OpenAI's SKILL.md location

[mrseeker]

OpenAI accidentally dumped its location for skills and this little gem:

bash -lc cat /home/oai/skills/docx/SKILL.md

---

DOCX Skill (Read • Create • Edit • Redline • Comment)

Use this skill when you need to create or modify .docx files in this container environment and verify them visually.

Non-negotiable: render → inspect PNGs → iterate

You do not “know” a DOCX is satisfactory until you’ve rendered it and visually inspected page images.
DOCX text extraction (or reading XML) will miss layout defects: clipping, overlap, missing glyphs, broken tables, spacing drift, and header/footer issues.

Shipping gate: before delivering any DOCX, you must:

• Run render_docx.py to produce page-<N>.png images (optionally also a PDF with --emit_pdf)
• Open the PNGs (100% zoom) and confirm every page is clean
• If anything looks off, fix the DOCX and re-render (repeat until flawless)

If rendering fails, fix rendering first (LibreOffice profile/HOME) rather than guessing.

Deliverable discipline: Rendered artifacts (PNGs and optional PDFs) are for internal QA only. Unless the user explicitly asks for intermediates, return only the requested final deliverable (e.g., when the task asks for a DOCX, deliver the DOCX — not page images or PDFs).

Design standards for document generation

For generating new documents or major rewrite/repackages, follow the design standards below unless the user explicitly requests otherwise. The user's instructions always take precedence; otherwise, adhere to these standards.

When creating the document design, do not compromise on the content and make factual/technical errors. Do not produce something that looks polished but not actually what the user requested.

It is very important that the document is professional and aesthetically pleasing. As such, you should follow this general workflow to make your final delivered document:

1. Before you make the DOCX, please first think about the high-level design of the DOCX:
   • Before creating the document, decide what kind of document it is (for example, a memo, report, SOP, workflow, form, proposal, or manual) and design accordingly. In general, you shall create documents which are professional, visually polished, and aesthetically pleasing. However, you should also calibrate the level of styling to the document's purpose: for formal, serious, or highly utilitarian documents, visual appeal should come mainly from strong typography, spacing, hierarchy, and overall polish rather than expressive styling. The goal is for the document's visual character to feel appropriate to its real-world use case, with readability and usability always taking priority.
   • You should make documents that feel visually natural. If a human looks at your document, they should find the design natural and smooth. This is very important; please think carefully about how to achieve this.
   • Think about how you would like the first page to be organized. How about subsequent pages? What about the placement of the title? What does the heading ladder look like? Should there be a clear hierarchy? etc
   • Would you like to include visual components, such as tables, callouts, checklists, images, etc? If yes, then plan out the design for each component.
   • Think about the general spacing and layout. What will be the default body spacing? What page budget is allocated between packaging and substance? How will page breaks behave around tables and figures, since we must make sure to avoid large blank gaps, keep captions and their visuals together when possible, and keep content from becoming too wide by maintaining generous side margins so the page feels balanced and natural.
   • Think about font, type scale, consistent accent treatment, etc. Try to avoid forcing large chunks of small text into narrow areas. When space is tight, adjust font size, line breaks, alignment, or layout instead of cramming in more text.
2. Once you have a working DOCX, continue iterating until the entire document is polished and correct. After every change or edit, render the DOCX and review it carefully to evaluate the result. The plan from (1) should guide you, but it is only a flexible draft; you should update your decisions as needed throughout the revision process. Important: each time you render and reflect, you should check for both:
   1. Design aesthetics: the document should be aesthetically pleasing and easy to skim. Ask yourself: if a human were to look at my document, would they find it aesthetically nice? It should feel natural, smooth, and visually cohesive.
   2. Formatting issues that need to be fixed: e.g. text overlap, overflow, cramped spacing between adjacent elements, awkward spacing in tables/charts, awkward page breaks, etc. This is super important. Do not stop revising until all formatting issues are fixed.

