Version 2.0 separates ordinary files from formal letterhead
The skill's most useful decision happens before it writes a paragraph. Ordinary reports default to an ordinary profile and do not gain a red agency heading merely because an organization name appears. Formal documents require an explicit formal choice. That branch defaults to preprinted letterhead, leaving room for the paper's existing red mark and line. A complete digital letterhead must be requested separately.
That separation prevents a common automation mistake: producing an official-looking file when the user wanted a formatted report. The generator also handles document numbers, signer lines, attachments, colophons, and centered or alternating page numbers. Letter, command, minutes, and horizontal-table formats have dedicated branches. Word and WPS heading styles are written through 4 levels so a user can insert or refresh a table of contents later.
What happened when we ran it
Our lab classified commit 6a5e47a as not runnable in the standardized sandbox. The repository is identified as JavaScript, but the harness found no supported ecosystem and no Dockerfile. We did not produce install, build, test, package, timing, vulnerability, file-count, or source-line measurements. Any claim that it passed in our environment would be false.
The repository does contain direct Node scripts, shell tests, fixtures, a DOCX verifier, and visual-audit assets. Those facts come from the source, not from a completed lab run. The README says version 2.0 was checked with 18 generated DOCX files and 37 rendered pages. We did not reproduce those counts, inspect their underlying runs, or convert them into a pass rate.
The verifier checks structure, while final layout stays manual
The generator creates DOCX XML directly and the companion verifier inspects the elements it promises to control. The documented workflow also checks the ZIP container, converts the file to PDF with LibreOffice, and renders pages for visual inspection. Strict font mode stops when standard Xiaobiaosong or Fangsong faces are unavailable instead of silently accepting a substitute. That behavior is useful when a fallback would change line breaks.
Several boundaries remain outside the verifier. The project tells users to inspect horizontal-table orientation, actual seals and signatures, paper, binding, page breaks, attachments, and letterhead alignment in the target Word or WPS setup. A file that passes XML checks can still paginate differently on another workstation. The deliverable should be judged on the machine and printer that will produce it, not only in a headless conversion.
Security starts before the file reaches the model
The README opens with a data-handling warning. The skill does not decide whether internal, sensitive, restricted, or confidential material may leave an organization. Users must confirm the current model and processing environment are authorized, remove sensitive content where required, and apply their own access, storage, logging, transmission, and output controls. Local or intranet deployment is suggested for organizations with that capability.
This warning matters more than the formatting accuracy. Government and enterprise documents often contain names, decisions, schedules, or attachments that cannot be pasted into a public model. The skill can arrange the text after the user supplies it; it cannot grant permission to process that text. A safe rollout needs a written data classification rule and an approved execution environment before anyone installs the agent integration.
Nine agent directories share one source skill
The cross-platform installer supports Codex, Claude Code, OpenCode, Trae Code, Trae CLI, Kimi Code, TraeWork, WorkBuddy, and ZCode. On macOS and Linux it links the source directory into user-level skill locations. Windows copies it by default to avoid symbolic-link permission trouble. Packaged ZIP files cover TraeWork and WorkBuddy import paths.
A shared source makes rule changes easier to maintain across several assistants. It does not make their model behavior identical. Each agent still has its own instruction loading, file access, confirmation rules, and document context. Teams should test the exact agent plus Word or WPS combination they will use, then pin that setup instead of assuming one successful platform trial proves all 9 paths.
The September 13 push shows current maintenance
GitHub reported 743 stars, 1 open issue, and a last push on September 13, 2026. There is no tagged GitHub release even though the skill metadata says version 2.0.0 and the repository includes packaged distribution files. The only open issue is positive AutoClaw usage feedback, which is useful as a user report but is not independent conformance testing.
For the right team, this is a practical bridge between Markdown-like input and editable Chinese formal documents. Its audit matrix and screenshot evidence are more candid than a promise of one-click compliance because they name what automation cannot prove. Our missing lab run keeps the recommendation conditional. Trial it with harmless sample content, compare every page against an approved template, and retain a person who knows the standard as the final sign-off.