A government-document writing skill got copied wholesale by Tencent, and it gained 700 more stars

The "Official Document DNA" repo on GitHub (lieflat-gongwen) runs statistics on 1.02M characters of real official documents and turns "what good official writing looks like" into a set of measurable numbers, with a self-check script. In two months it grew to 2,400 stars and became the top repo among similar official-document skills. Neither surge came from the product itself. The first came from a big Chinese tech account sharing it. The second came from the author publicly accusing Tencent's WorkBuddy of copying it onto its skill marketplace. Below is what to learn from it, and what can't be copied.

"47 files fully identical"

On Nov 2 at 12:30 noon (Beijing time) 2026, an X account called "Lieflat" (@lieflat_3) posted, opening with:

So infuriating! Today I found several of my Skills renamed and relisted under WorkBuddy.

WorkBuddy is Tencent's AI agent aimed at office work. The author found an skill called "Corpus-Based Official Writing Coach" in its skill marketplace. The author field said "WorkBuddy", and once installed the package name was wb-lieflat-gongwen. That night he published a long post, "I Was Plagiarized by WorkBuddy," comparing files one by one. 47 files matched his open-source repo word for word, including the self-check script, which was identical at 461 lines and 22,005 bytes. The repo is under a non-commercial license that prohibits integration into commercial products.

Three other skills of his were copied in the same batch. Chinese AI accounts shared the post, reviving the old story of "Tencent copying open source again." That night WorkBuddy responded that it was a "third-party upload," and it took down all four skills.

Over those days the repo's star count moved like this (GitHub daily snapshots, Beijing time): 11-01 on 1,673,11-02, 1,937,11-03 on 2,229, and 11-06 on 2,378. The week before, it had gained only 66 stars in total. This week it gained 720. One dispute worked as free publicity across the internet.

What it actually is

A skill package for agents like Claude Code and Codex: a writing guide, a set of parameters, 21 model essays, and a Python self-check script. No dependencies. git clone it into your skills directory and it works.

It handles only seven kinds of material with no fixed template: research reports, leadership speeches, work opinions, experience write-ups, work plans, experience summaries, and Party-building materials. Statutory documents such as notices, requests for instruction, and replies are explicitly excluded, since their formats are fixed and there is no style to learn.

Writing a piece follows six steps, two of which closely match real institutional workflows: first produce an outline and stop for your confirmation, then run the self-check after drafting and revise based on the report. Any figure it can't verify is left as "【TBD: xx】" rather than invented.

Why this one, and not the other official-document skills

GitHub has many official-document skills, and some launched earlier (GitHub, 11-06):

It differs from the others in three ways.

The README overturns conventional wisdom before offering a solution. Most official-writing tools teach templates and stock phrases: openings like "Based on... for..." and closings like "Please follow and implement." He tested each claim against the full corpus: basis-style openings account for only 6–8%; closings like "please implement" are under 2%; and 60–72% of real official documents use none of these at all. The conclusion is a line people love to share: the templated "official-writing formulas" are actually the signature of low-scoring documents.

Style is turned into numbers you can check. Average sentence length is 53–56 characters, the enumeration comma appears 14–31 times per thousand characters, and the "first... second..." pattern varies by 6x across genres. Each genre gets a range rather than an average, and the self-check only flags "wrong genre" as an error, with everything else as suggestions. His reasoning: "Good writing isn't neat. Neatness is usually mediocrity." Anyone who has actually written official materials will recognize that this was made by someone who has.

Mistakes are kept on record. Five statistics were overturned on review (one, for example, went from 84% to 60%), and the process is documented. Revision records for the ten model essays are also in the repo. For a tech audience, that is a strong trust signal.

Differentiation explains why big accounts were willing to share, but not why the stars came.

Where the stars came from: two shares and one fight

The author isn't new. The same GitHub account had already published writing-dna-skill (3k stars), lieflat-charts (6.7k stars), and less-ai-tone (5–6k stars), all with the "lieflat" prefix. Together its 8 repos total 17.5k stars. The author says he was a journalist for several years and keeps posting on X and Jike ("姜汁猴"). By the time the official-document version launched, a group of people were already waiting for his next project.

Wave 1: Two big Chinese tech accounts, September 9–10. @geekbb posted about it on 9-09 and it gained 374 stars that day, reaching #21 on GitHub's all-site trending. @GitHub_Daily followed on 9-10, writing "this Skill has something to it," and it gained another 188 stars that day (daily data from trendshift). Two days later, 'Namei Knowledge Base' recommended it: "Anyone who writes materials inside the system should take a close look." For the next month-plus, it kept adding 5–25 stars a day.

Second wave: November 1–3, the dispute. This is the event from the opening. On trendshift, 11-01 and 11-02 each rose by 280 stars. Three of the author's other copied repos also jumped at the same time, and the official-document repo gained the most relative to its size (about a third more), because the long post put it first.

