Hwpx editing

Skill kangdacool/hwpx-editing-skill/skills/hwpx-editing

LLM 에이전트(Claude Code·Codex 등)가 한글 .hwpx 파일을 안 깨고 편집하게 해주는 검증된 규칙 + 파이썬 도구 스킬. A portable Agent Skill that teaches Claude Code, Codex, and other LLM agents to edit HWPX (Hangul .hwpx) files without corrupting them.

Install

npx -y skills add kangdacool/hwpx-editing-skill --skill hwpx-editing

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

6 stars6 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Safely read, edit, and convert HWPX (Hangul / 한글 .hwpx) word-processor files with Python + lxml without corrupting them. Use this whenever a task involves a .hwpx file, a 한글 / Hangul / 한컴 (Hancom Office) document, HWPML, or a Korean government / academic / 논문 / 보고서 document — including reading or extracting text and tables (e.g. exporting complex merged tables to Excel / .xlsx), editing paragraphs, tables, images, equations, footnotes/endnotes, or memos, adding or positioning captions, fixing layout (orphaned headings, blank pages, columns), building a table of contents, or repackaging the zip. Also trigger when a 한글 file "won't open" / "is corrupted" (파일이 깨졌다 / 한글에서 안 열린다), or when the user hands you a legacy .hwp (this skill detects it and tells them to convert to .hwpx first). Trigger even if the user only says "edit this 한글 file" or ".hwpx" and doesn't mention the internals — naive edits (re-zipping, stale line caches, cloned ids) make 한글 refuse to open the file.

The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.

SKILL.md

8.0 KB, as published. Nobody here has run it

HWPX Editing

HWPX is a zip of XML (HWPML). The traps that corrupt a file aren't obvious from the outside, so read the relevant section of references/hwpx-guide.md before editing, and run the scripts in scripts/ to repack and verify — don't hand-roll the zip or eyeball correctness.

HWPX only. This handles .hwpx (zip + XML) exclusively. A legacy .hwp (OLE binary, signature D0CF11E0) must first be converted in 한글 via 다른 이름으로 저장 → HWPX(.hwpx). The scripts detect .hwp and say so.

The one rule that matters most

Never re-zip an HWPX with a normal zip writer. 한글 rejects a file whose unchanged entries were re-deflated. Use the raw-preserving repacker (scripts/hwpxlib.py:repack_preserve): it byte-copies every entry you didn't touch and re-deflates only what you changed, so a no-op repack is byte-identical to the source — meaning "if the original opens in 한글, your edit opens too."

Workflow

Inspect first. python scripts/inspect_hwpx.py FILE.hwpx --breaks — see per-section counts (paragraphs, tables, pic, equation, fields) and, crucially, which paragraphs carry a hidden pageBreak/columnBreak. If a heading is split from its content or a page/column is blank, hunt those breaks first (§6-A) — don't reach for keepWithNext. Body-paragraph breaks are usually leftover cruft.
Read the matching guide section in references/hwpx-guide.md (map below). The XML ids/refs (charPrIDRef, paraPrIDRef, borderFillIDRef, …) differ per file — always read them from the actual file, never assume.
Edit the XML with lxml, following the invariants:
- After editing or creating any paragraph, remove its <hp:linesegarray> (cached line layout goes stale → broken spacing). After structural edits, strip linesegarray from the whole section so 한글 fully re-lays-out. Use hwpxlib.strip_linesegarray.
- When you clone a node (endnote, table, equation, image), it inherits the original's id/instId → duplicates → instability. Reassign fresh ids with hwpxlib.make_uid, including nested subList>p / tbl/tc/p ids.
- Reuse existing charPr/paraPr definitions instead of adding new ones; if you must add, update the itemCnt or 한글 rejects the file.
Repack with hwpxlib.repack_preserve(src, changed, out, added): changed = edited entries (keep the XML declaration on top), added = new entries like BinData/imageN.png or a new sectionN.xml (also register these in content.hpf).
Verify every build: python scripts/verify.py EDITED.hwpx --orig ORIG.hwpx. All hard checks must pass (byte-identity self-check, well-formed XML incl. content.hpf, zero duplicate ids, zip integrity + mimetype first/STORED). It also prints a minimal-change diff so you can confirm only intended changes are present.
Round-trip in 한글. LibreOffice can't render HWPX, so render-dependent judgments (which heading orphans, whether spacing looks right) need the user to open the file in 한글 and, for equations/TOC, run 도구→차례 새로 고침 or double-click→close to finalize. Say so.

