orange2ai/orange-line-illustration: a style skill with no program behind it

There is no program here, only a folder of instructions

The repository is a skill package, not a library. The tree holds a main specification file, five reference documents, a directory of example images, a license file and the readme, with no source code in any language, which is why the primary language is recorded as unknown. Nothing compiles, nothing runs, and there is no entry point.

What SKILL.md contains, per the front page, is the style rules, the metaphor method and the prompt templates. The references split those out: one document defines the default character, one holds that character's image prompt template, two more define the other two characters, and a fifth documents the slide deck workflow.

That means the actual image generation is delegated entirely. No image model is named anywhere in the visible documentation, no API is called, and no parameter is pinned. What you install is a description of a style, and whatever agent and image model you already have does the drawing.

The practical consequence is that results are not reproducible across setups. Two people who both follow the instructions can produce visibly different pictures, because the constraints are expressed as prose guidance rather than as a program with fixed inputs.

Three agents are named, and only two get an install path

The stated requirement is any AI agent that supports the SKILL.md convention, with Cola, Claude Code and Codex given as examples. The install step then supplies two paths:

# Cola
~/.cola/skills/orange-line-illustration/

# Claude Code
~/.claude/skills/orange-line-illustration/

Codex is named in the same sentence as the other two and then never mentioned again. There is no directory path for it, and no note about where that agent expects skills. If you are on Codex you have to work out the location yourself before anything works.

There is no package manager step, no lockfile and no version selector, because there is nothing to resolve. Installation is a directory copy. Updating is a directory copy.

That last point has a cost. The repository has no GitHub releases at all, so there is no tagged version to pin against and no changelog to diff. The last push landed on 2026-06-15 and the repository is not archived, but with a copy-based workflow a user who updated six months ago has no way to tell whether their copy matches the current specification other than reading it.

The repository's own directory listing omits its English readme

The front page is written in Chinese and carries a language switch at the top pointing to an English version. That file is listed among the top-level entries in the repository, and the working examples directory contains seventeen PNG files.

The directory tree printed in the documentation itself does not match. It shows SKILL.md, the references directory with its five named files, an examples directory, LICENSE.md and README.md. README.en.md is absent from that listing, even though it exists in the repository and is the link target of the language switch at the top of the page.

So the structure a reader is given as the canonical map of the project omits one of the two files that exist. The same block also does not mention a .gitignore, a changelog or any workflow directory, which is consistent with a repository whose only tracked assets are prose and images.

Worth stating plainly: the front page text reproduced in this repository listing is the Chinese version only. The English file is named as present in the tree and as the language-switch target, and its contents are not set out here.

No standard license identifier, and a dual licensing arrangement

The license metadata records no standard identifier, which is consistent with what the documentation actually says. The terms are described as dual: open source and personal use are free, while closed source or proprietary use requires purchasing a commercial license, with the details in LICENSE.md.

That arrangement has a practical consequence for adoption. The illustration style itself, the character designs and the prompt templates are the kind of asset a commercial product might want to reproduce consistently, and under these terms that use is precisely the case routed to a paid license. Whether a given deployment counts as open source is the question the split turns on, and it is not answered anywhere in the visible documentation.

No version of the license text is reproduced in the front page, only the summary of the split. The repository metadata declining to name an SPDX identifier is the honest signal here rather than an inconsistency: this is not MIT, not Apache and not Creative Commons, and treating it as a standard open-source license would be a mistake.

For anyone evaluating this beyond personal article illustration, the file to read is LICENSE.md, not the summary.

The scale rule is the style's signature, not a suggestion

One rule carries more of the look than any other: figures are always extremely small and objects are always enormous. The documentation calls that imbalance the signature drama of the style, and it is stated as an invariant rather than a tendency.

