file-organizer

Sorts a disorganized folder into a target tree, by writing a move plan you read and approve before any file is touched. Trigger on "organize this folder", "sort these files", "clean up this directory"

SKILL.md

GitHub

File Organizer

Nothing moves until you have read the plan. The plan is a CSV, one row per file, and the run executes that file and nothing else.

Scope and limits — read first

What this does.

  • Moves or copies files from a source folder into a target tree you define.
  • Matches on filename glob, filename regex, extension, size threshold, and modified-date range, combined with and/or inside a slot.
  • Optional rename per slot, from a regex capture into a format string.
  • Finds duplicates by SHA-256 across the source set and routes them to a review folder, keeping one copy in the slot.
  • Writes move-plan.csv before touching anything, and executes only that file.
  • Appends audit.log per run: timestamp, action, source, destination, hash, result.

What this does not do.

  • It does not delete anything. Duplicates and collisions go to review folders. What to discard is your call, made afterwards, by hand.
  • It does not read file contents to classify. Filename, extension, size, and date only. For content-based sorting, extract metadata first with doc-reader or drawing-reader and feed the result in as rename pairs.
  • It does not invent target folders. It will draft a taxonomy from what it finds and show you, but a slot you did not approve does not get created.
  • It does not follow symlinks more than once, cross filesystem boundaries, touch anything outside the declared target root, or work on a network path unless you say so explicitly.
  • It does not execute a plan you have not read.

What you need before starting. The source folder, where the result should go, and either a taxonomy or ten minutes to review the one it drafts.

Workflow

Step 1. Set up the run folder

Every run gets its own folder so plans and audit logs never overwrite each other:

docs/file-organizer/<YYYY-MM-DD>-<run-id>/
  taxonomy.yml
  move-plan.csv
  audit.log

Create it before anything else. If the project has no docs/ folder, confirm before creating one.

Step 2. Identify source and target

Ask for the source root and the target root, as absolute paths. Confirm both exist. Confirm they are on the same filesystem, and say so if not, because a cross-device move is a copy followed by a delete and the rollback story changes.

List the source and report what is there: file count, total size, extensions present, date range. Show this before proposing any structure.

Step 3. Get the taxonomy

Three routes, in order of preference.

  1. The user has one. Read it, and show back the slots and rules as you understood them.
  2. The user names an existing folder to imitate. Read that tree and derive slots from it.
  3. Draft one. Group by what is actually in the source, propose slots with the rule for each, and show the draft. Do not create a slot you cannot write a rule for, and do not create a slot for a file type the source does not contain.

Write the agreed taxonomy to taxonomy.yml in the run folder. Its shape:

target_root: /abs/path
allow_network: false
slots:
  - path: 01-drawings/current
    match:
      any:
        - ext: [".dwg", ".pdf"]
          name_glob: "A-*"
    rename: null
  - path: 02-specs
    match:
      all:
        - ext: [".docx", ".pdf"]
          name_regex: "(?i)spec"

Step 4. Build the move plan

Walk the source. For each file, find the first slot whose rule matches, and write a row:

source destination slot action sha256 note
  • Files matching no slot get action: unmatched and no destination. They stay where they are.
  • Files whose hash already appeared get action: duplicate and a destination under _duplicates/<slot>/.
  • Two different files landing on the same destination name get action: collision and a destination under _collisions/<slot>/.

Write it to move-plan.csv. Touch nothing.

Step 5. Show the summary and get approval

Report, before asking:

  • Rows by action: move, copy, duplicate, collision, unmatched.
  • The slots that got nothing, which usually means a rule is wrong.
  • The unmatched list if it is short, or its extensions and count if it is long.
  • The largest ten files being moved.

Then ask plainly whether to execute, and say that nothing has been touched yet. If the answer is anything other than yes, stop and leave the plan on disk.

Step 6. Execute

Read move-plan.csv row by row. For each row with action: move or copy:

  1. Create the destination parent if needed.
  2. Perform the action.
  3. Verify the destination exists and the hash matches.
  4. Append a line to audit.log.

On any failure, stop. Do not continue past an error, and do not roll back automatically. Report the row number reached; the audit log records everything completed up to that point.

Execute only rows in the CSV. If the source changed since the plan was written, say so and offer to rebuild the plan rather than working around it.

Step 7. Roll back, if asked

Read audit.log in reverse and move each file back to its recorded source path. Verify each hash on arrival. Append the reversal to the same log rather than deleting lines, so the folder's history stays readable.

Step 8. Report

Counts by action, where the plan and the audit log are, the unmatched list, and what is sitting in _duplicates/ and _collisions/ waiting on a decision.

Rules

  • Nothing moves before the plan is approved. The plan is a file the user can read, not a summary in chat.
  • Execute the CSV, not the intention. If a row looks wrong, stop and ask; do not correct it in flight.
  • Never delete. Route to a review folder and let the user decide.
  • Verify by hash after every move, and stop on the first mismatch.
  • An unmatched file stays where it is. Never invent a slot at execution time to catch a straggler.
  • Confirm before creating docs/ in a project that does not have one.

Anti-patterns

  • Executing a plan the user has not seen because the summary looked clean.
  • Deleting a duplicate because it is obviously a duplicate.
  • Adding a catch-all slot so the unmatched count reads zero.
  • Continuing past a failed move to finish the run.
  • Rolling back automatically on error, which turns one problem into two.
  • Classifying by opening the file. This skill matches on names and metadata; if that is not enough, the answer is a different tool upstream, not guessing.

Resources

None. This skill is one file. The run folder, the taxonomy, the plan, and the audit log are written directly at the paths above.