Office Document Review

SkillFiles & storage

office-document-review — Review and proofread office documents in OneDrive or local file trees, including batch folder discovery, document text extraction, and spelling/grammar cleanup workflows.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Office Document Review skill

What this skill tells your AI

The instructions your AI receives, as published by atlasomnia/hermes-custom-pack in skills/office-document-review/SKILL.md and read by ahel’s review.

Use this skill when the user wants you to inspect a OneDrive/SharePoint-style folder tree, locate a subfolder by name/date, and proofread all documents in it for spelling and grammar.

This is a workflow skill, not a single-app skill: it covers local file-system browsing, document text extraction, and batch proofreading notes.

Triggers

  • "go to the OneDrive folder"
  • "look in folder X and then folder Y"
  • "proofread all those documents"
  • "check for spelling/grammar errors in the docs"
  • "review the files in this folder"

Core workflow

  1. Resolve the real path first
  • Prefer a filesystem path over guessing from the user’s wording.
  • Folder names in casual speech may not match the actual on-disk name exactly.
  • If the exact folder cannot be found, list the visible folder names and state what is missing before asking for clarification.
  1. Search the tree recursively
  • Use the live file tree to locate the target folder and confirm the actual path.
  • When the user gives a date-like folder name, search for variant spellings and punctuation (6-2-26, 6.2.26, 6_2_26, 6-2-2026) before assuming it does not exist.
  1. Inventory the documents
  • Identify all files in the target folder.
  • Group by extension and prioritize editable office docs (.docx, .doc, .odt, .rtf, .txt) before PDFs or spreadsheets.
  1. Extract readable text
  • For Word/Office files, extract the text in a way that preserves headings and paragraphs as much as possible.
  • If the file-reading tool rejects .docx as binary, do not classify the document as unreadable and do not repeat the same call across the batch. A DOCX is a ZIP package: validate it with Python zipfile, then use python-docx when available or extract word/document.xml and convert paragraph/table XML to bounded text. Record package validity, paragraph/table counts, and extraction method. Use native Word or PDF rendering separately when layout—not just text—is under review.
  • When validating generated fixtures, compare extracted text against the fixture specification or authoritative .txt source before opening Word. A valid ZIP alone does not prove the fixture contains the intended content.
  • For scanned or image-based files, use OCR or an extraction tool first.
  • If a file is not meaningfully text-reviewable after at least one appropriate format-specific fallback, say so explicitly instead of pretending to proofread it.
  1. Proofread for two passes
  • Pass 1: spelling, grammar, punctuation, capitalization.
  • Pass 2: obvious clarity issues, duplicated words, inconsistent dates/names, and broken formatting that would change meaning.
  • Keep business wording intact unless the user asked for rewriting.
  1. Report findings file-by-file
  • Summarize each document with:
  • file name
  • issue count
  • short list of fixes or concerns
  • If no issues are found, say that clearly.
  1. Do not silently modify files unless asked
  • The default action is review/report.
  • If the user wants edits, confirm the target files and whether they want tracked wording changes or a clean rewrite.

Pitfalls

  • Casual folder names may be approximate; always verify the actual folder names seen in the tree.
  • A folder search that returns the parent tree but not the exact folder is not a success; say what was found and what was not.
  • If the user says they downloaded the files, re-inventory the local download folder first; do not assume the original folder contents or the original extension set survived the move.
  • Do not claim “everything is fine” for a batch until each file was actually inspected.
  • If a document is large, inspect enough of it to make a real proofreading judgment rather than sampling one paragraph and extrapolating.
  • Keep the user informed if the folder path is ambiguous instead of asking generic questions.

Verification

Before closing the task, confirm:

  • the folder path was resolved,
  • every target document was identified,
  • each document was reviewed or explicitly excluded with a reason,
  • and the final report separates true findings from non-reviewable files.

Session note file

Practical notes

  • Always verify the exact business OneDrive root before drilling into a dated subfolder.
  • If the user says "OneDrive" in a work/document-review context and does not qualify which account, default to the work tenant first (OneDrive - <work tenant>) rather than the personal OneDrive. Keep the two trees mentally separate and say which root you are using when ambiguity could matter.
  • If folder recursion on a synced OneDrive tree is unstable, list the parent first and narrow the search in smaller steps instead of forcing a full tree walk.
  • When direct recursion throws Resource deadlock avoided, try a shallow parent find/inventory pass and then inspect the exact subfolder file-by-file.
  • If the user says they downloaded the files, re-inventory the local download folder first; do not assume the original folder contents or extensions survived the move.
  • Keep the final review file-by-file so the user can see what was actually checked.

Session reference

  • When the user asks for "edits needed," separate true spelling/grammar fixes from OCR or extraction artifacts and label each file accordingly.

Signals

GitHub stars
56
Forks
5
Last commit
Aug 2026
Advanced
Catalog kind
skill
Gateway key
office-document-review
Source
github.com/atlasomnia/hermes-custom-pack