Scripts (`scripts/`) — run these, don't reinvent

Script	Purpose
`hwpxlib.py`	Library: `repack_preserve`, `pick_template`/`clone_para`/`run_patterns` (문단 복제 — 직접 하지 말 것), `own` (real body text, excludes 각주·미주·메모 / footnote·endnote·memo), `make_uid`, `strip_linesegarray`, `find_duplicate_ids`, `structural_counts`, `replace_image` (swap a picture + update every geometry field incl. the easy-to-forget `imgDim`), plus §7 verify helpers. Import it.
`inspect_hwpx.py FILE [--text] [--breaks]`	Structure dump; find hidden page/column breaks.
`verify.py EDITED [--orig ORIG]`	Run the full §7 checklist; non-zero exit on failure (CI-gateable).
`selftest.py`	Prove the repacker is lossless without a real file.
`tables_to_xlsx.py FILE [OUT]`	Export tables to Excel (.xlsx), merged cells preserved (needs `openpyxl`). Cells with only an image/equation → `[그림]`/`[수식]`; legacy `.hwp` is refused with guidance.
`audit_typography.py FILE [--expect-face 이름] [--expect-body-pt N]`	서식 일관성 감사: 실제 쓰이는 charPr을 사용횟수·크기·글꼴로 집계. 글꼴 혼재와 JUSTIFY+KEEP_WORD 영문 문단(자간 벌어짐)을 잡는다. `--expect-*`를 주면 어긋날 때 종료코드 1.
`hwpx_to_markdown.py FILE [OUT.md]`	Extract the document's text + tables as Markdown (for an LLM to read/summarize; no OUT → stdout).
`hwpx_to_docx.py FILE [OUT.docx]`	Export to Word (.docx), tables + merged cells preserved (needs `python-docx`; for journals / non-한글 co-authors).
`data_to_hwpx_table.py DATA.(xlsx\|csv) TARGET.hwpx [OUT] [--sheet]`	Insert an Excel/CSV table into a 한글 doc, merged cells preserved (needs `openpyxl` for xlsx; TARGET must contain a table for cell styling; CSV encoding auto).

Scripts need lxml (pip install lxml); the table→Excel export also needs openpyxl. Python 3.10+.

Where to read in the guide (`references/hwpx-guide.md`)

Load only the section you need — the guide opens with a "흔한 실패 TOP" (top failure modes); skim that first, then jump to:

§1 Structure & parsing — namespaces, own(), "check all sections", table text per-cell (not bulk itertext).
§2 Repack (raw-preserving) — the crown-jewel repacker. Mirrors hwpxlib.
§3 Common rules — linesegarray removal, id-dedup after cloning.
§4 Content editing — paragraphs, tables (cell width sums must equal table width or 한글 rejects it; merges), images (orgSz=px×75, size math, swap via replace_image so imgDim is never forgotten — a stale imgDim crops the picture; verify by content/한글-render not extraction order, matplotlib legibility "small figure, big text"), formatting, endnotes/footnotes (<hp:ctrl> wrapping required), memos, and typo-audit scoping (no blanket regex).
§5 Columns · TOC · equations — colPr, TABLEOFCONTENTS from outline-level paragraphs, 한컴 equation script (not LaTeX; table of forms).
§6 Page/column management — hidden breaks first, orphaned-heading prevention (keepWithNext, no empty paragraph after a heading, conditional columnBreak), blank-page cleanup, wide tables → 1-column region + moving a secPr block across section boundaries.
§7 Verification checklist — what verify.py automates, plus the round-trip caveat.

Guardrails

This edits documents only; it never needs the user's credentials, and it doesn't fetch or execute remote content. Work on a copy and keep the original.
Preserve the author's content: only change what the user asked for; the minimal-change diff in verify.py is your proof.