I Connected Claude Cowork to My Gmail - 8 Slides That Show What Happened Next!

I documented this in a carousel because some things are easier to show than describe. Eight slides, each covering a different stage of what happened when I asked Claude Cowork to clean my Gmail inbox. This article walks through each one.

Slide 1: Connect

Connected Gmail to Claude Cowork in 30 seconds.

The first slide shows two things: the Connectors panel inside Cowork with Gmail listed under "Not connected," and the authorisation screen that appears when you click it.

There is no configuration involved. No API keys to generate, no developer settings to navigate, no documentation to read. You click Gmail in the Connectors list, a browser tab opens showing a straightforward screen explaining what Claude can and cannot access, you click Continue, sign in with Google, and it is done.

Thirty seconds. One click, one authorisation screen.

This matters because the barrier to connecting external apps is often what stops people from using integrations at all. Cowork removes that barrier almost entirely.

Slide 2: The Prompt

One instruction. Claude scanned 29,334 emails without touching a single one.

The slide shows the full prompt I typed:

"Connect to my Gmail inbox. I want to start organising it. Your rules are: NEVER delete any email without asking me first. Default action is always to archive, not delete. Start by scanning my inbox and giving me a summary of how many emails there are, the rough categories you see, and a proposed organisation plan. Don't touch anything yet."

Beneath it, Claude's response begins: "Here's what I found, Pritam. I've scanned your inbox without touching anything."

Total messages in the mailbox: 29,334 across 24,596 threads. Most recent 500 emails scanned. 490 of them unread.

The key thing this slide shows is that the instruction to not touch anything was followed precisely. Claude had access to the inbox. It chose to observe before acting because the prompt told it to. That distinction, between having permission and using permission, is what makes this feel different from handing over a password and hoping for the best.

Slide 3: The Plan

Claude mapped out every step before acting. Nothing touched yet.

Before proceeding, Claude produced a five-step organisation plan building on the label structure I already had in Gmail. Each step was clearly described, with the reasoning behind it:

Step 1 was to address Upwork notifications, which at 74% of inbox volume were the single biggest source of noise. Step 2 was to auto-archive transactional emails from banks and food delivery apps. Step 3 was to separate promotional emails from genuine shopping confirmations. Step 4 was to surface and protect the seven emails that actually needed attention. Step 5 was to label and archive everything else.

Then, before going further, it asked three questions: which step to tackle first, whether there were any senders to never delete under any circumstances, and whether there were specific client names to keep visible.

This is where the behaviour diverges from what most people expect from AI tools. It had a plan. It presented the plan. It asked for direction before executing any of it.

Slide 4: Always Asks First

Permission requested before accessing Gmail or Chrome.

This slide shows three things stacked together: a pop-up asking "Claude wants to use Google Chrome," a second pop-up asking "Allow Claude to use the browser on mail.google.com?", and a third showing the Gmail inbox with a banner reading "'Claude' started debugging this browser."

Every time Claude needed to take a new type of action, it asked. Not once at the beginning and then proceeded freely, but each time the scope of what it needed to do expanded.

"Claude started debugging this browser" is the line that stopped me. It is technically accurate and also genuinely strange to read in the context of your own email. It means Claude connected to the Chrome debugging protocol to interact with the browser programmatically, which is how it was able to select emails and click buttons in Gmail without a native API for those actions.

Seeing that line made the capability feel real in a way that a description of it does not.

Slide 5: To-Do

298 to archive. 187 to delete. Categorised and waiting for my go-ahead.

This slide shows the table Claude produced before touching anything in Step 2, the transactional noise cleanup:

Sender	1w–1m Archive	1m–6m Archive	Over 6m Delete	Subtotals
Kotak Bank	45	93	0	138 archive, 0 delete
Tata Neu	0	7	3	7 archive, 3 delete
Zomato	7	39	176	46 archive, 176 delete
BigBasket	5	21	6	26 archive, 6 delete
Tata 1mg	7	74	2	81 archive, 2 delete
Totals	64	234	187	298 archive, 187 delete

Total emails to be touched: 485.