While making and revising the DOCX, please adhere to and check against these quality reminders, to ensure the deliverable is visually high quality:

• Document density: Try to avoid having verbose dense walls of text, unless it's necessary. Avoid long runs of consecutive plain paragraphs or too many words before visual anchors. For some tasks this may be necessary (i.e. verbose legal documents); in those cases ignore this suggestion.
• Font: Use professional, easy-to-read font choices with appropriate size that is not too small. Usage of bold, underlines, and italics should be professional.
• Color: Use color intentionally for titles, headings, subheadings, and selective emphasis so important information stands out in a visually appealing way. The palette and intensity should fit the document's purpose, with more restrained use where a formal or serious tone is needed.
• Visuals: Consider using tables, diagrams, and other visual components when they improve comprehension, navigation, or usability.
• Tables: Please invest significant effort to make sure your tables are well-made and aesthetically/visually good. Below are some suggestions, as well as some hard constraints that you must relentlessly check to make sure your table satisfies them.
  • Suggestions:
    • Set deliberate table/cell widths and heights instead of defaulting to full page width.
    • Choose column widths intentionally rather than giving every column equal width by default. Very short fields (for example: item number, checkbox, score, result, year, date, or status) should usually be kept compact, while wider columns should be reserved for longer content.
    • Avoid overly wide tables, and leave generous side margins so the layout feels natural.
    • Keep all text vertically centered and make deliberate horizontal alignment choices.
    • Ensure cell height avoids a crowded look. Leave clear vertical spacing between a table and its caption or following text.
  • Hard constraints:
    • To prevent clipping/overflow:
      • Never use fixed row heights that can truncate text; allow rows to expand with wrapped content.
      • Ensure cell padding and line spacing are sufficient so descenders/ascenders don't get clipped.
      • If content is tight, prefer (in order): wrap text -> adjust column widths -> reduce font slightly -> abbreviate headers/use two-line headers.
    • Padding / breathing room: Ensure text doesn't sit against cell borders or look "pinned" to the upper-left. Favor generous internal padding on all sides, and keep it consistent across the table.
    • Vertical alignment: In general, you should center your text vertically. Make sure that the content uses the available cell space naturally rather than clustering at the top.
    • Horizontal alignment: Do not default all body cells to top-left alignment. Choose horizontal alignment intentionally by column type: centered alignment often works best for short values, status fields, dates, numbers, and check indicators; left alignment is usually better for narrative or multi-line text.
    • Line height inside cells: Use line spacing that avoids a cramped feel and prevents ascenders/descenders from looking clipped. If a cell feels tight, adjust wrapping/width/padding before shrinking type.
    • Width + wrapping sanity check: Avoid default equal-width columns when the content in each column clearly has different sizes. Avoid lines that run so close to the right edge that the cell feels overfull. If this happens, prefer wrapping or column-width adjustments before reducing font size.
    • Spacing around tables: Keep clear separation between tables and surrounding text (especially the paragraph immediately above/below) so the layout doesn't feel stuck together. Captions and tables should stay visually paired, with deliberate spacing.
    • Quick visual QA pass: Look for text that appears "boundary-hugging", specifically content pressed against the top or left edge of a cell or sitting too close beneath a table. Also watch for overly narrow descriptive columns and short-value columns whose contents feel awkwardly pinned. Correct these issues through padding, alignment, wrapping, or small column-width adjustments.
• Forms / questionnaires: Design these as a usable form, not a spreadsheet.
  • Prioritize clear response options, obvious and well-sized check targets, readable scale labels, generous row height, clear section hierarchy, light visual structure. Please size fields and columns based on the content they hold rather than by equal-width table cells.
  • Use spacing, alignment, and subtle header/section styling to organize the page. Avoid dense full-grid borders, cramped layouts, and ambiguous numeric-only response areas.
• Coherence vs. fragmentation: In general, try to keep things to be one coherent representation rather than fragmented, if possible.
  • For example, don't split one logical dataset across multiple independent tables unless there's a clear, labeled reason.
  • For example, if a table must span across pages, continue to the next page with a repeated header and consistent column order
• Background shapes/colors: Where helpful, consider section bands, note boxes, control grids, or other visual container[... ELLIPSIZATION ...]materialize SEQ/REF field display text for deterministic headless rendering/QA.

High-leverage utilities (also importable, but commonly invoked as CLIs):

• render_docx.py — canonical DOCX → PNG renderer (optional PDF via --emit_pdf; do not deliver intermediates unless asked).
• scripts/render_and_diff.py — render + per-page image diff between two DOCXs.
• scripts/content_controls.py — list / wrap / fill Word content controls (SDTs) for forms/templates.
• scripts/captions_and_crossrefs.py — insert Caption paragraphs for tables/figures + optional bookmarks around caption numbers.
• scripts/insert_ref_fields.py — replace [[REF:bookmark]] markers with real REF fields (cross-references).
• scripts/internal_nav.py — add internal navigation links (static TOC + Top/Bottom + figN/tblN jump links).
• scripts/style_lint.py — report common formatting/style inconsistencies.
• scripts/style_normalize.py — conservative cleanup (clear run-level overrides; optional paragraph overrides).
• scripts/redact_docx.py — layout-preserving redaction/anonymization.
• scripts/privacy_scrub.py — remove personal metadata + rsid* attributes.
• scripts/set_protection.py — restrict editing (read-only / comments / forms).
• scripts/comments_extract.py — extract comments to JSON (text, author/date, resolved flag, anchored snippets).
• scripts/comments_strip.py — remove all comments (final-delivery mode).

