Orange Line Illustration: a SKILL.md style contract that makes AI agents draw one idea per image

The problem is prompt drift, not image quality

Ask an image model for a minimalist illustration and you get something different each time: a different palette, a different sense of scale, a different density of detail. The failure is consistency across a set. An article with three illustrations that do not look like siblings reads as three unrelated commissions.

Orange Line Illustration attacks that directly. The repository is a style contract an agent reads before generating anything: thin black ink lines with a slight hand-drawn wobble, a pure white background with no paper texture, gradient or shadow, exactly one warm orange accent at #F97316, and a rule that the human figure is always tiny while the object is always large. The README states the intent plainly: one idea, one accent, a lot of silence. The audience is anyone producing editorial images at volume, articles, principle lists, covers, and slide decks, who wants the set to hold together.

The signature detail is the imbalance. In the README's own words, the person is always extremely small and the object is always grand, and that sense of imbalance is described as the style's signature drama. That is a compositional instruction, not a filter, and it is the kind of thing a generic prompt will not preserve on its own.

Three character IPs, each with a fixed anatomy

The skill defines three reusable characters, and the definitions are anatomical rather than descriptive. Xiao Cheng (小橙) is a geometric figure in pure black outline: a circular head with an unfilled outline, two black dot eyes, no mouth, a narrow rectangular body outline, thin line limbs, and a single orange dot on the chest at #F97316. That dot is the entire accent budget for the image.

Xian Ren (线人) is more abstract: a circular head, one curved line for the body, single-line limbs. The README calls it the minimum number of strokes that still says someone is here. Xian Mao (线猫) is a kitten symbol drawn in ten to twelve strokes, an unclosed curve, two dot eyes, two ear strokes, with the expectation that the viewer's brain completes the rest.

The constraint that matters is the one the README states explicitly: these characters are not decoration, they have to carry the core action. A character standing beside the concept instead of performing it is off-spec. That single rule is what separates this from a sticker pack, and it is also the hardest part to enforce, because whether a figure carries the action is a judgement a model cannot make reliably on its own.

Two workflows, and only one of them has a written procedure

The skill covers two paths. Workflow A produces a single 16:9 conceptual illustration for a judgement, principle or metaphor in an article. The README gives the trigger phrase (给这篇文章配三张橙线风格插图) and the sequence: extract the key line, design a concrete scene, generate candidates, and let a human choose. That last step is not optional in the description. The pipeline ends with a person picking.

Workflow B turns an article or outline into an HTML slide deck with one scene illustration per page. The trigger is 把这篇文章做成橙线 PPT, and the documented sequence is longer: read the article, write the slide outline, confirm it, batch-generate illustrations, let the user select, then output the HTML. The README points to references/ppt-workflow.md for the full procedure, so the detail lives in that file rather than in the top-level document.

The asymmetry is worth noting. Workflow A is described in four steps in the README. Workflow B is described in six, with a confirmation gate in the middle. If you are evaluating the slide path, the README alone will not tell you what the outline format looks like or how the selection step is presented. You have to open the reference file.

Installing it means copying a directory into an agent's skill folder

There is no package to install and no build step. The README states the skill works with any AI agent that supports the SKILL.md convention, naming Cola, Claude Code and Codex as examples, and gives two target paths:

~/.cola/skills/orange-line-illustration/ ~/.claude/skills/orange-line-illustration/

You place the directory at one of those locations and then address the agent in natural language. The two documented invocations are the Chinese trigger phrases quoted above. The README does not give an English trigger phrase, so if your agent is configured for English prompts you are inferring the equivalent rather than following a documented one.

The repository layout is the specification: SKILL.md holds the style rules, the metaphor method and the prompt templates; references/ holds one definition file per character (xiao-orange-ip.md, xianren-ip.md, xianmao-ip.md), a prompt template for Xiao Cheng, and the PPT workflow; examples/ holds the sample PNGs. There is no code directory, no dependency manifest and no test suite. That is consistent with what the project is: a written contract plus reference images, not a program.

The human-in-the-loop step is the design, and also the ceiling

Both workflows end with a person choosing among generated candidates. The README does not describe an automated selection step, a scoring mechanism, or a way to regenerate a rejected image without a new instruction. If you need unattended image generation, this skill does not offer it, and the documentation does not suggest it is planned.

The style rules also create a specific failure mode. A single orange accent on a white field works when the image carries one idea. The README's own framing is that one blank illustration may focus attention on a single point in an era when AI infographics and HTML text blocks are everywhere. That is a deliberate restriction. For a diagram with three interacting concepts, or a chart with a legend, the one-accent rule and the tiny-human rule are in tension with the content, and there is no documented fallback for that case.

The third constraint is the language of the material. The README, the style rules and both trigger phrases are in Chinese, with an English version at README.en.md. The example set is drawn from Chinese-language sources: twenty illustrations for a 75,000-character internal Alibaba post, each paired with a quoted line. How well the metaphor method transfers to English-language articles is not something the repository demonstrates.

Against a general-purpose image prompt or a diagramming tool

The obvious alternative is writing your own image prompt per article. The difference in approach is where the constraints live. A per-article prompt encodes the style inside the request, so the style is only as consistent as your memory of it. Orange Line Illustration puts the style in a file the agent reads every time: the palette is a hex value, the background is specified as pure white with no texture, gradient or shadow, and the character anatomy is written down stroke by stroke. That is the actual distinction, a persistent contract versus a repeated instruction.

The other alternative is a diagramming or vector tool, where you place shapes yourself. That gives you exact control and exact reproducibility, and it removes the generation step entirely. What it does not give you is the metaphor: turning a line like the technical debt AI must cross through old systems into a concrete scene. The skill's documented process is extract the key line, design the scene, generate, pick. A drawing tool covers only the last two steps, and you supply the scene.

Neither comparison is a knock. The skill is worth using precisely because it sits between those two: more consistent than ad hoc prompting, less deterministic than placing shapes by hand.

Licence, maintenance and what the repository does not tell you

The README describes a dual licence: free for open-source and personal use, with a commercial licence required for closed-source or proprietary use. The repository's licence metadata reports NOASSERTION, which means GitHub could not classify it automatically. The README directs readers to LICENSE.md for the terms. I have not read that file, so the boundary between free and paid use is not something I can state here. Treat that boundary as the first thing to check, and if your use is commercial and closed-source, treat it as a question for the licence holder rather than for this article.

On maintenance, the material supports only a limited read. The last push is dated 2026-06-15 and no releases were retrieved. The README describes weekend iteration with the goal of making the illustrations more stable, better looking and more interesting, which reads as active but informal development. There is no changelog, no versioned release and no stated compatibility matrix beyond the SKILL.md convention.

Upgrade cost is correspondingly low and correspondingly undefined. Because the skill is a directory of markdown plus example images, updating means replacing files, and any local edits you have made to SKILL.md or the character definitions would need to be reapplied. The repository does not document a migration path or a version number to compare against.

Editorial conclusion

Adopt it if you already write articles or slide decks in Chinese or English and want a repeatable visual grammar rather than one-off image prompts, and if you can live with a skill that is a specification plus examples rather than an installable library. Do not adopt it if you need generated images without a human picking between candidates, or if you need the licence question settled before you ship. Before anything else, read LICENSE.md yourself: the repository metadata reports NOASSERTION, the README describes a dual licence with free open-source and personal use and a paid commercial licence for closed-source or proprietary use, and the README's own pointer to LICENSE.md is the only place the boundary between those two is defined.

Official sources

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