I was in the middle of a live class when I tried to guess one month of screenshots. 150 to 200 files. The archive goes back to 2014, so you’d think I’d have a feel for it.
I asked the assistant to count. It came back with 361.
I didn’t sort them myself. I told it to inspect each image, give every file a descriptive name, and move it into a monthly folder. Then I went back to teaching. Renamed screenshots started showing up in the organized folder while class was still going. At the point I checked, the task was still running. I never verified that all 361 were done.
Have you ever been that wrong about a pile you look at all the time?
The target I organize around is finding what I need within seconds. The question I care about is whether I can reliably find this later. AI lets me shift the effort from filing everything by hand to defining how retrieval should work.
Duplicate the folder first
Screenshots are where I start. You can duplicate the folder and watch what the assistant does.
The instructions stay short. Group the screenshots into YYYY-MM folders. Inspect what each image contains. Rename each file with the date, then three to five keywords that describe the picture. Keep that date format identical every time. The keywords keep the filename useful when you browse the folder with no assistant open.
In class, the assistant interpreted the picture itself.
Run the first pass on a copy. Begin with default permissions. Observe what comes back. Expand access only once you understand the workflow. Preserve a backup. If something is going to be deleted, you approve it. Then compare the output with the originals on that first batch.
Screenshots can still contain sensitive information. A clean test still leaves prompt injection and data exposure in play. I’m no security researcher. Local file access is a separate question from local processing and from privacy. I demonstrated that folder access in Claude Cowork and in a Codex desktop app. Check current features and permissions before you copy the setup.

Let it execute, then repeat it at 5 p.m.
The step past chatting is a bounded task the assistant carries out.
I gave a desktop assistant one selected folder. It could read the files, create files, rename screenshots, and move them. I stopped explaining the system and delegated the execution.
Try it on one low-risk folder. Ask for an inventory before any renames or moves. Read what it lists. Then decide what it may change.
A cleanup you never repeat just waits for the next backlog. After the screenshot rules were in place, I created a daily 5 p.m. task to sort new screenshots with that same structure. I created that scheduled task in Claude Cowork during the class, and I said the app needed to remain open. Review the first few runs yourself.
Build the index, then ask something you already know
I’ve kept a long-standing note archive in Evernote. In that class I had the assistant turn it into a library I could search. It reported extracting 1,700 notes from 22 notebooks and creating an index. I was looking at Markdown notes and indexes. Those counts are the assistant’s report from the demo. I did not validate every note.
I opened a fresh task and asked for a summary of my December 2025 weekly reviews. It located the relevant material and summarized it within seconds. The speed was my observation from the room. I never ran it as a benchmark. I also have not independently checked every attachment or every sentence of that summary.
It held locations and metadata. The assistant consulted the index, then opened the source notes needed to answer. Build that map once. Consult it before you open a stack of documents.
Export one notebook. Ask for a table of contents of the folder. Test it with three questions whose answers you already know.
One category, on purpose
Where do you lose the most time looking for things?
Participants kept coming back to the same problem. Information scattered across multiple locations. I’m not naming anyone.
Pick one category you regularly struggle to retrieve. Meeting notes, screenshots, files, photos, or email. Gather a safe working copy, and decide the one trusted location for that material. Write down what successful retrieval looks like. A focused retrieval problem gives the first workflow a purpose you can test.
Email and photos need a slower start. Confirm the export, the access you are granting, and the privacy requirements. Direct folder access is something you verify for each app.
One suggestion, and I did not verify this one live. A small sample of your sent email can become a writing-voice profile. Redact it, and remove other people’s personal details. Read the profile it writes. Test the profile on an unsent draft. Leave the rest of the mailbox alone. Approve anything before it is sent.
Do this on a copy, this week
Duplicate one month of screenshots. Point Claude Cowork at that copy. Stay on default permissions. Ask for an inventory before it changes any names or locations. On a small batch, group the files into YYYY-MM folders, inspect each image, and rename the file with the date plus three to five keywords. Compare the batch with your originals. Keep the backup. You approve any deletion.
One copy. If the batch holds up, then you can talk about a second folder.
Recommended for you
Want the full system? 25X is the flagship productivity system we teach.
Explore 25X →