Audits / conversions / niche helpers:

• scripts/fields_report.py, scripts/heading_audit.py, scripts/section_audit.py, scripts/images_audit.py, scripts/footnotes_report.py, scripts/watermark_audit_remove.py
• scripts/xlsx_to_docx_table.py, scripts/docx_table_to_csv.py
• scripts/insert_toc.py, scripts/insert_note.py, scripts/apply_template_styles.py, scripts/accept_tracked_changes.py, scripts/make_fixtures.py

v7 additions (stress-test helpers):

• scripts/watermark_add.py — add a detectable VML watermark object into an existing header.
• scripts/comments_add.py — add multiple comments (by paragraph substring match) and wire up comments.xml plumbing if needed.
• scripts/comments_apply_patch.py — append/replace comment text and mark/clear resolved state (w:done=1).
• scripts/add_tracked_replacements.py — generate tracked-change replacements (<w:del> + <w:ins>) in-place.
• scripts/a11y_audit.py — audit a11y issues; can also apply simple fixes via --fix_table_headers / --fix_image_alt.
• scripts/flatten_ref_fields.py — replace REF/PAGEREF field blocks with their cached visible text for deterministic rendering.

scripts/xlsx_to_docx_table.py also marks header rows as repeating headers (w:tblHeader) to improve a11y and multi-page tables.

Examples:

• examples/end_to_end_smoke_test.md

Note: manifest.txt is machine-readable and is used by download tooling. It must contain only relative file paths (one per line).

Coverage map (scripts ↔ task guides)

This is a quick index so you can jump from a helper script to the right task guide.

Layout & style

• style_lint.py, style_normalize.py → tasks/style_lint_normalize.md
• apply_template_styles.py → tasks/templates_style_packs.md
• section_audit.py → tasks/sections_layout.md
• heading_audit.py → tasks/headings_numbering.md

Figures / images

• images_audit.py, a11y_audit.py → tasks/images_figures.md, tasks/accessibility_a11y.md
• captions_and_crossrefs.py → tasks/captions_crossrefs.md

Tables / spreadsheets

• xlsx_to_docx_table.py → tasks/tables_spreadsheets.md
• docx_table_to_csv.py → tasks/tables_spreadsheets.md

Fields & references

• fields_report.py, fields_materialize.py → tasks/fields_update.md
• insert_ref_fields.py, flatten_ref_fields.py → tasks/fields_update.md, tasks/captions_crossrefs.md
• insert_toc.py → tasks/toc_workflow.md

Review lifecycle (comments / tracked changes)

• add_tracked_replacements.py, accept_tracked_changes.py → tasks/clean_tracked_changes.md
• comments_add.py, comments_extract.py, comments_apply_patch.py, comments_strip.py → tasks/comments_manage.md

Privacy / publishing

• privacy_scrub.py → tasks/privacy_scrub_metadata.md
• redact_docx.py → tasks/redaction_anonymization.md
• watermark_add.py, watermark_audit_remove.py → tasks/watermarks_background.md

Navigation & multi-doc assembly

• internal_nav.py → tasks/navigation_internal_links.md
• merge_docx_append.py → tasks/multi_doc_merge.md

Forms & protection

• content_controls.py → tasks/forms_content_controls.md
• set_protection.py → tasks/protection_restrict_editing.md

QA / regression

• render_and_diff.py, render_docx.py → tasks/compare_diff.md, tasks/verify_render.md
• make_fixtures.py → tasks/fixtures_edge_cases.md
• docx_ooxml_patch.py → used across guides for targeted patches

Skill folder contents

• tasks/ — task playbooks (what to do step-by-step)
• ooxml/ — advanced OOXML patches (tracked changes, comments, hyperlinks, fields)
• scripts/ — reusable helper scripts
• examples/ — small runnable examples

Default workflow (80/20)

Rule of thumb: every meaningful edit batch must end with a render + PNG review. No exceptions.
"80/20" here means: follow the simplest workflow that covers most DOCX tasks reliably.

