Doc tools

Document tools

Four extractors on the PATH, plus Python libraries for authoring and editing. All four share the same interface: TOOL INPUT [-o OUT] [-f txt|md|json]. - as INPUT reads stdin. Default output is plain text on stdout.

Extracting

pdftotext report.pdf                      # all pages, reading order
pdftotext report.pdf --pages 3-7 --tables # page range + table extraction
pdftotext scan.pdf --ocr --ocr-lang hun   # scanned pages with no text layer
pdftotext report.pdf -f json -o out.json  # per-page records

doctotext memo.docx -f md --tables        # headings as markdown, tables included
doctotext draft.docx --comments           # append reviewer comments
doctotext deck.pptx --notes               # slides + speaker notes
doctotext legacy.doc                      # routed through LibreOffice

xlstotext model.xlsx                      # all sheets, computed values
xlstotext model.xlsx --formulas -s Data   # formulas from one sheet
xlstotext old.xls --csv -o data.csv       # legacy .xls to CSV
xlstotext big.xlsx --max-rows 50          # peek at a large sheet

latextotext paper.tex --bodyonly          # prose only, no preamble
latextotext paper.tex --sections          # outline of the section tree
latextotext paper.tex --pandoc -f md      # via pandoc; better cross-refs

Pick -f json when the output feeds another program — you get page/sheet/slide boundaries as structure instead of guessing at separators.

--formulas matters when auditing a spreadsheet: without it you see cached values, and a file never opened by Excel has no cached values at all (formula cells come back empty). Prefer --formulas when the logic is the point.

Editing and generating

Libraries, not CLIs — write Python against these:

Need	Library
Read/write .docx	python-docx (import docx)
.docx from a template with placeholders	docxtpl (Jinja2 syntax inside Word)
Merge several .docx into one	docxcompose
Read/write .xlsx, formulas, charts	openpyxl, xlsxwriter
Read/write .pptx	python-pptx
PDF manipulation (split/merge/annotate)	pymupdf (import fitz), pypdf
PDF tables and layout	pdfplumber
LaTeX → text, macro expansion	pylatexenc
LaTeX AST parsing and rewriting	TexSoup
Any-to-any fallback conversion	pandoc (CLI, already installed)

For .docx work that needs styling beyond python-docx's API (page numbers, letterheads, tracked changes), see the anthropic-skills:docx skill — it goes deeper on OOXML manipulation. This skill covers extraction and the plumbing.

Compiling LaTeX

TinyTeX lives in ~/Library/TinyTeX, binaries symlinked into ~/.local/bin.

latexmk -pdf -interaction=nonstopmode paper.tex   # handles reruns, biber, refs
tlmgr install <package>                            # add a missing package

latexmk exit 0 means a clean build. A PDF can appear even on a failed run — always check the exit code, then grep -E "^!" paper.log for the real error.

Hungarian is set up (babel-hungarian + hyphen-hungarian, formats rebuilt). If you add a language, install its hyphen-* package and run fmtutil-sys --byfmt pdflatex or hyphenation silently fails.

Legacy binary formats

.doc, .ppt, .odt, .odp are converted by shelling out to LibreOffice, installed at ~/Applications/LibreOffice.app (no sudo, not on the PATH — the tools find it themselves). .xls is read directly by xlrd, no LibreOffice needed. Every XML-based format works without it.

Feeding documents into memo-index

For a large pile of documents, don't extract them one by one — the memo-index skill's scripts/doc_ingest.py batches every format through these extractors and also rasterizes PDF pages for its vision pass. Use that when the goal is "read all of these and tell me X" rather than reading one file.

Notes

• The shebangs point at /Users/szili/anaconda3/bin/python3 on purpose, not at env python3. A login shell resolves python3 to the python.org framework build, which does not have these libraries; pip3 installs into Anaconda. If the Python setup ever changes, repoint the shebangs and reinstall the libs.
• OCR: Tesseract 5.5.2 lives in the ocr conda env, symlinked onto the PATH, with 125 languages including hun, eng, deu. Only pdftotext --ocr uses it. --ocr is per-page and lazy: a page that already has a text layer is read normally and never rasterized, so passing it on a mixed scanned/digital PDF is safe and costs nothing on the digital pages. Default --ocr-lang is eng+hun. Expect OCR to mangle math and to drop accents from capitals (É → E) — if a number matters, verify it against the image rather than trusting the OCR text.
• The implementations are in bin/ next to this file; _doccommon.py holds the shared argument parsing, LibreOffice discovery and output handling.