deck for Friday. It's going to the seller's CFO. She reads the exec summary first, and what she actually cares about is the TSA run-rate, so lead with that. The point of the deck is to get a decision on the two open Day 1 items; it is not a status update. This week's call notes are in notes/, and the costing model in models/tsa-costs.xlsx is the number of record, so if the notes and the model disagree, the model wins. Match the last pack in templates/steerco-2026-08.pptx and keep it to five slides. One thing not to touch: the go-live date is agreed, don't reopen it. And if anything's unclear, flag it rather than guess.

Why one chat can't read a whole data room, and how subagents split the reading.

The context window is everything Claude can see at once: your instructions, every file it has opened and its own replies. A data room is bigger than that.

How to ask Claude to read a data room: say what you need, then split the reading.

Read the data room and give me everything for the standalone diligence.

I need the facts for the standalone diligence from the data room in Data room/. Before reading anything, list the folders, how many files each has, and the file types.

Then use subagents, one per functional folder. Give each one this brief:
- Read only your folder.
- Pull headcount and cost by function; services the parent provides and what they cost; contracts and systems shared with the parent; anything needing consent to transfer.
- One row per item: what it is, the figure, the file and page. Mark each confirmed or inferred.
- List every file you opened, every file you skipped and why, and any file you couldn't read.
- If two documents disagree, report both. Don't pick one, and don't guess.

When they're done, save each report to working/dataroom-review/, combine them into one tracker, and list anything in one folder that affects another.

The first draft is a starting point, not the final deliverable.

Some deliverables are one-shot, when the context is complete. Most need your judgment: that's the work, not a failure.

When AI gets it wrong, the output looks better, not worse.

The failure hides inside good writing. Six ways it happens, all of them on real engagements.

The counter is always something you can check: which documents it actually read, what it cut, a number re-derived from its source.

Three questions to ask any AI output on your engagement.

When you’ve done it twice, keep the method.

If it’s context you’d otherwise explain again, ask Claude to write it down. Next time, you tell Claude where to find it.

After a call: you heard it, Claude didn't.

If you don't tell it, it won't know.

A file, a chat message, a voice note. Any format works. The only way to get this wrong is not to do it.

What relaying a call actually sounds like.

Priya 4:15
- readout too long → 8 slides
- CFO is reading it, not deal team
- add Day 1 risks
- TSA costs as ranges
- Thurs 9am

Just got off with Priya. She's seen the draft readout and it's too long; she wants it down to eight slides. The reader has changed: it's going to the client's CFO, not the deal team, so less on process and more on what it costs and when. Two new things: a slide on Day 1 risks, and the TSA costs shown as ranges, not point estimates, because Tom's model is still moving. She wants it Thursday 9am.
Save this to the folder as the steer from today's call, then tell me what you'd change first.

The last line is the one that makes the rest stick: ask Claude to write the steer to the folder, so the next session and your teammate start from it too.

When things get busy, most people drop AI and go back to the old way.

We hear it a lot. But the fire drill is the moment AI should help the most, as long as your workspace is set up and the draft was built there.

If the first draft that went to the partner was built in the folder, every time-sensitive ask after it is a one-sentence change to something that already exists.

The five time-sensitive asks in action: click one, see the change.

None of these is a new brief. Each one is a change to work Claude already knows.

Ask for three versions of a slide. Pick one.

Same numbers, same folder, three ways to say it.

You choose in thirty seconds. You didn't write three decks.

The fundamentals on one page: drop these, build these.