Claude Code Is a Tedium Solvent

Claude Code is a coding tool. I have used it for four things this year that are not coding.

I organized 120 files and 11 GB of wedding media out of my Documents folder. I swept 35 stray items out of my home directory and reclaimed 9 GB. I audited my Obsidian second-brain and consolidated hundreds of orphan notes. I sorted 473 credentials in 1Password into six new vaults, archived 21 dead accounts, and renamed 37 multi-account titles.

None of that is what Anthropic built Claude Code for. All of it worked. The pattern that made it work is not specifically about AI. It is about one specific friction point.

The Problem AI Actually Solves Here

Organization work fails at scale for one reason. The decisions are individually cheap and collectively infinite.

You know where the file goes. You know the old credentials can be archived. You know the email thread can be deleted. Each of those is a two-second decision. You have 500 of them. That is sixteen minutes of thinking and two hours of clicking.

Humans quit around minute forty.

We quit because the individual decisions, while cheap in isolation, cost full cognitive context to evaluate. Each file forces you to remember what the project is, where it belongs, whether the destination folder exists, whether this is a duplicate. The reload cost is constant. The mental tax is linear in the number of items. Humans are not built for this.

This is the gap AI tooling actually fills. Not replacing judgement. Absorbing the cost of exercising judgement 500 times in a row.

Four Episodes

Each of these cleanups looked different on the surface. Underneath they were the same shape.

Documents folder (March). ~/Documents/ had 48 items. Numbered folder system fighting generic file-type dumps. 6.7 GB of forgotten ML models. Three copies of my own app's installer. I wrote a structured cleanup plan, pointed Claude Code at it, and watched 120+ files get sorted, verified byte-for-byte against an external backup before any deletion, and routed to their numbered homes. Two hours. 11 GB freed.

Home directory and Obsidian vault (March). Home had 35 items that had no business being there. Downloads had 70 files, three of them the same DMG. Pictures had 9.3 GB of wedding media. The vault had hundreds of orphaned notes. Same pattern: audit, back up, propose, execute. Same outcome: drastically smaller footprint, every file with a reason to be where it is.

Emails (April). My Gmail was thousands of threads deep. Hundreds of unread. I wrote a CLI wrapper (claude-gmail) so Claude Code could search, label, archive, and draft through IMAP and the Gmail API. Batch label rules. Draft with dry-run. Context-tagged activity logging that mirrors into my second-brain. The email chaos did not survive the week.

1Password (April). 494 items in a single Private vault. Eight years of passwords, eight years of never pruning. A Python classifier routed items to six new vaults based on title, username, URL, and category patterns. An offline markdown proposal for me to review. A batch mover with audit log. 119 items repositioned. 21 archived. 37 renamed.

Four domains. Different data, different CLIs, different destinations. Same shape.

The Shape

Five steps. Every session that has worked has followed them in this order.

1. Audit before acting. Inventory the current state. Count items, categorize by type, surface duplicates. Write the summary to a file, not a chat message. Files do not scroll off the top.

2. Back up before mutating. Metadata snapshot to the second-brain vault, git committed. Not the secrets. Not the sensitive content. The structural information needed to reverse any move. This step takes five minutes. It is the step I would skip if I were doing this myself. Claude Code proposes it every time, unprompted.

3. Propose before executing. Produce a markdown file or a table in the chat with every proposed action. Group by destination. Include the reason. Let me edit any cell before anything runs.

4. Execute in small batches with dry-run first. Never all 500 items at once. Ten to fifty per batch, with the exact commands printed before execution. Approval gate on destructive actions.

5. Log every mutation. Timestamp, operation, item identifier, before and after states. One line per action. Append-only. Greppable six months from now when something feels off.

These steps are not specific to AI. You could do them yourself. The point is: you will not. I would not have. The overhead of setting up the audit log, the backup, the proposal file, the execution script is what makes organization feel like an engineering project, which is what makes humans avoid it.

Claude Code does not mind.

The Boundaries That Matter

Three rules I have landed on. Every session observes them.

Archive, never hard-delete. If 1Password or Gmail or the filesystem supports "soft delete with recovery," use that. If it does not, move to a backup directory first and verify before removing. The cost of keeping archived items forever is trivial. The cost of losing one important thing is not.

Human decides, AI executes. Every destructive action is gated on approval. The AI will not delete anything without me saying the specific word. "Follow the plan" does not mean "delete." "Archive the dupes" does. This rule survives every session.

Flag instead of guess. When Claude Code sees a credential owned by a teammate from a job I left, it flags and asks. It does not try to be helpful by archiving it autonomously. It does not try to figure out whether "Wifi Isabela" is Isabela the person or Isabela the network. It asks. The number of disasters this pattern has prevented is not measurable. It is not zero.

Where This Does Not Work

The pattern does not work for tasks where the decisions are not cheap.

Migrating a codebase across a framework is not a tedium problem. The individual decisions are hard. Rewriting a failing test suite to passing is not a tedium problem. You need to understand why the tests were failing.

The cleanup pattern is for work where you already know what you want, there are too many items to face manually, and the tooling to do each action exists. When those three conditions hold, Claude Code turns hours into minutes.

When they do not hold, it just helps you write code.

The Broader Implication

For a decade the promise of "digital organization" has been productivity apps. Notion, Obsidian, Things, OmniFocus. Every one of them makes the capture and the view better. None of them address the actual bottleneck, which is that periodic cleanup requires the patience of a saint or the tedium tolerance of a data-entry clerk.

I am neither.

What AI tooling does, when applied to these domains, is turn cleanup from a two-hour boring task into a fifteen-minute conversation with a file manager that happens to speak English. That is not a breakthrough. That is a small, mechanical reduction in a specific kind of friction.

But small reductions in friction, applied to work you keep avoiding, compound faster than any productivity app I have ever adopted.

My Documents folder is clean. My emails are triaged. My vault is organized. My 1Password is sorted.

None of these stays clean forever. Entropy is the rule. But the recovery cost went from "weekend project that never happens" to "afternoon session when I feel like it." That is a different life.

The next time someone tells you AI is overhyped, ask them what they have actually organized with it.