Breaking down 60 award-winning papers sentence by sentence into a paper-writing skill: one WeChat article earned 100 stars

gulucaptain/Best-Paper-Writing-Skill is a paper-writing skill pack for Claude Code and Codex. The author annotated and counted sentence by sentence the abstracts and introductions of 2021–2025 60 CVPR, ICCV and NeurIPS award-winning papers, turned the patterns into rules, and lets an agent draft a full manuscript following the "award-winning paper style". Created Oct 28, reached 6 278 stars by Nov. Nearly all of its stars came from one Chinese WeChat Official Account article, published two weeks before the CVPR 2027 deadline. Below: what it is, where the stars came from, why it has plateaued since that wave, and how to borrow its playbook for a project that can earn stars too.

The evening of Nov 2, an article titled "One abstract generates a 6-page paper"

On Nov 2 21:22, 2026, an article titled "One Abstract Generates a 6-Page Paper: The Strongest CVPR Draft-Writing Skill Is Here" was published on the WeChat official account. The Sohu account "Technology Playground" reposted the full text that evening. The opening line:

"No full paper text needed, no web search needed. One abstract is enough to generate a six-page CVPR draft."

The demo used the three award-winners just announced at CVPR 2026: best paper D4RT, best student paper O-Voxel, and honorable mention SAM 3D. Taking D4RT as the example, the agent received only the original abstract and was not allowed to browse. It produced a draft of about six pages that could be compiled directly into the CVPR template. Experiment numbers, baselines and citations that were not provided were all marked [TODO]. The second half of the article showed a comparison table: in a generic AI-written draft, about 32% of sentences use the sentence-final participle pattern "..., making ..."; in award-winning papers, only about 8% do.

The timing was well chosen. The CVPR 2027 paper registration deadline was Nov 10, and the full submission deadline was Nov 16 (per the CVPR website). Readers of this article were exactly the CV graduate students holding a draft two weeks from the deadline.

What it is

In one sentence: a skill that makes AI write paper drafts according to the playbook of award-winning papers.

There are two ways to use it: fill in an offline form in the browser (5 required fields: a one-sentence summary of the work, the task, the situations where existing methods fail, the method components, and the headline result), or paste an abstract directly. The skill first breaks the abstract back into these fields, tells you which are present, vague or missing, asks one follow-up question, and then writes the title, abstract, introduction, method and experiments, plus a report.md for the author (core claims, a claims-to-evidence table, and a to-do list of gaps). v1.1 added figure planning: based on the figure statistics of 44 award-winning papers, it plans Figure 1 and the method overview figure, and can draw schematic diagrams directly. It also includes two check scripts that catch AI-sounding prose, em dashes, stacked catchy phrases and numbers that appear from nowhere. The repo ships with the CVPR 2027 LaTeX template.

The README has many numbers that work as headline material: 1,929 sentences annotated one by one, 475 searchable exemplars, 20 writing metrics, and 8 paper types each with one argument route (method improvement, new capability, scaling laws, re-examination, mechanism analysis, theory, dataset, measurement); "100% of introductions first give specific credit to prior work, and about 98% then point out the gap and its cause".

The author's homepage is credited to Haoyu Zhao, who works on video generation and camera control. He has his own paper code repos for ICML'26 and ACM MM'26 (a couple hundred stars), and 149 GitHub followers. The workflow and template were modified from Master-cai/Research-Paper-Writing-Skills (8,134 stars), with credit given in the README. License CC BY 4.0, free.

How the stars came in, day by day

(Daily GitHub star counts, fetched once a day.)

One article brought roughly 100 stars and 7 forks. No evidence was found for where the first 63 stars came from; the inference is the author's own circle, same-lab peers and project homepages. X, Reddit, Hacker News, Zhihu and Xiaohongshu return no hits for it. A few later GitHub aggregator sites (skillsllm, vibecrowd, etc.) list it, but that was a result of the stars rising, not the cause. Which WeChat account originally published the article could not be verified (the WeChat page requires a captcha); the inference is a large AI-academia account. The article's material and the README correspond almost one-to-one, suggesting the author supplied the content.

Why this one got written about

Paper-writing skills are not new, and the category is already crowded. This one earned a standalone article because of two things others lacked.

First, the rules were counted, not "felt". Most rival skills rely on experiential rules, from a teacher's notes or the author's own tips. This one says "across 60 award-winning papers, 100% do X and 98% do Y" and "32% of AI drafts versus 8% of award-winning papers". Media can use numbers like that directly as subheadings, and readers find them credible.

