Md2latex

md2latex

Convert a Markdown file to LaTeX. The heavy, mechanical work is done by a bundled Python script (zero model tokens); you only spend tokens on the few semantic fixes the script reports.

Guardrails (always)

• Never modify the source Markdown. Read it only.
• Convert grammar, not content. Never reword, add, or drop the author's text.
• No personal data in output. Leave \author{names} as-is; never insert real names, emails, IDs, or passwords (outputs may be published).

Workflow

Let SKILL_DIR be the directory containing this SKILL.md.

1. Run the converter (it writes into the current working directory):

   python3 "$SKILL_DIR/scripts/md2latex.py" <target.md>
   

   It writes ./<basename>.tex, ./latexmkrc and — for templates that split their preamble — ./config.tex, and prints a POST-EDIT HINTS report to stderr. It auto-selects the template; override with --template general_jp|general_en|exercise only if the user asks.

2. Read the hints from stderr. They tell you exactly which lines still need a human decision — do not re-scan the whole file.

3. Apply only the reported edits to the generated .tex (use Edit):

   • Equation labels: the script turned each \tag{N} into a placeholder \label{eq:eqN}. Optionally rename to a meaningful label (e.g. eq:sin_function) — if you do, update every reference to it.
   • Cross-references: rewrite the listed prose references to \eqref{eq:...} (equations) or \ref{...} (figures/tables/sections), using the label map in the hints. Verify each candidate — some matches (e.g. a number that appears inside a caption) are not references. For a referenced section, add a \label{sec:...} to that \section first.
   • Empty captions: figures/tables with no adjacent > caption block get \caption{}. Leave them empty (do not invent text) unless the user provides the caption.
   • Subproblems / lists: the script folds each subproblem's full body (math, continuation prose, nested lists, figures) into its \item; just confirm the (a)/(b)/... order matches the source.
   • Callouts: Obsidian callouts (> [!theorem] …) are converted automatically; act only on flagged ones (a title-less [!law], an unmapped type rendered as info, or a degradation to quote in the exercise template).

4. (Recommended) Compile to confirm it builds, then report the PDF:

   latexmk <basename>.tex
   

   general_jp and exercise use LuaLaTeX; general_en uses pdfLaTeX (the bundled latexmkrc selects the right engine). Re-run once after adding references so the .aux cross-references resolve. Fix any compile errors.

5. Report the generated files and any captions left empty / references you were unsure about, so the user can finish those.

Conversion rules & template selection

The full, authoritative grammar mapping and the template-selection rules live in reference/conversion-spec.md. The semantic post-edit details live in reference/post-editing.md. Read them when you hit a case not covered above.