A anime-drama studio broke its production pipeline into 12 skills: how drama-skills earned 3,800 stars in a crowded niche

There are over eighty AI short-drama skills on GitHub. drama-skills isn't the first, and it doesn't have the most stars, yet it has held a top spot steadily: 113 days since the repo was created, 3,825 stars and 688 forks (GitHub, 11-06), gaining thirty to forty stars a day with no break. What it runs on is a real production pipeline from an anime-drama studio, a sister project already at 8,000+ stars, and a distribution approach that covers every channel. This piece covers what it is, where the stars came from, why this project and not others, and which parts of its playbook you can copy and which you can't if you want to build a star-earning skill pack.

"80k lines of code, and we can't maintain it anymore"

In Aug 25 2026, Jason Zhu (@GoSailGlobal), who makes AI-going-global content on Chinese X, posted with a 31-second sample clip:

If you make AI short dramas, you'll eventually use this open-source workflow. drama-skills: 10 skills distilled from a studio's own pipeline, 1000+ stars. They've made over a thousand short-drama projects since 2025, and the custom tool they built, frontend and backend, came to nearly 80k lines of code. Against the pace at which models and requirements change, they slowly couldn't keep it up. So they dropped the all-in-one graphical tool and distilled everything into skills an agent can read. ... The most important design choice is putting confirmation before production ... you look at the preview and clearly nod, and only then does it get sent off to burn money on generation.

This post got 28k views, 457 bookmarks and 287 likes (on X, 11-06 reads). More bookmarks than likes means people plan to come back and use it.

The post almost copies the "origin story" section of the repo README word for word. The last line of the original is blunter: "What's left in our in-house tooling now is just the queue for pulling cards." A studio that has shipped over a thousand projects admitting its 80k lines of tooling were worse than an agent plus a pile of Markdown: that story is the project's best ad.

What it is

12 skills you install into coding agents like Claude Code and Codex. One command, npx skills add zenstory-ai/drama-skills, and you go from an idea or a source novel all the way to a finished cut:

source review → episode scripts → character / scene / prop designs → image prompts and storyboards → video prompts → after your confirmation, model generation → edit → review.

A few design choices that are rare in this category:

• Only five Markdown files per episode: script, visual design, storyboard, image prompts, video prompts, plus an edit list at the editing stage. Every creative decision lives in a file you can version with git and edit by hand.
• Continuity locks: looks that must hold across shots are written as a short phrase you can paste verbatim into prompts, with a check script that verifies every shot references them correctly.
• Preview before you spend: for any image, video or voice-over job, the exact count, content and parameters of the batch are laid out first. Paid APIs are only called after you explicitly confirm.
• Model-agnostic: writing runs on whatever model is in your agent; video works with Seedance, MiniMax H3 and several others.

The first screen of the README is a 68-second vertical sample: the opening of episode 1 of "Let You Manage the Account", generated shot by shot in 20 shots and cut together, with the note "video generation cost about ¥40" beside it. Below are compilations of fight scenes, crowd scenes and ink-wash comic dramas. MIT licensed, with 6 releases from mid-October to early November, going from 11-06 to v0.11.0.

Where the stars came from

On every day we can see, the growth stays between 26 and 57. No single day spiked, and it never hit GitHub Trending. This wasn't pushed up by one viral post. Several pipes were pouring into it at once:

Cross-promotion from a sister project. The same author's earlier project, oh-story-claudecode (a skill pack for AI-written web novels, created in April), has 8,334 stars. Its README has a "related projects" section that links straight to drama-skills, and the author's profile pins the two side by side. People who write web novels are the natural next step toward short drama, so this audience was ready-made. The organization's 6 repos total 14.5k stars.

Automatic listing in skill directories. The install command is written as npx skills add, so a single install lands it on the skills.sh install leaderboard. On skills.sh, its 13 skills total 25.3k installs, with the three core ones (script, storyboard, video prompts) at about 2,400 each (skills.sh, 11-06). A rough estimate is that a couple thousand people have installed the full set (estimate). SkillsLLM, skills-hub.ai, TypingMind and a dozen or so other aggregators also picked it up automatically, so it shows up everywhere when you search for "short drama skill".

Lists and evaluation endorsements. The awesome-short-drama-skills list (1,845 stars) that appeared in mid-October features it as a curated pick in the full-pipeline category. In a "blind review of script skills" (10-18), it was one of 4 winners in three rounds against the control group with no skill installed. There is also a Zhihu article, "I reviewed 198 AI short-drama skill repos and only kept 82". The timing of the list's appearance matches the steady 40+ daily star rate from mid-October on, but single-day star counts can't be precisely attributed to that list.

The official site wins search. zenstory.ai hosts 30 bilingual (Chinese and English) guides on writing short drama: how to write payoff moments, where to place paywall cliffhangers, reincarnation plots, cute-baby plots, vertical script format, and how to adapt novels into short dramas. Every article links to drama-skills. The site also publishes an llms.txt file for AI search engines to read.

