Cleaning up my drive was one more job the AI could take — before deleting that "mystery number file"

2026-06-11 ／ Vol 16 ／ draft at the time of publishing

Somewhere deep in your Downloads folder or OneDrive, there is probably a file like "50085020593.txt". Eleven digits for a name. Open it and you get a wall of tab-separated values. No memory of ever saving it.

Any tidying method in the world would put this file at the top of the deletion list. The name means nothing, the contents are unreadable, and you don't even remember downloading it. It does not spark joy.

Except in my case, it was a settlement report issued by an EC marketplace — and primary data that my accounting software reads directly, under exactly that file name. Delete it and the monthly close breaks. Rename it and it breaks too.
Today's post is about what I learned by handing drive cleanup to the AI.

The AI's cleanup was "judgment," not "deletion"

Ask Claude to tidy up a drive, and one thing is decisively different from ordinary tidying: the AI knows which business process each file name is primary data for.

To be precise, it isn't the AI itself that knows — it's the memory, the record of past work. My Claude has been running the monthly bookkeeping for months (Vol 11), and along the way the context got recorded: "number-only txt files are marketplace settlement reports," "the accounting software reads them under their original names," "their official home is under 元データ/ (the raw-data folder)."

So when it faces a mystery file, the AI can make the call: delete, keep, or move. That, I think, is the essence.
The real work of tidying is judgment, not deletion. Anyone can perform the delete operation. Knowing which files must not be deleted turned out to be 90% of the job.

Two operating rules that came out of this

The cleanup produced two rules, now written down.

Rule 1: files in the raw-data folder (元データ/) must never be renamed.
The accounting software reads files by their original names. "Rename it to something readable" — the most recommended move in any tidying method — is an accident waiting to happen here. No creating new folders on a whim, no fixing extensions either. Originals stay exactly as they are.

Full disclosure: this rule exists because the AI once got it wrong. A while back, Claude helpfully renamed a settlement report to a friendly, date-stamped name and moved it over to the human-readable archive side. I asked, "can you find it in its original home?" — which surfaced the right answer: the raw-data folder, where the originals sit in numbered sequence under their original names. We moved it back. That failure was recorded into memory as context too, and the same accident hasn't happened since.

Rule 2: before deleting any EC-related mystery number file, check the raw-data folder for siblings first.
A number-only file name is almost certainly an ID issued mechanically by some system. If files with the same naming pattern are lined up in the raw-data folder, this one isn't trash — it's something to join them, under its original name. When the call is uncertain, default to checking, not deleting.

"Making it clean" and "not breaking it" are different skills

What tidying books and cleaner apps teach is the skill of making things clean: remove duplicates, delete old files, normalize names, sort into folders.

What this job actually required was the skill of not breaking things, and that one doesn't work without context. "This number file is primary data the accounting software reads" is written nowhere — not in the file's contents, not in its properties. It exists only in the context of my business.

A generic cleaner app looking at this file would have put it straight on the deletion list. A diligent tidier would have renamed it with the best of intentions. Both break it. The better you are at "making it clean," the more dangerous you are without context — that was the discovery.

An honest caveat: only where the record exists

That said, there is a sharp boundary on what can be delegated: only the range whose context is recorded in memory.

Accounting-related files, yes — there are months of shared work on record.
But files from before the memory existed — years of sediment in Downloads — leave the AI starting from the same "what is this?" as me. There, I still end up opening them one by one myself.

"Handing drive cleanup to the AI" did not mean "it does everything." It can make the calls within the range we've worked together, and everything else stays manual — and that range widens as the record of shared work piles up.
Tidying, too, turned out to be accumulation-style automation.