---

name: latex-formula-to-word description: End-to-end pipeline that turns a LaTeX Chinese thesis into a Word docx with native OMML formulas + clean Word layout. Scans TeX, generates placeholder PDF, waits for user to convert PDF→Word, replaces placeholders with OMML, then surgically cleans up PDF→Word sectPr artifacts (Word-specific blank-page bug). Use ONLY when user invokes /latex-formula-to-word. Manual trigger; never auto-fire. argument-hint: "[optional: project directory; defaults to cwd]" allowed-tools: ["Read", "Bash", "Glob", "Edit", "Write", "AskUserQuestion"]

Thesis Formula Replace (LaTeX → Word with Native OMML)

End-to-end pipeline for Chinese thesis projects that converts a LaTeX thesis into a Word docx where every formula is a native Word OMML object, while preserving PDF-consistent layout.

Designed-in defaults (do not skip without user override):

1. Pandoc compile failure → terminate (fail-fast)
2. Heading detection skips empty BodyText spacer paragraphs
3. Plan F applied: delete redundant type=continuous sectPrs (Word treats them as soft page breaks even though spec says they shouldn't)
4. LibreOffice pre-render verifies per-page fill rate at the end

Inputs the user must provide before/during the run

• Project directory containing TeX files (typically with body/, thesis.tex, csust*.cls)
• A PDF→Word conversion at one point (you tell the user when; they use Word/Acrobat/Mathpix/etc. themselves)

The six phases

Phase 1 — Scan TeX, generate placeholder copies

Run:

python3 ~/.claude/skills/latex-formula-to-word/scripts/scan_tex.py \
  --project <PROJECT_DIR> \
  --tex-entry <MAIN_TEX>      # e.g. thesis.tex

What it does:

• Walks all .tex files reachable from the entry point (resolves \input/\include)
• Extracts every formula: inline $...$ / \(...\), display $$...$$ / \[...\], environments equation, equation*, align, align*, gather, gather*, displaymath
• Assigns F001, F002, … sequentially in source order
• Writes formula_placeholders.json (list of {id, file, line, column, type, source})
• Writes modified TeX copies under <PROJECT_DIR>_formula_placeholders_only/ where each formula is replaced by [公式 Fxxx] (numbered display gets the equation number to its right; the numbering macros are preserved)

Skip rule: if the user already has formula_placeholders.json, ask whether to reuse or rebuild. Default: reuse.

Phase 2 — Compile placeholder TeX to PDF

Run XeLaTeX 4-pass on the placeholder copy:

cd <PROJECT_DIR>_formula_placeholders_only
xelatex -interaction=nonstopmode <MAIN_TEX>
biber <MAIN_TEX without ext>
xelatex -interaction=nonstopmode <MAIN_TEX>
xelatex -interaction=nonstopmode <MAIN_TEX>

If make is available with the right target, prefer make <target> from the original project after symlinking — see project's CLAUDE.md.

Phase 3 — Hand off to user for PDF→Word conversion

Tell the user exactly:

占位符 PDF 已生成：<path>.pdf

请用你偏好的 PDF→Word 工具（Word、Acrobat 等）把它转成 docx，不要用 Mathpix（会丢页眉版式）。完成后告诉我 docx 路径。

Wait for the user to come back with the docx path. Don't try to auto-convert.

Phase 4 — Replace placeholders with OMML in docx

Run:

python3 ~/.claude/skills/latex-formula-to-word/scripts/replace_docx.py \
  --docx <USER_DOCX> \
  --json <PROJECT>/formula_placeholders.json \
  --out <USER_DOCX_DIR>/<NAME>.replaced.docx

Internally this script:

1. Dry-run pandoc on all formulas. If any fail, stop immediately and report the failure list with file/line. DO NOT auto-fix or silently skip. User must fix the TeX macro or accept exclusion before re-running.
2. Extracts OMML for each formula into cache/. For display formulas, keeps only <m:oMathPara> and drops the standalone <m:oMath> (pandoc emits both due to its display-math wrapping; double-insertion causes the formula to render twice in Word).
3. Walks <w:body> and every <w:tc>, finds [公式 Fxxx] occurrences:
   • Numbered display (paragraph matches ^\s*\[公式 Fxxx\]\s*\(\d+[-–]\d+[a-z]?\)\s*$): replaces whole paragraph with a 1×3 borderless table, column widths 200 / 8200 / 1100 twips. Right cell carries the equation number with <w:noWrap/> and tight <w:tcMar> (50/50/0/0). Middle cell carries the OMML, paragraph spacing line=240, before=0, after=0.
   • Inside table cell: lock the parent table's layout to fixed and copy gridCol widths into each <w:tcW> BEFORE inserting OMML (otherwise widening cells will rebalance the table).
   • Inline: run surgery — split runs at placeholder boundaries, replace the matched range with OMML siblings, preserve surrounding rPr.

Run surgery basics for split runs (placeholders are typically split across multiple <w:r> by PDF→Word):

1. Concatenate run texts → build offset map (run_idx, offset_within_run) per char.
2. For span [start, end]: split run at `end` first, then at `start`.
3. Delete runs between the two splits, insert OMML node(s) at that position.

