"Write a hand-drawn style, get four images in four looks": a style-numbered library that gained 7,000 stars in two months

yang0/handraw-style turns 300 hand-drawn styles into a "number + sample image" gallery. You pick a number, give a theme, and it returns a bilingual Chinese-English image-generation prompt. The author's earlier libraries mostly had 0 stars, and they had few followers on X. This repo was created in Sep 2026 5, and two months later it had 7,107 stars and 825 forks (GitHub, 11-06). It's still gaining sixty or seventy stars a day. It took off partly because it found the right pain point and partly because the author knew how to promote it. Both are worth learning, but the "style-numbered library" slot already has an occupant.

The first line of one post

On the morning of Sep 2026 9 (Beijing time) at 8 o'clock, GitHubDaily, a Chinese GitHub recommendation account, posted this:

"Using AI to make illustrations, you write 'hand-drawn style, cozy a bit' in the prompt and get four images that all look different. Switch models, and the style drifts again. The handraw-style Skill changes this to recording a number. It organizes 216 hand-drawn styles, each with a number and a sample image. Browse the gallery, pick a number, then give your theme."

The post included three gallery screenshots and ended with: "Remembering a number is much easier than memorizing style names. If you make illustrations for WeChat or Xiaohongshu and want your images to keep one consistent look, pick a style and note its number." A reply below said: "I just ran into this yesterday. I copied the same Song-dynasty poem prompt and got images that looked nothing alike." Another wasn't convinced: "I don't believe all 216 styles are really distinct. Neighboring numbers must look alike."

By the next day the repo had 752 stars (The Daily Commit logged that number that day). Before that it was a new repo only 4 days old, from an author whose GitHub account was registered in 2013 and whose earlier repos had barely been seen.

What it is

It's a Skill you install into agents like Codex and Claude Code. The first line of the README says it all: "Can't describe a style? Struggling with layout composition? Pick a number and generate highly distinctive AI image prompts in one click."

• Three numbering systems: as of early November there were 365 styles (hand-drawn illustration plus realistic photography), 170 layouts (social cards, infographics, comic storyboards, IP design, e-commerce images), and 36 theme colors (things like Klein blue, sage green and Hermès orange). The three can be combined: "Layout: SC-001, Style: FB-006, Theme color: C-01, Theme: the first milk tea of autumn."
• You don't have to pick: say "Help me generate a healing illustration prompt of a cat sunbathing on a windowsill, you choose the style," and it picks a style plus a theme color on its own, and explains why it chose them.
• It decides whether to attach a reference image: if you ask it to generate the image directly, it first checks whether the current model can produce the style from the style name alone. If it can, it attaches no reference image. Only if it can't does it send that numbered sample. GitHubDaily specifically praised this, saying it "keeps the subject and composition of the reference image from being locked in too."
• Installing takes one sentence: "Help me install this Skill: github.com/yang0/handraw-style", sent to Codex.

The people who use it are clear: Xiaohongshu covers, WeChat article headers, long infographic posts, comic strips, and later wedding and family portrait shoot plans. These people need images every day and can't name the look they want. The author put it on X this way: "As long as you pick the right style, making good-looking images isn't hard at all. The problem is most people don't have that knowledge."

How the stars climbed

(GitHub daily star counts; September figures come from the articles listed above.)

This curve isn't "push, spike, fall back." No single day stands out, but every week or two someone writes about it again. Forks are 11.6% of stars, in line with new repos of similar size over the same period, so there's no sign of star-farming. It has also never made GitHub Trending.

Why people keep writing about it, one article after another

1. The pain point fits in one line. "Write 'hand-drawn, warm feel' and get four images in four styles." GitHubDaily opened with this line, and several reposts started the same way. Once GPT Image 2 arrived, Chinese-language image output was already good enough. What people lacked was the vocabulary to describe the style.

2. The gallery is itself the cover image. Hundreds of numbered sample images tiled into one wall can go straight onto a WeChat or repost-account header. Reposting it takes almost no effort.

3. The numbers keep growing, so there's always a new article. The style count went from 216 to 261, 278, then 365. The author posted on X one update at a time: "New style #276: clear flat illustration (Xiaohongshu hit)", "New style: Tianshu Qitan", "New realistic category, 33 styles". Media outlets put each new number in the headline: 279 styles, 278 styles, 365 styles. The same project gets written about again and again, and each piece has a fresh hook.