There's also a Zhihu tutorial on using WorkBuddy, chapter 8 of a DeepSeek tutorial book, and a Telegram group. I found no sign of paid promotion or influencer product reviews on Douyin or Xiaohongshu.

Why this one

Timing. In 2026, AI short drama and comic dramas were the hottest areas for Chinese creators. The all-in-one open-source platform Toonflow had 20k stars and Huobao Drama had 17k. In the second half of that year, the Agent Skill spec and the npx skills distribution system were just taking shape. The idea of packaging an industry's craft as a skill opened up right when this project was created.

The story is first-hand. "Over a thousand projects, 80k lines of code that couldn't be maintained, and turning it into skills made things easier" is not something a prompt collection can fabricate. The big accounts shared exactly that paragraph, paired with a real finished cut that states its cost, and readers find it hard not to believe it.

Cold start isn't from zero. oh-story already proved the team's skills work well, and drama-skills is its sequel.

Distribution covered every surface. The install leaderboard, aggregator sites, awesome lists, long-tail articles on the official site, and a bilingual README. Each pipe is small, but together they make a steady long tail of roughly thirty to forty stars a day.

But one thing needs to be clear: the most complete, rigorous pipeline isn't the main reason it gained stars. eternityspring/shuohao-skills, launched 3 weeks later, covers only script-to-storyboard in 7 skills and doesn't handle production or editing. It has 6,075 stars now, +390 over the last ~7 days, and has already overtaken it (GitHub, 11-06). In this niche, simple and easy to start gets stars more easily. drama-skills' full pipeline and verification scripts are more a reason for professional studio users to stay. A 18% fork ratio (new repos of similar size are around 11%) suggests real people took it and modified it.

Who it's competing with

The awesome lists count 83 short-drama skills so far, most of them appearing after July. The "short drama skill pack" slot is already crowded.

If it were you

C. Its playbook is worth learning from, but the short drama slot is already packed. It got going on sister-project traffic and the production-line experience from over a thousand studio projects, which followers can't copy. Another short drama skill pack would struggle to stand out.

Who it suits: people who have actually run a production line in some industry, with dozens or hundreds of real projects behind them (outsourced game art, e-commerce product pages, commentary videos, audiobooks), who want to open-source that craft and build a reputation.

Who it doesn't suit: anyone wanting to build yet another short drama or comic-drama skill pack (83 already ahead, and the top ones are still growing at 6,000 stars); people with no first-hand experience who can only stitch prompts together (you can't write the "origin story" that makes this project valuable); anyone hoping a single post will blow up (its stars are all long tail, with at most 57 in a day).

How to do it:

1. Pick a vertical industry that AI is rewriting in production but that doesn't yet have a decent skill pack. Signals: people in the industry already mass-produce with AI, a few all-in-one platforms have tens of thousands of stars, yet only a handful of skills that plug into agents exist.
2. Distill from your own real projects instead of writing from scratch. Put a real finished piece on the README's first screen and state what it cost. Then write an "origin story" section explaining how you used to do it and why you switched to skills.
3. Copy a few of its engineering details: use a few Markdown files to record each step's decisions; for every step that costs money, preview first and wait for user confirmation before calling the API; include a check script so the agent's output can be verified.
4. If you want stars, start with a simple slice. shuohao-skills beat the full pipeline by covering just script-to-storyboard. Build the most common stage first, then expand outward in both directions.

How to get users: write the install command as npx skills add your-name/repo, which automatically lands on the skills.sh install leaderboard and the aggregators. Turn the origin-story section into an intro people can copy directly and send it to Chinese AI recommendation accounts on X. Find the maintainers of awesome lists in your category and file issues asking for inclusion. Build a bilingual page and write a dozen or two long-tail articles on "how this industry's work gets done", each linking back to your repo. Once you've built one, chain sister projects that follow your users' next step, which is how web novels → short drama came about.

Deadline: while the skill packs in your chosen industry are still in the single digits. Short drama went from a few to eighty-some over two or three months starting in July, and that is probably how long a vertical category's window stays open.

How you could be wrong: (1) new short-drama skill packs appear in the coming months and, without a parent brand or studio backing, still quickly pass 3,000 stars, which would mean the slot isn't saturated and should be revised upward; (2) you apply this playbook to another industry, ship a real finished piece and an origin story, yet stars stay stuck in the dozens, which would show that drama-skills' stars mainly came from oh-story's existing users and the playbook itself isn't that strong; (3) install leaderboards like skills.sh stop giving new projects exposure, and the free distribution route narrows.

Data and sources: star counts, forks and daily numbers come from GitHub (10-09 to 11-06); install counts come from skills.sh (11-06); the text and engagement numbers of Jason Zhu's post come from X (posted 2026-08-25, read 11-06); the sample description and the origin story text come from the repo README (read 11-06); competitor star counts come from GitHub, 11-06; awesome-short-drama-skills and the blind review come from that repo's README and a Zhihu article (10-16). Figures marked "estimate" are estimates.