Output report under <USER_DOCX_DIR>/thesis_replace_report.md listing:

• Counts (inline / numbered / in_table)
• Any IDs missing from OMML cache
• Pandoc failures (should be zero if Phase 4 didn't abort)

Phase 5 — Clean Word layout artifacts (Plan F + Heading detection)

Run:

python3 ~/.claude/skills/latex-formula-to-word/scripts/fix_sectpr.py \
  --docx <USER_DOCX_DIR>/<NAME>.replaced.docx \
  --inplace

Internally:

1. Backup the docx to <NAME>.before-sectpr.docx.
2. Scan all embedded <w:sectPr> in the body. For each, find the next non-empty paragraph (skip empty BodyText spacer paragraphs).
3. Classify each sectPr:
   • Keep nextPage if either:
     • It already has <w:headerReference> or <w:footerReference> (real chapter break), OR
     • The next non-empty paragraph has pStyle in {Heading1, Heading2, Heading3, Title} plus the project's localized heading style names (auto-detected from word/styles.xml — see "Heading style detection" below)
   • Delete the sectPr element entirely (do NOT just retype to continuous) when ALL of:
     • Type is or would be continuous
     • <w:pgMar> exactly matches the body-level default
     • No header/footer reference
4. Leave 1–3 "unique pgMar" continuous sectPrs alone (cover page, abstract, ToC area — they intentionally have different margins).

Why delete instead of retype: Word treats type=continuous sectPrs as soft page breaks even though OOXML spec says it shouldn't. LibreOffice respects continuous; Word doesn't. Only removing the element entirely makes Word stop creating a section boundary.

Heading style detection (multi-template support)

Different Chinese thesis templates name heading styles differently: Heading1, 标题1, 章, 一级标题, etc. Detect dynamically:

• Read word/styles.xml
• A style is a "heading" if its <w:pPr> contains <w:outlineLvl val="0..3"/>
• Add those styleIds to the heading detection set

The default set hard-codes Heading1, Heading2, Heading3, Title for templates that don't set outlineLvl.

Phase 6 — LibreOffice render verification

Run:

python3 ~/.claude/skills/latex-formula-to-word/scripts/verify_render.py \
  --docx <USER_DOCX_DIR>/<NAME>.replaced.docx

Internally:

1. Calls soffice --headless --convert-to pdf (requires LibreOffice installed; on macOS at /Applications/LibreOffice.app/Contents/MacOS/soffice or via Homebrew at /opt/homebrew/bin/soffice).
2. Uses PyMuPDF to inspect each page's content extent vs page usable area.
3. Reports any page with content fill < 80% (excluding header/footer band).
4. Maps PDF page number → header label "第 X 页" for user-friendly reference.

If verification finds blank-prone pages:

• A page filled 60–80%: usually figure-driven (image too tall, pushed to next page). Cannot fix without resizing images. Report and move on.
• A page filled < 40%: probably a kept-nextPage sectPr that user wants to convert to continuous. Show the next-paragraph pStyle and text, ask user whether to demote.
• Note: Word and LibreOffice paginate differently. LibreOffice reports are a sanity check, not gospel — final judgment is in the user's Word.

Hard rules

1. Never overwrite the user's original docx or TeX. Every phase outputs to a copy with a suffix.
2. Every destructive change has a backup: before-replace, before-sectpr, etc.
3. No silent failures. If pandoc rejects a formula, stop. If a sectPr has unexpected pgMar, leave it and report.
4. Don't unify pgMar across sections. Each section's original pgMar is preserved. Only sectPrs with the body-default pgMar are eligible for Plan F deletion.

Decision points where you should ask the user

• Phase 1: if formula_placeholders.json already exists — reuse or rebuild?
• Phase 4: if pandoc fails on any formula — show the failed IDs and source, stop and let user fix TeX macros.
• Phase 5: if any Heading3 sub-section break looks borderline (e.g., a section heading mid-chapter) — show its body_idx + next text, ask whether to demote.
• Phase 6: if any page has fill < 40% — show context, ask whether to demote a sectPr or accept.

Known limitations

• PDF→Word can pack multiple consecutive equation blocks into a single multi-column table (e.g., a 2×3 grid holding 3 formulas together). The replace_docx.py script handles them as "in_table" inline replacements (with cell width locked), but if a formula is wider than its cell, it will wrap. There's no automatic fix in this pipeline — those cases need a separate "split the merged formula table" pass (not implemented here).
• Pre-existing OCR alignment errors in the PDF→Word output (e.g., a paragraph that got split mid-sentence with one half right-aligned) are visible after replacement but were already present. This skill does not fix them.
• Word renders differently from LibreOffice for some sectPr edge cases. The pipeline targets Word behavior; LibreOffice verification is approximate.

What gets produced

In the user's docx directory after a clean run:

• <NAME>.replaced.docx — final output with OMML + clean layout
• <NAME>.before-replace.docx — pre-replacement backup
• <NAME>.before-sectpr.docx — post-replacement, pre-Plan-F backup
• thesis_replace_report.md — replacement count + any caveats
• thesis_render_report.md — LibreOffice page-fill audit