LaTeX ↔ Word

This skill moves academic manuscripts between LaTeX and Microsoft Word and assembles .tex from heterogeneous sources. It consolidates four former skills into three workflows: a footnote-preserving round-trip for co-author editing cycles, a high-fidelity one-way .tex → .docx engine for delivering finished papers to Word-only journals or coauthors, and a set of knowledge patterns for building a single .tex from PDF/docx/LLM-generated content. Read the deep doc for the workflow you need before running anything.

Routing

User intent	Workflow	Entry point	Deep doc
Deliver a finished LaTeX paper as high-fidelity Word (regression tables, math, cross-references must survive)	B — one-way tex→docx	scripts/convert.py	references/tex-to-docx-engine.md
Iterate on a manuscript with Word-based co-authors while editing in LaTeX (docx→tex→docx, footnotes preserved)	A — round-trip	gui.py, or scripts/docx_to_tex.py + scripts/tex_to_docx.py	references/roundtrip.md
Build a .tex from PDF / .docx / LLM-generated text	C — assemble from mixed sources	knowledge patterns (no scripts)	references/mixed-sources.md

Which workflow

• B is the default for "deliver my LaTeX paper as Word." The fidelity engine builds native Word tables, OMML equations, real footnotes, embedded figures, and resolves \cref/\Cref/\eqref from the .aux. Use B whenever the paper has regression/booktabs tables, math, or cross-references that must survive — plain pandoc drops or mangles all of these.
• A is for iterating with Word-based co-authors while you edit in LaTeX. Its tex → docx step is plain pandoc plus a style-templated scripts/reference.docx (fast, formatting-only), so it does not build native tables or resolve crefs. Choose A when the exchange is prose and footnotes and speed matters; switch to B once the document depends on tables/math/cross-references.
• C is knowledge-only — patterns for a pipeline that emits .tex, applied when you author the pipeline. No script to invoke.

Quick start

Workflow B (one-way, high fidelity) — compile the source once so a current .aux exists, then:

python scripts/convert.py path/to/main.tex --out main.docx
# optional: --aux main.aux --tables-dir tables --figures-dir figures \
#           --font "Linux Libertine G" --no-render

Workflow A (round-trip) — either launch the GUI:

python gui.py

or drive the two version-aware helpers (they wrap pandoc + reference.docx and run the footnote-count sanity check), using the input/ intermediate/ output/ layout:

python scripts/docx_to_tex.py v1     # input/input_v1.docx  -> intermediate/intermediate_v1.tex
# edit intermediate/intermediate_v2.tex ...
python scripts/tex_to_docx.py v2     # intermediate/intermediate_v2.tex -> output/output_v2.docx

Workflow C — no command; read references/mixed-sources.md and apply the two-track escaping / code-fence-stripping / \DeclareUnicodeCharacter patterns in your own generator.

Tool requirements

• pandoc (3.x for the engine; ≥ 2.11 for the round-trip).
• xelatex / latexmk (TeX Live or MiKTeX) — required to produce the .aux the engine reads, and for TikZ/PGFPlots/algorithm rendering.
• python-docx, lxml — native table and OOXML construction.
• Optional PyMuPDF (fitz) + Pillow — QA rasterization and image rendering.
• Optional Microsoft Word (COM, via win32com) or LibreOffice (soffice) — final render-to-PDF check.
• scripts/toolcheck.py resolves the pandoc / latexmk / xelatex executable locations for scripts/convert.py, scripts/render_latex_env.py, and scripts/gen_reference.py; gui.py, the round-trip helpers, and tests/run_tests.py still call bare pandoc/xelatex/latexmk from PATH.

Consolidation notes

This skill merges four predecessors; their old names route here via the description above:

• manuscript-editing-template-latex → workflow A (references/roundtrip.md).
• latex-to-docx-fidelity → workflow B (references/tex-to-docx-engine.md).
• latex-from-mixed-sources → workflow C (references/mixed-sources.md).
• tex2docx (retired). Its pandoc-postprocessing fix registry was diffed against the fidelity engine: six of seven fixes were already subsumed architecturally, because the engine prevents upstream what tex2docx patched after the fact — tables, figures, and captions never pass through pandoc, so the damage tex2docx repaired downstream never occurs. The two genuine gaps were ported INTO the engine: the notes-punctuation normalizer (normalize_note_text, applied to table-notes text in both table engines) and \include{} expansion (in expand_latex_inputs). See the "Ported from tex2docx" section of references/tex-to-docx-engine.md.

tests/ holds the 46-fixture (t01–t46) regression harness driven by tests/run_tests.py and scored by tests/qa_metrics.py. It ships source .tex fixtures only — compiled PDFs, rendered PNGs, and .docx artifacts regenerate on each run. Fix the converter under scripts/, never the fixtures.

Caution: gui.py bypasses permission prompts

gui.py invokes the Claude CLI with --dangerously-skip-permissions for its auto-edit and auto-fix loops (the "Edits" box and the automatic XeLaTeX-error repair). This bypasses the normal permission prompts, so any tool call Claude makes in those loops runs without confirmation. Review gui.py (see the two --dangerously-skip-permissions invocations) and understand what the auto loops can do before enabling that path.