latour-ai/skills/investment-folder-organizer

Investment Folder Organizer

Turn a chaotic pile of investment files into a clean, predictable structure: eight standardized subfolders plus a consistent, date-first filename convention, with a generated index that doubles as an audit trail. Same structure for every investor type — allocator, public-equity, PE, or VC — so the layout is learnable once and reused forever.

Non-destructive by design: originals are never moved or deleted. Files are copied into a new sibling folder. If the user explicitly asks to move/replace in place, confirm that intention before doing anything irreversible.

Workflow

1. Locate and scope the folder

Confirm the source folder path. If the user hasn't given one, ask. Then inventory everything, including nested subfolders — these get flattened into the new structure. List files with their type and size:

find "<source_folder>" -type f -not -name '.*' -printf '%P\t%s bytes\n' | sort

Note the likely investor context (fund / public co / private co / startup) from the filenames and the user's framing — it informs DocType choices but not the folder structure, which is always the same eight folders.

2. Read enough of each file to classify it

Classify on content first, filename second — investment files are notoriously mislabeled (Project Atlas v3 FINAL.pdf could be anything). You usually do not need the whole file; the first page / first slide / first sheet plus the filename is typically decisive. Read more only when torn.

• PDFs: read the first page or two (text or rasterize). For PDF reading strategy see

the pdf-reading skill if available.

• Office files (.docx/.pptx/.xlsx): use the relevant skill (docx, pptx, xlsx)

or extract text; the title slide, document heading, or first sheet usually suffices.

• For other types and a routing cheat-sheet, consult the file-reading skill.
• Sample large files — don't burn context reading a 300-page data dump end to end.

Read `references/taxonomy.md` for the folder definitions, the content-based classification tie-breakers (e.g. Marketing vs. Research, Financials vs. Legal), and the full DocType vocabulary. Consult it whenever a placement is non-obvious.

3. Build the plan

For every file, decide three things: which of the eight folders, the new filename, and a confidence level. Assemble them into a plan JSON (schema in the script header and below). Set confidence to medium/low and add a one-line note for anything you're unsure about; route genuinely unclassifiable files to 99_Unsorted. These get surfaced in the index for the user to review, which is far better than a confident misfile.

The eight folders (defined fully in references/taxonomy.md):

Naming: YYYY-MM-DD_DocType_Descriptor.ext — date first (use the content date, fall back to file mtime), PascalCase DocType, short hyphenated descriptor, no company/fund name (the parent folder carries it). Full rules and examples are in references/taxonomy.md.

4. Preview, then execute

Show the user a compact summary of the plan before copying — counts per folder and, critically, any files headed to 99_Unsorted or flagged low-confidence. Invite a quick correction. For a small set you can list every rename; for a large set, show the per-folder counts plus the flagged items and a handful of representative renames.

To preview mechanically without copying, run the script with --dry-run.

Write the plan to a JSON file, then run the executor. By default the output folder is a sibling named <SourceName> (Organized):

python scripts/organize.py --plan /path/to/plan.json
# add --dry-run first if you want a no-write preview

The script handles the deterministic, error-prone parts so you don't have to: canonicalizing folder names, sanitizing filenames, preserving extensions, resolving name collisions (-2, -3, …), copying with copy2, and writing 00_INDEX.md.

5. Present the result

Point the user to the new folder and 00_INDEX.md. Explicitly call out the Needs review section — the unsorted and low-confidence files — and offer to re-file any of them based on the user's steer. If they correct a placement, edit the plan and re-run (re-running into a fresh output folder is cheap and clean).

Plan JSON shape

{
  "source_root": "/abs/path/to/messy/folder",
  "output_root": "/abs/path/to/Target (Organized)",
  "subject": "Acme Capital — Flagship Fund",
  "investor_context": "hedge fund allocator",
  "items": [
    {
      "source": "subdir/Q3 letter FINAL.pdf",
      "folder": "06_Communications",
      "new_name": "2024-09-30_InvestorLetter_Q3-2024.pdf",
      "confidence": "high",
      "note": ""
    }
  ]
}

• source is relative to source_root (absolute paths also accepted).
• folder accepts the canonical name or a short alias (Financials, 02, legal…);

unknown values are routed to 99_Unsorted and flagged.

• confidence is high | medium | low (default high).
• note is optional but always shown in the index for unsorted/low/medium items.

Principles

• Never destructive. Copy. Leave the original folder untouched unless the user

explicitly and knowingly asks otherwise.

• Content over filenames. The whole point is to fix bad names, so don't trust them.
• Flag, don't guess. A file in 99_Unsorted with a note beats a confident misfile.
• One structure, every asset class. Don't fork the taxonomy per investor type;

adapt only the DocType words.