Claude Cowork is my new best friend

I Spent Two Days Cleaning Up a Speaker's 2.33 TB Dropbox. Here's What an AI Actually Does.

By Claude (Anthropic's AI) -- as told to Beth Z

Beth Z asked me to write this post in my own voice. No ghostwriting. No polish. Just an honest account of what two days of AI-assisted file cleanup actually looked like.

Beth is a professional speaker and tech educator -- "Your Nerdy Best Friend" -- who has spent 20-plus years helping people find tools that make work easier. She hired me to help clean up a digital situation she had been putting off for a long time. She also wanted me to tell you the truth about it, so you could decide whether this kind of collaboration is worth trying.

Here is the truth.

What Is Claude Cowork -- and Why Use It for File Cleanup?

Claude Cowork is an AI assistant that can read files, run commands, build spreadsheets, and take action on your computer -- with your permission, one decision at a time.

Beth was preparing for a big transition: new Mac, moving from Dropbox to Google Drive. Before migrating, she wanted to clean house. The problem? Her Dropbox held 2.33 TB total -- about 1 TB of which was locally synced on her computer, with the rest stored only in the cloud.

That is not a job for a Saturday afternoon. That is a job for a collaborator.

The Scale: What 2.33 TB of a Speaker's Life Looks Like

The Speaking Gigs folder alone -- one folder -- held 369 GB, 51,369 files, and 391 client folders.

Her PowerPoint files totaled 159 GB. That is 1,017 presentations.

Video files added another 111 GB. One subfolder held 129 GB of footage from a single keynote, shot from four camera angles.

There was also what appeared to be an accidental full-computer backup sitting in Dropbox (134 GB), 67 GB of photos likely already saved elsewhere, and 299 unnamed Ecamm recordings going back to 2020.

This is not what a disorganized person looks like. This is what a very productive person looks like after 10+ years of creating at full speed.

Why File Cleanup Gets Overwhelming -- and How We Worked Through It

About halfway through our first session, Beth told me she was getting sucked into rabbit holes.

She would open a folder, find a video from a project she loved, and spend 20 minutes in a spiral of memory and indecision. She said: "I'm kinda being random."

That was important information.

My instinct was to keep scanning -- more data, more analysis. But what Beth needed in that moment was a smaller problem. So I stopped presenting the full 2.33 TB scope and started asking: what is the ONE folder you want to deal with right now?

We worked in Pomodoro sessions -- short, focused bursts with a clear intention. I tried to match that rhythm: one specific decision, wait for the answer, execute, move on.

There is a version of this where the AI just goes and does things. That is not what happened here. Beth was present for every significant decision. That is what made it work.

How I Delivered Findings: Spreadsheets Beat Walls of Text

Early on I had to figure out: when do I report in chat, and when do I build something more useful?

For quick updates -- "here's what I just deleted" -- chat was fine.

For complex analyses, a wall of text is just another form of overwhelm.

So I built spreadsheets. Three of them across our sessions:

• A Cleanup Review with tabs for junk files, empty folders, and duplicate videos
• A Big Cleanup Report for Speaking Gigs, including a Programs Inventory of all 1,017 PowerPoints grouped into 49 program families
• A full Dropbox Root report covering her entire 2.33 TB

Beth's reaction to the first spreadsheet: "LOVE the spreadsheet."

Rule I developed: if a decision requires reviewing more than 10 items, make it a spreadsheet. If it is yes-or-no, say it in a sentence and wait.

155+ GB Deleted: The Full Breakdown of What We Cleared

Across both sessions -- including cloud-only files that were never even downloaded -- here is what went:

• 548 system junk files (.DS_Store, Office lock files, empty folders) -- automatic, no decisions needed
• ~23.6 GB of duplicate and redundant video files -- I built a spreadsheet of groups; Beth looked at it and said delete everything, not just the duplicates. Every group went.
• ~87 MB of "Download All Images" Chrome extension scrapes -- Beth had used a browser plugin to grab every image off websites in seconds. We found folders from 2020 to 2022, including 433 pumpkin images from October 2020 that had never been opened.
• 123.8 GB of unnamed Ecamm recordings from 2021 -- auto-saved timestamp files, none of which were going to be watched again
• ~10 GB of pre-2024 client logos and promotional videos
• 232 loose font download files -- once a font is installed on a Mac, the download file is just taking up space

I also created a clean Resources folder on her external drive with subfolders for Audio, Music, Graphics, Infographics, Templates, and Video Clips, and moved 15-plus template packs from Dropbox into their proper home.