No traces found on Product Hunt, Hacker News, Xiaohongshu, or Douyin (not finding them doesn't mean they're absent). The fork-to-star ratio is about 13.6%, high for a writing skill. Many people apparently installed it into their own agents rather than just bookmarking it (inference).

The method can be copied, but the stars may not follow

The projects most worth watching are the ones already copying this approach of "distilling writing style from corpus statistics" (GitHub, 11-06):

Same method, but the stars differ by two orders of magnitude. Look at the official-document table from the other side: projects like gongwen-gbt9704-skill and official-document-skill have no author halo and ordinary methods, yet the topic alone got them 300–850 stars. Read the two tables together and two things stand out:

1. The topic matters more than the method. Genres everyone in Chinese workplaces has to write and finds painful (official documents, bids, civil service essays) naturally get searched and shared. Law and CHI papers have too narrow an audience, so even solid methods only reach a few dozen stars.
2. Between a few hundred and a few thousand stars, distribution is the difference. The extra 1,500 stars lieflat gained came from the author's audience, two big accounts, and one public dispute, not from more detailed statistics.

Who's following

• Copycats on Tencent's ecosystem: WorkBuddy's "Official Writing Coach" (now delisted). The author 11-03 also found that Chongqing-based "Tuguang Network Media" mass-published 2,000 skills on Tencent SkillHub, including his less-ai-tone.
• Later open-source entrants: dyz-gongwen-skills launched in October with "full coverage of 22 document types + anti-fabrication + confidentiality." Its README benchmarks against lieflat and scored it 5/10, and it gained 156 stars in a week. The topic of official documents is still bringing stars to newcomers.
• Content scraping: the author mentioned a WeChat Official Account article that rewrote his README and got over ten thousand reads.

If it were you

In one line: Grade B. The approach of "real corpus statistics → parameter ranges by genre → self-check script" transfers directly, and for a Chinese genre that everyone has to write, a few hundred stars is realistic. Passing a thousand depends on distribution, which means the audience the author built and that dispute, and you don't get those.

Who it fits: developers who have written some painful genre themselves (civil service essays, bids, performance reports, research reports, judgments, medical records, grant applications), who can gather tens of thousands of characters of corpus, and who know a bit of Python and agents. Under the X post, people are already asking "is there a civil service essay skill?" The top bid-document skill on GitHub, BiaoShu-SKILL, has 205 stars, and it works from industry norms plus templates. No one has done quantitative statistics on winning bid documents yet.

Who it doesn't fit: anyone who wants to build yet another official-writing skill (there are already five or six, and the top one has 2,400 stars, backed by an author with 17.5k stars); anyone picking a genre only a small circle writes (the law and CHI paper projects stalled at a few dozen stars); anyone without a real corpus who would pad the statistics with AI-generated samples. This approach depends entirely on the data being real.

How to do it: pick the topic first, then the method. Gather your own corpus, and start with the most counterintuitive statistics, the ones that show "people assume you should write this way, but the good ones don't." That page becomes your README's opening. Give parameters as ranges, not averages. Make the self-check block only hard errors, and keep the mistakes and review records in the repo. Build on what the others leave out: lieflat doesn't do formatting or statutory documents, and dyz scored only 5 when it benchmarked against lieflat. No one yet offers quantitative analysis plus full coverage across writing, checking, and formatting in one genre.

How to get users: follow his path. Use one fixed brand prefix, publish a series one after another so each new release draws on the last one's audience, and with each release bring your data conclusions to the Chinese X accounts that recommend GitHub projects (@GitHub_Daily, @geekbb, and similar). His first 560 stars came from two posts. Write the license and attribution in from the start, so you can publicly compare files line by line when a big platform copies your work.

How soon you need to move: there is no hard deadline. Now that agents support SKILL.md, professional Chinese writing skills are being filled in one genre at a time. Usually only the first one or two serious projects in a genre get shared by big accounts, and the later you come in, the harder it gets (inference).

Signs you misjudged: ① three months later, the existing projects in your chosen genre are still stuck at a few dozen stars, which means not enough people search that genre, so change topics; ② a big account shares your work and it still doesn't pass 200 stars that day, which means the opening page of your README isn't counterintuitive enough, and distribution isn't the problem; ③ platforms like WorkBuddy or WPS launch their own "quantified style by genre + self-check" features, and then the novelty of open-source skills like this fades quickly.

Data and sources

• Repo and stars: GitHub API (11-06). Daily stars come from GitHub daily snapshots (10-17 to 11-06, Beijing time) and trendshift (trendshift.io/repositories/209279, UTC days). The two use different day boundaries, so single-day figures don't line up exactly.
• Dispute timeline: @lieflat_3 11-02 post and long post, 11-03 post. WorkBuddy's response and takedown as relayed by the author.
• Shares: @geekbb (9-09), @GitHub_Daily (9-10), Namei Knowledge Base @fhwofjow51260 (9-12).
• Statistics and quotes: repo README (11-06).
• Similar and method-copying projects: GitHub search (11-06).
• Which post first sparked the 11-01 wave, and how many times the copied version was downloaded, have not been verified.