How I Used AI to Reorganize 100,000 Files (And Why I Let It Make Decisions)

Amy Westlake
May 31
4 min read

Situation

I was going on vacation in a few days and my phone was full.

Not kind of full. 96% full. No-more-photos full. And the reason I couldn't clear it was sitting on my external hard drive: 25 years of photos, documents, music, and camera cards, scattered across 30-plus root-level folders, with backups of backups of backups underneath. I'd been putting this off for a long time. Every time I thought about starting, the scope stopped me.

I gave myself the weekend.

The AI Move

I didn't open Cowork first. I opened a regular Claude Chat thread and said I was thinking about using AI to reorganize a hard drive with around 100,000 files. Here was my actual first prompt:

We're going to do something fun and a little scary today. I have an external hard drive that is very unorganized and messy. I want to use Claude cowork to help clean everything up. However, I've never used cowork so this makes me nervous. This hard drive stores all of my photos from every single trip I've ever taken. There are copies of copies of copies of photos on there because I'm always afraid of losing them. But I know Cowork can help clean everything up. Before I go try to do anything in Cowork, I'd like your thoughts on how to best approach this since I've never used Cowork before. One golden rule to this entire process is that nothing ever is deleted. There is no circumstance where something should be deleted. I'm OK with files moving to an archive but nothing... I repeat nothing... should ever be deleted.

How should I approach this?

Then I started asking questions.

A lot of them. I kept coming back with "just one more thing." What should the folder structure look like? How should duplicates be handled? What happens when there's a quality difference between two copies of the same photo? What are the things that could go wrong?

I spent parts of two weekends working through every scenario I could imagine before I touched a single file.

At the end of that thread, I asked it to write me a starter prompt for Cowork. One with everything I'd decided already built in.

The most important rule: nothing gets deleted without my explicit approval. Not one file.

The Shift

I was nervous the entire time. More than I'd expected. I added "remember, do not delete anything" to my prompts more than once — not because I'd forgotten, but because I couldn't stand the idea of it losing that context. It never deleted a thing. It surfaced lists. I checked every list. Then I approved.

The moment I wasn't expecting came from Morocco.

I had 304 photos in my Morocco folder. A complete set, I thought. Then Cowork worked through a folder of unlabeled camera cards. It had been sitting in a backup for years, unlabeled. Cowork matched the EXIF dates. Morocco grew from 304 photos to 1,611.

I had the full Morocco trip. I just didn't know it.

That happened with other trips too. An Ireland folder I didn't know I had. A Dubai and India trip recovered entirely from camera card metadata. Files buried in a separate external drive that doesn't even power on anymore, pulled from a backup of a backup and matched to the right year, the right place.

By the time both sessions were done: 51,013 photos organized, 2,101 unique photos recovered, 43,052 file moves, zero data loss.

My photo collection, my pride and joy, from the last 25 years is actually complete now. I don't have a better way to describe what that feels like.

The Pattern

What made this work wasn't the capability. It was the structure I built before I started.

The rules I established weren't restrictions. They were the conditions that made delegation safe. Nothing deleted without approval. Work in batches. Flag uncertainty before acting. Surface edge cases (the borderline duplicates, the quality differences, the permissions walls) as questions, not decisions.

Every time Cowork hit something it wasn't sure about, it stopped and asked. I didn't have to monitor constantly. I just had to answer the questions it brought back to me.

That's the model. You're not stepping away from consequential work. You're staying in the loop at exactly the moments that require human judgment, and letting AI handle everything in between.

The Implication

You probably have a version of this somewhere. A folder you've been avoiding. A backlog that's been sitting long enough that opening it feels like a whole project.

The move isn't to open Cowork and start. The move is to think first. Build your requirements. Decide what "done" looks like and what the non-negotiables are. Figure out where you need to be the decision-maker. Write that down. Then hand it over.

The preparation isn't the delay. It's what makes the delegation real.

What I'm Testing Next

There's a companion post coming on the actual mechanics: the folder structure, how duplicate detection worked, how EXIF metadata sorts photos by trip when the filenames don't help. For people who want to try this themselves.

This post was for the question I kept asking myself in the days before I started.

Is this actually safe? Can I trust this with something that matters?

For me, the answer was yes. But only because I'd done the work to know what I needed it to respect.