The table shows exactly what the slide title says: everything categorised, everything mapped to an action, nothing moved. It sat there waiting for my confirmation.

I approved the plan.

Slide 6: Execution

Archiving and deleting one at a time. Reporting back after every step.

The slide shows the execution log as Claude worked through each sender. The entries read like a running commentary:

"46 conversations archived." Zomato done. Now BigBasket (26 to archive).

"24 conversations archived" and "No messages matched your search." BigBasket done. Now Tata Neu (7 to archive).

"3 conversations moved to Trash." BigBasket done. Now Tata Neu older than 6 months.

3 Tata Neu emails. Selecting all and deleting.

"3 conversations moved to Trash." Tata Neu done. Last deletion: Tata 1mg (2 emails).

2 Tata 1mg emails. Selecting all and deleting.

"2 conversations moved to Trash." All deletions complete. Now for the final step: creating inbox-skipping filters for all 5 senders.

What this slide shows is the reporting rhythm. After every batch, a confirmation. After every sender, a status update and the next target. The execution was not a black box that ran and returned a result. It was observable at every stage, which meant I could have stopped it at any point if something looked wrong.

Slide 7: Using My Computer

Claude took control of Chrome, navigated Gmail, and created the filters itself.

This slide shows the Cowork interface with a red banner at the top reading "Claude is using your computer," alongside Claude's running commentary as it worked through creating the filters:

"The filters settings page is open. Let me scroll to the bottom to find the 'Create a new filter' link and check existing filters."

"I can confirm BigBasket and 1mg filters are still missing. I can see the 3 filters from before (Tata Neu, Kotak Bank, Zomato) plus the 'Create a new filter' link."

"I can see the filters list and the 'Create a new filter' link. I can confirm 3 filters are saved. Now clicking 'Create a new filter'..."

Gmail's filter creation is not accessible via the standard Gmail API in the way that reading and archiving emails is. To create filters, Claude had to use Computer Use, taking direct control of the browser, navigating to the Gmail settings page, clicking into the filter creation form, typing the sender domain, checking the relevant checkboxes, and clicking Create filter. It did this five times, verifying after each one that the filter had saved correctly.

This is also where the tool showed its rough edges. Some clicks missed their targets slightly, requiring Claude to zoom in, identify the exact element coordinates, and try again. It handled those failures gracefully, describing what had gone wrong and correcting the approach, but they were visible. Cowork is not a finished product. It is a research preview, and the Computer Use feature in particular has moments where that shows.

Slide 8: Done

5 permanent filters live. These senders will never land in my inbox again.

The final slide shows the completed filter table:

Sender	Filter	Action
Kotak Bank	from:(@kotak.bank.in)	Skip Inbox + label "Kotak"
Tata Neu	from:(@info.tataneu.com)	Skip Inbox + label "Tata Neu"
Zomato	from:(@mailers.zomato.com)	Skip Inbox + label "Zomato"
BigBasket	from:(@bigbasket.com)	Skip Inbox + label "Big Basket"
Tata 1mg	from:(@mail.1mg.com)	Skip Inbox + label "Tata 1Mg"

All five senders now bypass the inbox and go directly to their own labelled folders.

The slide note underneath reads: "All old emails (1 week–6 months) were archived, anything older than 6 months was deleted, and going forward none of these senders will land in your inbox."

What the Eight Slides Show Together

Taken individually, each slide documents one stage of a task. Taken together, they show a sequence that is worth paying attention to:

Connect an external app in seconds. Scan without acting. Plan before touching anything. Ask permission before each new type of action. Present a to-do list with exact counts and wait for approval. Execute methodically with confirmation after every step. Use computer control as a last resort when no cleaner path exists. Finish by setting up infrastructure so the problem does not return.

That sequence, connect, scan, plan, ask, execute, report, build for the future, is not what most productivity tools do. Most tools perform an action when you tell them to. This performed a workflow: understanding the context, designing the approach, checking in at every decision point, and leaving the system in a better state than it found it.

The inbox is quieter now. These senders will not come back. And I was present for every decision without having to manage every action.