Golden path (don’t mix-and-match unless debugging):

1. Author/edit with python-docx (paragraphs, runs, styles, tables, headers/footers).
2. Render → inspect PNGs immediately (DOCX → PNGs). Treat this as your feedback loop.
3. Fix and repeat until the PNGs are visually perfect.
4. Only if needed: use OOXML patching for tracked changes, comments, hyperlinks, or fields.
5. Re-render and inspect again after any OOXML patch or layout-sensitive change.
6. Deliver only after the latest PNG review passes (all pages, 100% zoom).

Visual review (recommended)

Use the packaged renderer (dedicated LibreOffice profile + writable HOME):

python render_docx.py /mnt/data/input.docx --output_dir /mnt/data/out
# If debugging LibreOffice:
python render_docx.py /mnt/data/input.docx --output_dir /mnt/data/out --verbose
# Optional: also write <input_stem>.pdf to --output_dir (for debugging/archival):
python render_docx.py /mnt/data/input.docx --output_dir /mnt/data/out --emit_pdf

Then inspect the generated page-<N>.png files.

Success criteria (render + visual QA):

• PNGs exist for each page
• Page count matches expectations
• Inspect every page at 100% zoom (no “spot check” for final delivery)
• No clipping/overlap, no broken tables, no missing glyphs, no header/footer misplacement

Note: LibreOffice sometimes prints scary-looking stderr (e.g., error : Unknown IO error) even when output is correct. Treat the render as successful if the PNGs exist and look right (and if you used --emit_pdf, the PDF exists and is non-empty).

What rendering does and doesn’t validate

• Great for: layout correctness, fonts, spacing, tables, headers/footers, and whether tracked changes visually appear.
• Not reliable for: comments (often not rendered in headless PDF export). For comments, also do structural checks (comments.xml + anchors + rels + content-types).

Quality reminders

• Don’t ship visible defects (clipped/overlapping text, broken tables, unreadable glyphs).
• Don’t leak tool citation tokens into the DOCX (convert them to normal human citations).
• Prefer ASCII punctuation (avoid exotic Unicode hyphens/dashes that render inconsistently).

Where to go next

• If the task is reading/reviewing: tasks/read_review.md
• If the task is creating/editing: tasks/create_edit.md
• If you need an accessibility audit (alt text, headings, tables, links): tasks/accessibility_a11y.md
• If you need to extract or remove comments: tasks/comments_manage.md
• If you need to restrict editing / make read-only: tasks/protection_restrict_editing.md
• If you need to scrub personal metadata (author/rsid/custom props): tasks/privacy_scrub_metadata.md
• If you need to merge/append DOCXs: tasks/multi_doc_merge.md
• If you need format consistency / style cleanup: tasks/style_lint_normalize.md
• If you need forms / content controls (SDTs): tasks/forms_content_controls.md
• If you need captions + cross-references: tasks/captions_crossrefs.md
• If you need redaction/anonymization: tasks/redaction_anonymization.md
• If the task is verification/raster review: tasks/verify_render.md
• If your render looks wrong but content is right (stale fields): tasks/fields_update.md
• If you need a Table of Contents: tasks/toc_workflow.md
• If you need internal navigation links (static TOC + Back-to-TOC + Top/Bottom): tasks/navigation_internal_links.md
• If headings/numbering/TOC levels are messy: tasks/headings_numbering.md
• If you have mixed portrait/landscape or margin weirdness: tasks/sections_layout.md
• If images shift or overlap across renderers: tasks/images_figures.md
• If you need spreadsheet ↔ table round-tripping: tasks/tables_spreadsheets.md
• If you need tracked changes (redlines): ooxml/tracked_changes.md
• If you need comments: ooxml/comments.md
• If you need hyperlinks/fields/page numbers/headers: ooxml/hyperlinks_and_fields.md
• If LibreOffice headless is failing: troubleshooting/libreoffice_headless.md
• If you need a clean copy with tracked changes accepted: tasks/clean_tracked_changes.md
• If you need to diff two DOCXs (render + per-page diff): tasks/compare_diff.md
• If you need templates / style packs (DOTX): tasks/templates_style_packs.md
• If you need watermark audit/removal: tasks/watermarks_background.md
• If you need true footnotes/endnotes: tasks/footnotes_endnotes.md
• If you want reproducible fixtures for edge cases: tasks/fixtures_edge_cases.md