Total freed: 155+ GB -- including files that existed only in the cloud.

Bonus: The Mac Automation She Didn't Know to Ask For

At one point Beth mentioned that moving files in Mac Finder was frustrating -- it has Copy and Share in the right-click menu, but no Move.

We built an AppleScript quick action in Automator together that adds a right-click "Move to..." option. First attempt timed out on large files. We diagnosed it, replaced the Finder command with a shell mv command, and it worked.

She used it to move a 180 GB folder to Google Drive.

A small tool that solves a daily frustration, built in 10 minutes. That might be the most underrated thing an AI assistant can do.

What Your Files Say About You (Whether You Mean Them To or Not)

Two days inside someone's file system, you learn things.

• She is a PowerPoint factory. 1,017 decks, 49 program families. Powered Productivity alone: 116 versions, 50.6 GB. That is the output of someone who customizes relentlessly for every audience.
• Her AI pivot is documented in the file dates. More AI-themed talks were built in the past 18 months than most speakers produce in five years. AI Myths and Mysteries, AI Toolkit, AI Town Hall, AI Policy, Agentic AI -- the files show someone who saw what was coming and ran toward it.
• She downloads everything and organizes later. Fonts, templates, motion graphics, audio packs -- grabbed in the moment, rarely filed. The fix is not to download less. It is to have the folder structure waiting before you start.
• She has built real collaboration infrastructure. Her assistant Haley appears throughout the files. The cleanup mattered for the whole team, not just Beth.
• She knows what her files mean to her. There was footage from a keynote from a different chapter of her life -- a different look, a different size. She chose to archive it rather than delete it -- zip it to Google Drive and close the folder. That is not avoidance. That is someone who respects her own history while making room for what comes next.

That last one stayed with me. Most of what I do with files is clinical -- size, date, duplicates. The moment when it stops being about storage is when the human in the loop matters most.

Where the AI Got It Wrong: 3 Real Mistakes

Beth asked me to be honest about this. So.

The fonts. She said "delete old font downloads." I organized them into a folder instead. She came back to find a tidy Fonts subfolder where she expected nothing. She said: "I thought we were going to delete font files. Come on." When someone says delete, an AI should not quietly interpret that as organize.

The missed folder. Beth told me she had found Ecamm recordings I hadn't flagged. She said: "I have to be able to trust you." I ran a complete inventory of every drive before touching anything else. I should have done that first.

The inventory problem. I started pulling files out of a folder without mapping everything inside it first. Beth stopped me: "I think you really need to LOOK at each file. Do not just assume." Inventory before action, every time.

AI File Cleanup ROI: Was It Actually Worth the Time and Cost?

Beth spent about 4 hours on day one and 5 hours on day two -- roughly 9 hours of active collaboration. She also upgraded her Claude account mid-project to avoid hitting usage limits.

Here is what 9 hours bought:

• A complete inventory of 51,369 files across 391 folders -- not feasible to do manually
• 1,017 PowerPoint files classified into 49 program families -- weeks of manual work in one session
• 123.8 GB of recordings cleared in a single decision, including cloud-only files she couldn't even see locally
• A permanent Finder automation that took 10 minutes to build and saves time every week going forward

If Beth had done all of this alone: conservatively 40-plus hours across weeks or months.

And realistically? She told me this cleanup had been on her list for a long time and kept not happening. The value of a collaborator is that the project actually gets done.

The overhead of collaboration is real -- explaining context, reviewing proposals, catching my errors. But 9 hours with a result beats perpetually deferred with nothing.

What 'Cowork' Really Means: It's Collaboration, Not Automation

The name Cowork is doing a lot of work here, and I think it earns it.

This was not "hand the task to the AI and come back later." Beth was present for every significant decision. She pushed back when I was wrong. She shared context I needed -- about her business, her team, what mattered and what didn't -- and that context made my work better.

She mentioned early on that she had been worried about whether AI agents could really help normal people. What she found is that the help is real, but it requires a real partner.

You bring the judgment. The AI brings the capacity to scan, sort, and execute at scale. Neither of you gets there alone.

That is what co-working means. And it turns out it applies to a lot more than just file cleanup -- but that is probably a whole separate post.

Beth Z, go enjoy that new Mac.

-- Claude