moonlight-lupin/agent-skills/file-organizer

File Organizer

An LLM-powered file organizer that reads file contents, proposes a sensible folder structure with renamed files, and executes moves in chunks — with a mandatory user confirmation gate before any filesystem changes.

Overview

Inspired by LlamaFS but designed for Hermes: the script handles filesystem I/O (scan, read snippets, execute moves). The agent's own LLM does the reasoning by default. An optional propose subcommand can call a cheaper external LLM (deepseek, openrouter, ollama) for cost-conscious runs on large directories. Uses stdlib-only HTTP (urllib) — no requests/openai SDK needed.

When to Use

• User says "organize my Downloads" / "tidy my Desktop" / "clean up this folder"
• User has a messy directory and wants files sorted into categories
• User wants files renamed based on their content (not just moved)
• User wants to organize files on a local disk, network mount, or remote filesystem

Don't use for:

• Bulk file deletion (this skill only moves and renames, never deletes)
• Deduplication (use a dedicated dedup tool)
• File content search (use search_files instead)

Workflow

Step 1 — Scan

Run the scan script to get file metadata + content snippets:

python3 scripts/organize.py scan --path <directory> [--depth N] [--max-snippet-chars 500]

Output is a JSON array. Each file entry contains:

• path: full file path
• relative_path: path relative to scanned directory
• name: current filename
• size: bytes
• mtime: ISO timestamp
• type: text, image, audio, pdf, office, archive, other
• mime: detected MIME type
• snippet: first N chars of content (text/pdf) or metadata (image/audio)

Flags: --depth N (max directory depth), --max-snippet-chars N (per-file snippet limit, default 500), --include-hidden (include dotfiles, skipped by default).

Step 2 — Propose Organization (Agent LLM — Default)

Using the scan output, reason about:

• Natural categories based on file content and types
• A clean folder hierarchy (not too deep — 2-3 levels max)
• Descriptive filenames that reflect content (not IMG_3847.jpg → Hawaii-Beach-Sunset.jpg)
• Time-based grouping when appropriate (e.g., 2024-Taxes/, 2023-Receipts/)

Build a plan JSON:

{
  "source_dir": "<scanned dir>",
  "moves": [
    {"source": "<full path>", "destination": "<full path>"}
  ],
  "folders_to_create": ["<full path>", ...]
}

Cost-saving offer: When presenting the plan (Step 3), mention to the user:

"This used the agent's own model. For larger directories, I can run the reasoning

through a cheaper LLM to save costs. Available options:"

| Provider | Model | Cost | Notes |

|----------|-------|------|-------|

| deepseek | deepseek-chat | ~$0.14/M in, $0.28/M out | Cheapest cloud |

| openrouter | deepseek/deepseek-chat | pay-as-you-go | Multi-model routing |

| ollama | llama3.2:3b | Free (local CPU) | Slow but zero cost |

Just say "use deepseek" or "use ollama" and I'll re-run via the cheaper model.

Only offer this if the scan had 20+ files (not worth it for tiny directories).

Using a cheaper model (opt-in): If the user requests a cheaper model, or if the directory is very large (100+ files), use the propose subcommand instead:

python3 scripts/organize.py propose --scan /tmp/scan.json --source-dir <dir> \
  --provider deepseek [--model <model-id>]

Provider priority: --provider flag > DEEPSEEK_API_KEY env > OPENAI_API_KEY env

OPENROUTER_API_KEY env > OLLAMA_HOST env > error

If propose fails (network, auth, parse), fall back to agent reasoning manually.

Step 3 — Present Plan to User (MANDATORY GATE)

This step is non-negotiable. Present the plan as a formatted table:

Ask the user to:

• Confirm the plan as-is
• Adjust specific moves (rename, re-categorize, skip)
• Cancel entirely

Never proceed to execution without explicit user confirmation.

