gimi-illustration-skill turns articles into readable, opinionated illustrations

Illustration that explains instead of decorates

Most article illustration is decorative: generate something adjacent to the topic and hope it fits. gimi-illustration-skill, from GiMi-Xiaomi, takes the opposite brief. You paste an article's body text, and the skill extracts the judgments, processes, comparisons, states and metaphors inside it, then draws those as clean, readable images. The README's one-line pitch: readable, with a bit of humor, and never clip-art.

The stated audience is Chinese-language content: long-form articles, Xiaohongshu posts, blogs, Notion pages, spoken-script drafts and workflow write-ups. This is an agent skill rather than an app; it installs into a coding agent and drives that agent's built-in image generation. The output contract is specific, which is what separates it from a prompt template: images land in outputs/ with a shot-config.md recording the illustration strategy for later traceability.

Three styles, selected in plain language

The skill ships three style families, and switching is a sentence, not a settings panel. Quirky hand sketch, the default, is white-background line art with soft blue accents, aimed at knowledge content and methodology. Warm storybook brings watercolor textures for travel and narrative pieces. Product proposal is flat editorial illustration for decision documents and business explanations.

The README demonstrates the same subject drawn in all three, which doubles as calibration: you can see what each style does to a process diagram before choosing. Style directives stack with the other controls: say vertical for portrait output, add title with your text for an in-image heading, or just three for a capped batch. Changing style on an existing piece is a one-line request, with no need to rewrite the requirement.

A character you can swap out

The default illustrations include Gimi, a girl in a horse hat, acting as the explainer. The design rule is worth quoting in spirit: when a character is present, it participates in explaining the thing, it is not a mascot parked in the corner. That constraint keeps the IP from degrading into a sticker.

The character system is pluggable in both directions. Uploading a front-view image and requesting an IP entry registers your own character, after which illustrations follow that persona without re-describing it each time; several characters can coexist. Saying no IP removes characters entirely and produces structure-and-metaphor images only. For teams with their own mascot, that swap is the difference between adopting the skill and redrawing everything it produces.

Installing into Codex or Cursor

Installation copies the repository into the agent's skills directory. For Codex, the README's recommended path:

git clone https://github.com/GiMi-Xiaomi/gimi-illustration-skill.git
cd gimi-illustration-skill
mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills"
cp -R . "${CODEX_HOME:-$HOME/.codex}/skills/gimi-illustration"

For Cursor, the equivalent is a symbolic link into ~/.cursor/skills:

mkdir -p ~/.cursor/skills
ln -s "$(pwd)" ~/.cursor/skills/gimi-illustration

After that, usage is conversational: send the trigger word for illustration plus the article body, with optional style, character, size or title directives. Image generation itself runs on the host platform's built-in tools, Codex Image Gen or Cursor's GenerateImage, so the skill carries no model of its own.

The machinery behind the prompt

The repository is small but structured like a product. SKILL.md is the entry point and workflow. references/ holds the working contracts: style definitions for the three families, composition rules, visual promises, prompt construction, QA checks and an IP protocol with a template for registering custom characters. assets/ carries the Gimi character sheets plus calibration examples for every style, with and without the character, so the model has anchors rather than vibes.

That architecture explains the skill's consistency claims: each image is checked against references before delivery, and the shot-config.md beside every output batch records the strategy. It also tells you the maintenance cost: this is prompt engineering as versioned documents, and improvements land as edits to references, not code releases. There are no releases or tags, and the last push was on 2026-08-03.

Limits to know before adopting

Three boundaries matter. First, the licence is split: the skill's code and document structure are MIT, but the Gimi character image is explicitly excluded and governed by IP-NOTICE.md, so commercial use of the mascot needs a separate look. Second, the skill is built for Chinese content end to end, from its trigger words to its style examples; English articles are outside its design center, even if the mechanics would run.

Third, quality depends on the host platform's image model, since the skill generates nothing itself; the same skill on a weaker generator produces weaker work. The README's own tips concede the edges: partial text yields vague images, over-illustrating turns an article into a picture book, and不满意 pieces should be called out for a redraw rather than regenerated blind.

What to stack it against

The alternatives form a spectrum. Generic image generation with your own prompts is more flexible and has none of the structure: no QA pass, no reproducible style, no character continuity across a set. Diagram-as-code tools like Mermaid or Excalidraw produce precise, maintainable diagrams, but their visual register is engineering documentation, not editorial illustration, and they will not draw a metaphor.

Commissioned illustration sits at the far end: better taste, real licensing, slower and pricier. The skill occupies a defensible middle for one specific job: turning written Chinese reasoning into a consistent, explainable image set at zero marginal cost. If your job is closer to documentation, use Mermaid. If it is closer to brand work, hire someone. If it is content with a deadline, this occupies exactly the gap.

Editorial conclusion

gimi-illustration-skill is worth installing if you publish Chinese long-form or script content regularly and use Codex or Cursor, provided you check IP-NOTICE.md before using the Gimi character commercially. Skip it for English-first content, for precise technical diagrams, or if your platform's image model is weak, since the skill inherits all of its rendering quality. First run: paste a finished article with the default style, no character, and judge the batch against your own published artwork before adopting it into a workflow.

Frequently asked questions

Official sources

1. GiMi-Xiaomi/gimi-illustration-skill on GitHub
2. Issues
3. License: MIT
4. README