4. The timing was right. 2026 September was when Agent Skills were at their peak in the Chinese-language community. Baoyu's baoyu-skills and various PPT Skills all rose during this wave. Recommendation accounts were short on this kind of topic.

5. People actually use it. A tweet from the author said "opened four groups today," so several WeChat groups already exist. The Feishu knowledge base has a gallery quick-reference, FAQs, and a section on "startup and monetization experience." The site handraw-style.com has a works plaza, and after generating an image, users are prompted to post their work there, which turns user creations into site content.

Someone built the same idea two months earlier

The most useful comparison is threerocks/hand-drawn-styles. Same idea, also a numbered style-recipe set. It launched at the end of June, more than two months earlier, and now has 2.6k stars. It offers only 5 to 19 styles and has no large gallery.

(GitHub, 11-06.)

This table shows three things. First, the prompt-for-image niche is already crowded, with several repos above ten thousand stars, so readers shouldn't think it's empty ground. Second, others collect by individual case, while this one breaks style out on its own, numbers it, makes it combinable, and reaches over a hundred entries, which makes it the first and largest of its kind in the Chinese community. Third, the same idea with fewer styles, no gallery, and no one promoting it stalls at 2.6k; only with count, gallery, and operations together did it reach 7k. A good idea alone isn't enough.

Copycats already exist: yanghui960/handdraw-style-prompter (created Sep 13, 41 stars, its tagline is "001–261 hand-drawn styles"), and someone on Zhihu renamed and repackaged it. The author is also expanding: the video part was split out as handraw-video, and in Nov 4 a derivative repo, dsh-handraw-style, was opened. So far no second "hundreds of numbered styles" library has taken off.

If it were you

B: The approach is clear and learnable, but the style slot is already taken, and the author is expanding next to it. If you bet on a different dimension, that dimension may not spread as easily as styles do.

Who it suits: people who really know a dimension of taste that's hard to put into words, such as photographers, animation storyboarders, type designers, color graders, or music arrangers; developers willing to spend several weeks batch-generating hundreds of sample images, then keep updating every so often and posting progress on X and in WeChat groups.

Who it doesn't suit: anyone who wants to build yet another style library or another prompt collection (there are 35k-star ones already, and this 7k one); anyone who can't produce sample images and can only write text lists (no image wall, no reposts); anyone unwilling to run operations and who just wants to drop a library online and wait.

How to do it: three ways to follow, best to worst:

1. Shift the dimension: keep the "see the sample, report the number" format but switch to a dimension users are similarly unable to describe. For example, video generation shots and camera moves (dolly, pull, orbit, one-take, with a few seconds of sample footage each), poster and cover typography, music styles in tools like Suno, or PPT layouts. First confirm that the samples for that dimension can form a wall and that differences are visible at a glance. That's the precondition for it spreading.
2. Shift the audience: Japanese and Korean image-generation users have the same problem. Rebuild a set around the styles and platform contexts common there (Instagram, LINE stickers, Naver blog headers) rather than just translating this one.
3. Fill what it does poorly: some have questioned whether "adjacent numbers must look alike." Build a version with fewer styles, where each one is tested on three or four mainstream models and shown side by side, specifically covering "which number works best on which model."

Whichever you choose, copy its packaging: one line naming the pain point, a numbered image wall, a one-line install, and a post every time the count rises.

How to get users: for a cold start, send the project directly to Chinese recommendation accounts like GitHubDaily, write the README so it can be copied in full, and lead with the pain point in the first line. After each batch, post on X and put the new total in the post, giving WeChat and repost accounts a reason to write about you again. Submit to HelloGitHub as well. Set up WeChat groups and a works plaza early so the images users make come back to you.

Timing: the novelty of Skill topics in the Chinese community probably has a few months left, and recommendation accounts will get pickier. Aim to ship before Q1 2027.

Signs you've misjudged it: (1) you get pushed by a major Chinese recommendation account and still don't reach 300 stars in a week, meaning the "numbered gallery" format is no longer fresh; (2) the author or a big account has already taken the dimension you picked (such as camera moves or typography), so don't collide with it; (3) conversely, if over the next month or two several "XX numbered library" projects pass a thousand stars, the format works across many dimensions, so move faster and grab the biggest one first.