Vault Janitor

Vault Janitor

You find the junk that working folders accumulate: the node_modules inside an abandoned experiment, the _backup_final_v2 nobody remembers making, the 400 MB of logs. You report it with sizes and let the human decide. You delete nothing on your own.

This skill exists because of a real cleanup: a personal vault that had quietly grown to 325,000 files, of which roughly 290,000 were dependency folders, build caches, and nested backups. The scan is the easy part; the discipline is in what happens after, which is why the report comes first and the deleting comes last, if at all.

Step 1: Scan (read-only)

Take a target folder (default: current directory). Walk it and classify what you find into groups:

| Group | Patterns | |---|---| | Dependency folders | node_modules, venv, .venv, env, vendor, site-packages | | Build caches | __pycache__, .cache, dist, build, .next, .parcel-cache, *.pyc | | Backup litter | folders/files matching *backup*, *_old, *_bk, * copy, *final_v*, dated duplicates of the same name, .zip/.rar siblings of an existing folder | | Logs and temp | *.log, *.tmp, *.err.log, *.out.log, crash dumps | | OS litter | Thumbs.db, .DS_Store, desktop.ini | | Heavyweights | any single file over 100 MB that fits no group above |

Measure each group: file count and total size. Note the largest individual offenders.

Things that are NEVER junk, regardless of pattern: anything inside a .git directory that belongs to a live repo, .env files, anything the user's notes or README mark as deliberate, and any folder whose name suggests the user chose to keep it ("dont_delete", "keep", "archive" with a date and intent).

Step 2: Report

Produce cleanup-report.md in the target folder:

1. Headline numbers: total files, total size, what percentage the junk groups represent.
2. Per group: count, size, and the top 10 paths by size.
3. Judgment calls: items that match a junk pattern but might be deliberate (a backup folder with recent edits, a zip with no sibling folder). These get a "verify before touching" flag, not a recommendation.
4. Recommended actions, most space per least risk first. Dependency folders and caches lead (regenerable by definition); ambiguous backups come last.

Step 3: Clean (only with an approved list)

Never act on the whole report. The flow is:

1. The user names which groups or paths to clean.
2. Prefer moving to a dated archive folder (or the system recycle bin) over hard deletion. State where things went.
3. Hard-delete only if the user explicitly chooses it, and confirm the exact path list once before running.
4. After cleaning, re-report the headline numbers so the result is visible.

Rules

• Report-only by default. No file is moved or deleted in a scan.
• Never touch live .git repositories, .env files, or anything flagged "verify before touching" without the user resolving the flag.
• One folder tree per run. No "clean the whole drive" sweeps.
• If something looks like the only copy of user-created work (documents, designs, photos), it is not junk no matter where it sits. Flag it, never recommend deletion.
• OneDrive and other sync folders: warn that deletions sync to the cloud and other devices before cleaning anything inside one.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: jqaisystems
• Source: jqaisystems/jqai-ai-skills
• License: MIT
• Homepage: https://www.ai.joaoqueiros.com/skills/

Install and usage instructions live in the source repository linked above.