sboghossian Import: DOCX Processing (Anthropic)
What it does
This import adapter migrates a DOCX document-processing skill originally built for Anthropic's Claude API into the mini-claude-for-legal standard format. Legal teams working with Microsoft Word documents — contracts, briefs, pleadings, board resolutions, regulatory submissions — rely on reliable DOCX ingestion as a prerequisite for all downstream review and analysis skills.
The adapter normalises how Claude reads .docx files: extracting body text, table content, header/footer metadata, comments, revision marks (tracked changes), and document properties, then feeding them as structured context into whichever downstream skill (contract review, NDA triage, risk assessment) was originally wired to the Anthropic API.
Import config
| Field | Source mapping | Default if absent |
|---|---|---|
extraction_mode |
Legacy mode or extraction field |
full_text |
tables |
Legacy include_tables boolean |
true |
tracked_changes |
Legacy track_changes or revision_marks |
accept_all |
metadata_fields |
Legacy doc_properties array |
[author, created, modified, title] |
language |
Legacy lang |
auto-detect |
output_format |
Legacy format |
markdown |
chunk_size |
Legacy chunk_tokens |
2000 tokens per chunk |
overlap |
Legacy overlap_tokens |
200 tokens |
Dry-run preview
IMPORT PREVIEW — docx-processing-anthropic
Source shape : Anthropic DOCX extraction config
Extraction : full_text + tables
Track changes: accept_all (hidden; use 'show' to surface redlines)
Metadata : author, created, modified, title
Output : markdown with table serialisation
Chunk size : 2000 tokens / 200 overlap
Confirm before the adapted skill is written.
Extraction pipeline (post-import)
Once imported, the skill processes a DOCX file in these steps:
- Parse document XML — unzip the
.docxcontainer; readword/document.xml,word/styles.xml, and relationship files. - Body text extraction — walk paragraph elements, preserving heading hierarchy (H1–H6 →
#–######). - Table serialisation — convert Word tables to markdown pipe tables; flag merged cells.
- Tracked changes — depending on
tracked_changessetting:accept_all(clean),reject_all(original), orshow(inline[+added]/[-deleted]markers). - Comments extraction — append as footnote block with author + timestamp.
- Metadata header — prepend a YAML block with document properties.
- Chunking — split by token count with overlap for large documents before passing to downstream skill.
Legal document considerations
- Arabic / bilingual contracts: DOCX files from MENA jurisdictions often contain right-to-left text (Arabic) alongside left-to-right (English/French). Set
rtl: trueto preserve column order in tables and prevent paragraph reordering. - Certified translations: some UAE Ministry of Justice and Lebanese Notary submissions require the certified translator's stamp metadata to be preserved. Map
doc_properties.customfields to retain this. - Redlines in M&A: tracked changes in DOCX are legally significant in negotiated drafts. Default is
accept_all(clean final read); switch toshowwhen the user sends a redlined counterparty markup. - Password-protected files: prompt the user for the document password before extraction; never log it.
- Macro-enabled files (.docm): strip VBA before processing; flag to the user.
Failure modes
| Error | Likely cause | Resolution |
|---|---|---|
corrupt_docx |
File not a valid ZIP/XML structure | Ask user to re-save as .docx from Word |
encoding_error |
Arabic/Persian characters garbled | Force encoding: utf-8 and re-extract |
table_parse_fail |
Nested or merged cell tables | Fall back to raw cell dump |
rtl_paragraph_reversal |
LTR parser reorders RTL paragraphs | Set rtl: true |
password_protected |
File encrypted | Prompt for password |
Related skills
- [[import-pdf-processing-anthropic]]
- [[import-pptx-processing-anthropic]]
- [[import-contract-review-anthropic]]
- [[multimodal-document-ingestion]]
- [[review-contract-generic]]