my AI deleted 28 files but I lost nothing

a folder in my vault had quietly grown to 38 files. daily notes, morning fixes, half of it resolved and the rest just sitting there looking like clutter every time I opened it.

I told claude: consolidate it.

it didn’t just DELETE everything. what it did was a clean demo of the one rules I set that makes a destructive operation like ‘deleting files’ safe to actually utilize.

"clean up this folder" is a dangerous instruction if you don't specify what's allowed to disappear.

some of those 38 files were genuinely stale. some of them were the only record of decisions I'd made weeks earlier. an agent that interprets "clean up" as "delete what looks old" is one bad guess away from deleting something that mattered.

before touching anything, claude read the vault's own conventions file first.
(that’s also a rule. read the ‘rules doc’ before you take any action)

one of my rules:

that one read changed the entire plan. it noticed the files in that folder was explicitly flagged as canon. if that read hadn't happened, the "obvious" cleanup would have deleted things that were important.

here's the part that made 28 deletions something I didn't have to babysit. before deleting a file, claude wrote the consolidated archive first, compressed, dated, indexed. then it read that new archive file back, in full, to confirm it had actually captured everything correctly. only after that verification passed did it delete the 28 stale originals it had just summarized.

never delete first and hope AI ‘will remember’.

I didn't have to worry about anything important being deleted. I had a system that proved it to itself before doing anything irreversible.

claude didn't just start deleting on its own judgment either. before doing anything, it gave me a multiple-choice question:

I picked "plan first, no deletions yet" and "leave that file alone." it built the full plan, showed me exactly what would happen to every file, and only executed after I said "do it."

that's the second half of the rule, and it's the part people skip when they're excited about automation. a destructive operation with a human confirmation gate in the middle isn't slower in any way that matters.

it's more the difference between an agent that acts and an agent you can actually hand real authority to, because the failure mode of "guessed wrong" gets caught before it costs you anything real.

38 files became 13. two clean archive files replaced 28 scattered ones, each old note compressed into a dated section instead of its own file. the folder's index got rebuilt too. that's its own lesson: the same append-only discipline that keeps a system's history intact is exactly what makes the index unreadable over time.

integrity and readability pull in opposite directions, and you need a deliberate pass to reconcile them.

none of this was about writing a smarter deletion script.

every piece of the safety here was procedural:

an agent doesn't need to be more careful in some abstract sense to be trusted with destructive operations. it needs a process that makes carefulness the only path through, so that trusting it stops being a leap of faith and starts being a checked fact.

38 files, 28 deletions, zero facts lost. the number that actually matters isn't 13. it's zero.

If closing gaps like this across your whole content system sounds like something you’d rather have built for you than DIY forever, that’s literally what I do.

SEE YOU ON THE NEXT ONE
- Auny 🧡