Second, the demo used a famous benchmark everyone recognizes. It took the best paper of that year, used only the abstract, reconstructed a six-page draft, and showed the results side by side. Readers don't need to install or try anything. Seeing the picture is enough to understand what it does.

Add publication two weeks before the deadline, with three hooks in the title ("Best Paper", "CVPR", "one abstract"), and it was practically a direct call-out to Chinese CV/ML graduate students.

Why growth stalled

After that wave passed, daily gains fell back to the twenties, close to where they were before the article. The reason is simple:

• The category is crowded, and this is a very small slice of it. Similar projects range from tens of thousands of stars and are still growing:

Its upstream gained in one week about as many stars as this project has in total since it was created.

• It had only one hook. The "60 award-winning papers analyzed" hook was used up in one go. No second article or channel picked it up, and the English-speaking community never touched it.
• The author did not keep updating. No new commits after 10-29. Energy shifted to the sibling projects Best-Paper-Review-Skill (22 stars) and Best-Paper-Polish-Skill (no stars yet).
• Nobody has said whether it works well. Stars don't mean people are using it, and there is no third-party evaluation of the quality of the drafts it produces.

I haven't found anyone else building a "best-paper corpus" version of this skill. The dozen-plus same-named 0-star repos on GitHub are just re-uploaded copies.

If it were you

The tactic is worth studying. Copy it and expect a few hundred stars. It shows that "authoritative samples + stats + a showcase demo + pitching WeChat Official Accounts before the deadline" can earn attention in a crowded niche, and it's cheap: one person can do it in a week or two. But its own result is the ceiling: about 100 stars for the main article, then it faded. Paper-writing is dominated by tens-of-thousands-star full suites, so a standalone tool is hard to remember.

Good fit for: people who publish at top venues or are PhD students in a given field, already have a corpus on hand, and have the patience to annotate sentence by sentence; people who want a small project they can put on a CV and that peers will share, without chasing big star counts.

Bad fit for: anyone aiming for thousands of stars (that niche's stars flow to the full suites); anyone unwilling to do corpus annotation and only wanting new prompts (without the stats there's no hook); anyone who only promotes in English-speaking communities (its traffic came entirely from Chinese WeChat accounts).

How to do it, best to worst:

1. Swap in a "gold-standard corpus" nobody has counted yet: ACL/EMNLP outstanding papers (NLP), ICLR outstanding papers, abstracts from Nature/Science main journals, funded NSFC proposals, successful rebuttals at top venues. As long as the "gold standard" is recognized and the samples are obtainable, copy the method: annotate sentence by sentence, extract a few rules with percentages, and build an "AI draft vs. gold standard" comparison table.
2. Fill the gaps it left: it only counted abstracts and introductions. Methods, experiments, and rebuttals have no corpus. It led with a Chinese README and never touched the English community. Covering "how award-winning papers write their experiments" or "how accepted rebuttals are written" is the open space.
3. Change the audience: Chinese thesis writing, medical SCI, and social science papers each have a crowd anxious about deadlines. A social-science version already exists (fakerqwq/social-science-paper-writing-skill, 463 stars), but none have corpus statistics.

How to get users: follow its rhythm, but don't post once. Work backward from the deadline: publish two to three weeks before it, use a recent award-winning paper as the demo ("abstract only, reconstruct the draft"), package the comparison table and side-by-side figures, and pitch them to Chinese AI academic accounts. Each venue's deadline gives another hook (CVPR November, ICML January, ACL February, NeurIPS May), which it didn't do. Write an English README and post the same comparison table to X and r/MachineLearning.

Timing: the most recent window is before the ICML 2027 deadline (usually late January; check the official site for the exact date), so it needs to go out in early January of 2027. Longer term, the "agent-written papers" wave is still going, but the full suites keep absorbing niche features, so the later you launch, the harder it is to stand out.

Signs you've misread it: ① your project gets a Chinese WeChat article but under 50 stars in a week, which means readers are already tired of "award-winning paper statistics" hooks; ② Best-Paper-Writing-Skill itself doesn't rise again around the CVPR deadline (11-16), meaning deadline anxiety doesn't move this kind of tool; ③ a top full suite (like academic-research-skills) adds award-paper-based rules, and the niche project loses its position.