Delegating File Cleanup and Repetitive Work to Agents

Automation with agents starts by picking closed-rule repetition — not “do everything.” Choose local versus cloud by permission, backup, and auditability. Prevent mistakes with dry-runs, samples, and rollback. Paid RPA/agent prices are omitted.

Replace “clean up my files” with S04: target, rules, done-check.

Which repetition fits?

One-line answer: First candidates have similar inputs, rules that fit in sentences, and a way to undo failure.

Good fits:

  • Propose subfolders in Downloads by extension/date (move after approval).
  • Build the same weekly table from logs/CSV with a fixed filter.
  • Fill issue templates/checklists with the same schema.
  • Rename screenshots with date and ticket id.

Too early:

  • Taste-heavy labels (“only important mail”).
  • Jobs that combine ACL changes, payment, delete, and external share.
  • Exception-heavy work whose rules change every week.

Instruction skeleton: “goal / target paths / allowed ops (read, copy, move) / banned (delete, external send) / done-check (counts, samples) / rollback.”

Automate one kind per week. Document the winning rule, then take the next kind.

Local vs cloud?

One-line answer: Prefer local (or self-hosted) for secrets, bulk, and offline. Prefer cloud when collaboration, inherited ACLs, and audit matter — still at least privilege.

Selection cues:

  • Local / ADE worktree: source, configs, personal notes on disk. Easy to bound the agent to a directory.
  • Cloud (Drive, etc.): team shares, link ACLs, version history — and higher cost of a bad share.
  • Hybrid: clean and verify locally, upload results only; humans manage source ACLs in cloud.

An ADE such as Orca helps isolate repetitive scripts/agent sessions in a worktree. When attaching cloud MCP/API, start read-only; open write/share in a separate session.

Do not pin plan prices here. Prefer scope limits, audit logs, and rollback over sticker price.

How do you prevent mistakes?

One-line answer: Always dry-run (proposal list) → human sample of N → execute → keep a rollback path.

Required guards:

  • Dry-run: current→planned table before move/delete/rename. Stop if counts look wrong.
  • Samples: random plus edge cases (long names, spaces/Hangul, permission-denied files).
  • Quarantine/backup: isolate instead of delete; confirm cloud trash/version restore.
  • Denylist: .env, keys, wiping node_modules, production buckets.
  • One direction: do not mix moves and external sharing in the same batch.
  • Logs: what changed, so next week you fix the rule.

Sample done-check: “50 proposals, 5 samples approved. After execute, counts match. One item restored from quarantine within 10 minutes as a drill.”

Minimum path: (1) write the rule, (2) dry-run table, (3) sample approve, (4) execute and log. “Finish it in one shot” like a search query makes recovery harder than search.

Sources

  • S03/S04 applied to repetitive file work.
  • RPA/agent/cloud add-on pricing: vendor pages (not pinned here).