DOCX Documents

This skill covers Word-compatible .docx work in Canvas Notebook. It is a clean-room Canvas skill; do not copy vendor skill material, hidden prompts, or proprietary helper code into this folder.

Core Contract

• Always edit a checked-out working copy. Call checkout_docx with the workspace-relative input path, edit only its returned workingPath, and publish with commit_docx. This applies even when overwriting is requested: the commit tool checks the original revision and edit lease.
• For a new document, use checkout_docx with createOnly: true and the desired .docx path. Create the returned working file with Python, then commit its checkout ID. An existing output is never silently overwritten.
• Use CANVAS_AGENT_TEMP_DIR for scripts, working copies, unpacked OOXML, conversions, and render previews. Publish DOCX working copies through commit_docx; use copy_path or move_path for requested non-DOCX exports.
• Preserve existing document style for targeted edits. For new documents, set explicit page size, margins, paragraph spacing, font choices, headings, tables, headers, and footers instead of relying on Word defaults.
• Prefer editable Word-native content: paragraphs, runs, tables, lists, sections, headers, footers, comments, and fields. Do not turn pages into screenshots unless the user asks for an image-only deliverable.
• Keep intermediate render images, PDFs, unpacked XML, and scratch files out of the final response unless the user asks for them.
• Store final user-facing files in the requested workspace path using a stable descriptive filename.

Recommended Tools

• python-docx for normal document creation and structured edits.
• Standard library zipfile plus lxml for OOXML edits that python-docx cannot express, such as true comments, content controls, fields, or targeted XML repair.
• Pandoc for simple, text-oriented Markdown-to-DOCX or DOCX-to-Markdown conversion. Do not use it for format-faithful edits to an existing DOCX.
• LibreOffice soffice for DOCX-to-PDF conversion. Before invoking it, read references/headless-libreoffice.md.
• Poppler pdftoppm or pdftocairo when available for PDF to PNG rendering.

Use the installed runtime dependencies. Agent shell subprocesses can write only to the current session temporary directory; workspace files are read-only there. Do not try shell redirects, direct Python saves, copies, or custom scripts to bypass publication. Never store API keys or secrets in a skill directory.

Checkout, publication, and recovery

• Keep the checkoutId, workingPath, and lockExpiresAt returned by checkout. Use inspect_docx_checkout with renewLease: true before the lease expires during longer edits. Expired or replaced leases are never recreated by renewal.
• Finish and await all Python/render subprocesses before commit_docx reads the working copy. Pass only the checkout ID; never supply or replace its baseline.
• A committed result is idempotent. If publication returned a transient error with status prepared, retry the same checkout ID: its exact proposed bytes remain fixed even if scratch files changed afterward.
• If status is conflict, report the conflict and its recoveryId. The proposal is retained outside temporary cleanup. Use inspect_docx_checkout with restoreWorkingCopy: true to recover it in the same task; this replaces only the scratch copy. To keep a separate proposal, create a new createOnly checkout and copy the recovered content into that checkout's working path.
• Use release_docx_checkout when stopping without publication. It retains the current draft and releases the lease. Released checkouts cannot commit.
• If the lease belongs to another editor, explain who is editing when that information is available and preserve any proposed work. Never force release another editor or overwrite a newer version to complete a document task.

Workflow

1. Identify whether the task is read/review, create, targeted edit, structural repair, comments/redlines, or conversion.
2. Check out the input or reserve a new output. If editing an existing document, inspect the working copy's current structure before changing it: headings, sections, tables, images, comments, tracked changes, headers, footers, and metadata.
3. Make the smallest reliable edit that satisfies the request. Use styles and reusable helper functions instead of scattered one-off formatting.
4. For visual deliverables, render the DOCX when possible and inspect page images for clipping, overlap, missing glyphs, broken tables, bad page breaks, header/footer issues, and inconsistent spacing.
5. If render output is wrong, fix the DOCX and render again. If rendering is unavailable, perform structural checks and clearly state that visual QA could not be completed.
6. Publish the validated working copy with commit_docx and verify that the returned status is committed before reporting the workspace file as saved.

Common Tasks

• Read or summarize: extract text with python-docx; inspect tables and images separately; render if layout matters.
• Create a new document: build with python-docx, define styles up front, then add content. Use real tables and lists.
• Targeted edit: preserve existing formatting. Avoid document-wide restyles unless asked.
• Comments: use OOXML comments only when true Word comments are required. Otherwise, a review memo or inline markup may be safer.
• Tracked changes: true Word redlines require OOXML patching. If exact revision semantics matter, validate in Word-compatible tooling after writing.
• Google Docs target: create a clean .docx first. Avoid Word-only visual effects that import poorly, such as decorative title rules, complex floating objects, and brittle fields.

Final Checks

• Final DOCX opens without repair warnings.
• No placeholder tokens, internal notes, or unresolved TODOs remain.
• Tables fit within margins and have readable column widths.
• Headers, footers, page numbering, links, and references are intentional.
• For high-stakes documents, mention any environment limitation that prevented render or Word-native validation.