The rest of the quick reference is a short list of constraints. Lines are thin black ink with a hand-drawn wobble so they do not read as mechanical. The background is pure white, with no paper texture, no gradient and no shadow. There is exactly one accent color, a warm orange at `#F97316`, and it lands on the single most meaningful element in the frame rather than being sprinkled. Pull-quotes appear as small grey text in the bottom right corner, quiet and legible without competing with the drawing. The intended temperament is named as witty, restrained and intelligent.

The strictness of the accent rule is what makes the design system work. One color, one element, no exceptions, which is why the default character can carry that same hex value on its chest without competing with the scene around it.

Three reusable characters, each with a fixed stroke budget

The character system is presented as three reusable identities with fixed shape definitions, and the constraint stated on them is that they must carry the central action rather than serve as decoration.

The default is Xiao Cheng, a geometric figure drawn in pure black line with an orange dot on its chest, described as quietly working without announcement. Its construction is fully specified: a round head outline left unfilled, two black dot eyes and no mouth, a narrow rectangular body outline, and thin line limbs. The chest dot is the same `#F97316` as the global accent, so the character doubles as the palette's anchor.

Xianren is the minimal figure, a round head, a single arc for a body and single-line limbs, described as the fewest strokes that still read as someone being present.

Xianmao is the kitten, budgeted at ten to twelve strokes: unclosed arcs, two black dot eyes and two ear strokes, relying on the viewer to complete the shape. The documentation frames that reliance as the point.

The stroke budgets matter because they bound how much an agent can improvise. A shape specified to twelve strokes cannot become a detailed animal no matter how the prompt is worded.

Both workflows end with a person choosing images

There are two workflows, and they differ in output rather than in philosophy.

Workflow A produces a single concept illustration at sixteen by nine for a judgment, a principle or a metaphor inside an article. Its stated sequence is: extract the pull-quote, design a concrete scene, generate candidates, and a person picks. The trigger phrase is short.

Workflow B turns an article or an outline into HTML slides with one scene illustration per page, using the file references/ppt-workflow.md for detail. Its sequence is longer: read the article, write a slide outline, confirm, generate illustrations in bulk, have the user select images, then output the HTML.

Both sequences place a human decision near the end, and both have an explicit confirmation step before the expensive phase. Workflow B asks for confirmation of the outline, workflow A does not ask for confirmation of the scene.

The shipped demo follows workflow A at volume. Twenty illustrations were generated for a 75,000-character internal Alibaba article called 置身钉内, each paired with one pull-quote from the text. Seventeen example images ship in the repository, so the rest of that set is visible only in the linked public article.

Seventeen shipped images against a demo that claims twenty

The examples directory holds seventeen PNG files, and their names map back to the captions used in the documentation. Several are named after the pull-quotes they illustrate, covering a sender's position, a flag overloaded with too much, a user who is simultaneously real and hypothetical, a constant holding firm inside a storm, context inequality, a greedy scale, an inbox overload, technical debt in old systems, a trophy that weighs, a balloon tethered to the ground, an empty table, and three hundred days of continuous flight.

Two more are named for the secondary characters, one for the kitten and one for the figure at an old city wall. Three are not part of that set: a cover image, an amplifier and a subtraction, which read as generic style samples rather than demo output.

So the shipped set is a mix: some images demonstrate the characters, some demonstrate the general style, and the remainder of the twenty-illustration demo is not in the repository at all. The full set is published as a public article behind a WeChat link, which is Chinese-language and outside the repository.

For a style defined largely by proportion, accent placement and stroke economy, the shipped images are the only calibration material available, and seventeen is what you get.

Editorial conclusion

Use it if you write articles or decks and want one recognizable illustration style applied consistently, and if you are willing to select the output yourself each time. Skip it if you need reproducible generation, a pinned model, or a licensing story that is settled, because the terms here route closed-source use to a paid license and the repository has no releases to version against. Before adopting it commercially, read LICENSE.md rather than trusting the free-for-open-source framing.

Frequently asked questions

Official sources

1. Issues
2. orange2ai/orange-line-illustration on GitHub
3. README