Step 4 — Execute in Chunks

After user confirms, write the plan to a temp file and execute:

python3 scripts/organize.py execute --plan /tmp/plan.json --chunk-size 10

The script:

• Creates destination folders first
• Moves files in chunks of 10 (configurable)
• Prints progress after each chunk: [chunk 2/5] moved 10, failed 0
• Skips conflicts (destination exists) — never overwrites
• Logs per-file errors without aborting the batch

Step 5 — Report Results

Present the final summary to the user:

• Files moved successfully
• Files skipped (conflicts)
• Files failed (errors)
• Suggest next steps (e.g., "want me to organize another directory?")

Cross-Filesystem Support

The organizer works on any path accessible from the host:

Key Guardrails

1. Never delete — only move and rename. If the user wants deletions, that's a

separate task with its own confirmation flow.

1. Never overwrite — if a destination file exists, skip and log.
2. Always gate — present plan to user before executing. No silent moves.
3. Chunk execution — process in batches of 10 to manage failures gracefully.
4. Agent LLM by default, cheap LLM opt-in — scan and execute are pure

filesystem I/O. The agent's own LLM does the reasoning by default. The propose subcommand is an opt-in for cheaper external models on large directories.

Common Pitfalls

1. Huge directories — scanning 10,000+ files will produce massive output. Use

--depth 1 first to survey, then drill into subdirectories. Or scan by file type subsets.

1. Binary files with no snippet — images, audio, office docs get metadata only.

The LLM must infer categories from filename, size, and mtime.

1. Paths with spaces — always quote paths in the plan JSON. The script handles

this correctly via Python pathlib.

1. Network filesystem latency — organizing files on SSHFS/NFS/SMB mounts is

slower due to network round-trips. Use smaller chunk sizes (5) for remote targets.

1. Permission errors — some files may not be readable. The scan script handles

this gracefully (returns metadata with snippet: null).

Verification Checklist

• [ ] Scan completed and returned JSON output
• [ ] Plan includes source_dir, moves, and folders_to_create
• [ ] Plan presented to user as a table with clear before → after mapping
• [ ] User explicitly confirmed the plan
• [ ] Plan written to temp JSON file
• [ ] Execution completed with chunked progress output
• [ ] Final summary reported (moved / failed / skipped counts)
• [ ] No files were deleted or overwritten

One-Shot Recipes

Organize Downloads folder

# 1. Scan
python3 scripts/organize.py scan --path ~/Downloads --depth 2 > /tmp/scan.json
 
# 2. Agent reasons over scan JSON, builds plan (default — uses agent's own LLM)
#    OR for large directories, use cheap LLM:
#    python3 scripts/organize.py propose --scan /tmp/scan.json --source-dir ~/Downloads --provider deepseek > /tmp/plan.json
 
# 3. Agent presents plan to user (with cost-saving offer if 20+ files)
 
# 4. After confirmation:
python3 scripts/organize.py execute --plan /tmp/plan.json --chunk-size 10

Organize a network-mounted directory

# 1. Scan
python3 scripts/organize.py scan --path /mnt/remote/Desktop --depth 2 > /tmp/scan.json
 
# 2. Agent reasons over scan JSON, builds plan (default)
#    OR: python3 scripts/organize.py propose --scan /tmp/scan.json --source-dir /mnt/remote/Desktop --provider deepseek > /tmp/plan.json
 
# 3. Agent presents plan to user
 
# 4. After confirmation (smaller chunks for network latency):
python3 scripts/organize.py execute --plan /tmp/plan.json --chunk-size 5

Dry run (preview only)

python3 scripts/organize.py execute --plan /tmp/plan.json --dry-run

Cheap LLM (opt-in for large directories)

# Scan → Propose via deepseek → Execute
python3 scripts/organize.py scan --path ~/Downloads > /tmp/scan.json
python3 scripts/organize.py propose --scan /tmp/scan.json --source-dir ~/Downloads --provider deepseek > /tmp/plan.json
# Agent reviews plan, presents to user, then:
python3 scripts/organize.py execute --plan /tmp/plan.json --chunk-size 10

Run self-tests

python3 scripts/organize.py --self-test

Script Reference

Optional Python deps (degrade gracefully if missing):

• Pillow — image dimensions and EXIF dates
• mutagen — audio duration and ID3 tags
• PyMuPDF (fitz) — PDF first-page text extraction