guizang-material-illustration: an agent skill for explainer images with Chinese labels inside the picture

The gap it targets is the middle image, not the cover and not the layout

Most content pipelines already have a cover and a layout. What they lack is the one image in the centre of a document that has to carry meaning: a diagram of a mechanism, a labelled process, a rebuilt chart. The README frames this precisely, describing the problem as the middle image that social cards, slides, articles and documents keep needing, and contrasting it with a decorative image that looks good but communicates nothing. The project is scoped to three jobs: explainer diagrams that turn abstract concepts, flows and system relationships into labelled pictures; chart beautification that extracts semantics from a screenshot or raw data and regenerates a more shareable chart; and reference-assisted generation for obscure concepts, brands, models, scientific apparatus and historical objects, where the skill looks up background first and then converts everything into one consistent material style. It is the companion to guizang-social-card-skill, with that project owning the title, body text, theme colour and dimensions of the whole card while this one owns the illustration inside it. The audience is narrow and identifiable: people writing Chinese-language content who need diagrams, not photographers, not poster designers.

The skill is a routing layer plus a prompt template, and the file tree is the whole architecture

There is no runtime here. The repository is a set of Markdown instructions an agent reads. The layout given in the README is a single SKILL.md holding trigger conditions, workflow and delivery rules, with the visual system and quality rules split across references/visual-style.md, references/prompt-patterns.md, references/chart-beautify.md, references/use-cases-and-routing.md, references/reference-gathering.md and references/qa-checklist.md. A reusable image prompt template sits in assets/prompt-template.md, and agents/openai.yaml carries the Codex and OpenAI skill display configuration. The documented workflow runs in eight steps: understand the material and find the relationship that actually needs drawing; judge the diagram type internally rather than making the user pick a mode; gather reference information when the subject is obscure; compress the copy into one sentence plus three to five short labels; write the generation prompt covering labels, data, aspect ratio, safe area, visual style and reference cues; call GPT-Image, imagegen, or whatever image capability the current agent has; inspect and regenerate if Chinese labels, data, cropping, legend or reference cues are wrong; then save the image path and the prompt. The design principle that matters most is the second one: the image is allowed to contain text. Explainer diagrams are not treated as decoration, so short labels, arrows, legends and data annotations are supposed to be rendered inside the picture rather than overlaid afterwards.

Chart beautification extracts semantics and discards the original layout

The chart path is the most concrete mechanism in the material. Input can be a screenshot or a bare set of data. The agent first extracts the chart type, the title and conclusion, the horizontal and vertical axes with units and ticks, the category order, the values, percentages and error bars, and the highest value, lowest value, bottleneck or anomaly that needs emphasis. It then regenerates the chart in a form suited to distribution, where the chart itself may be smaller and surrounded by small scenes, icons, explanatory labels and a visual focal point. The README is explicit that the goal is not reskinning the original screenshot but helping the reader understand it faster. That distinction has a practical consequence: because the agent is told to preserve values, axes and units and to keep the category order, a beautified chart can be checked against the source. It also means the output is a redrawing, so it will not be pixel-identical to the input, and anyone who needs the original raster for an audit trail should keep it. The supported chart list is broad on paper: bar, line, Gantt, Sankey, heatmap, funnel and cumulative flow. The roadmap admits the gaps, listing radar, matrix, swimlane, timeline, org chart and map-based data as still to come, so treat that list as aspirational rather than shipped.

Reference gathering is for factual accuracy, not for borrowing a style

The reference step is easy to misread, and the README pre-empts the misreading: reference search is not for finding an art style and not for copying external images. It answers three questions only, namely what the thing is, which structures, components, processes or icons must not be drawn incorrectly, and what visual cue lets an audience recognise it at a glance. The examples given are telling in their range: PKCE, Andon, Zettelkasten, the Kirkpatrick model, the Panopticon, a model logo, a scientific instrument, a historical object. These are subjects where a plausible-looking but wrong diagram is worse than no diagram, because the reader will trust the labels. The workflow places this step before prompt writing, and the roadmap notes that a template for recording what was searched, what was extracted and which cues were rejected is still pending. That is a real gap: without a recorded trail, you cannot audit why a diagram shows a particular component, and the skill currently gives you no artefact for that. The FAQ also states plainly that the reference material is used to understand facts and visual cues, not to copy.

Installation is one npx command or one git clone into the skills directory

Three installation routes are documented. The short one is `npx skills add https://github.com/op7418/guizang-material-illustration --skill guizang-material-illustration`. The manual one clones into the Claude skills directory: `git clone https://github.com/op7418/guizang-material-illustration.git ~/.claude/skills/guizang-material-illustration`. The third route hands the job to an agent with shell access, instructing it to create `~/.claude/skills/` if missing, run the clone, then verify with `ls ~/.claude/skills/guizang-material-illustration/` and confirm that `SKILL.md`, `assets/` and `references/` are present. Updating is a `git pull` inside the same directory, and the README suggests asking the agent to report the current commit afterwards. Triggers are natural-language phrases in Chinese, such as asking for a labelled explainer image, a beautified chart, or a material-style illustration for a report, plus a phrase that explicitly asks the agent to search for reference information first when the concept is obscure. Because the repository ships no releases, the update path is the commit history on main, and the last push recorded is 2026-07-07. There is no versioning scheme to pin against, so a `git pull` can change behaviour under you.

The failure modes are label corruption, unreadable downscaling, and the wrong category of work

The README's own FAQ concedes the central risk: if the Chinese characters come out wrong, the recommended fix is to shorten the label and regenerate, and it advises against patching text onto the image with HTML. That is an honest answer and also a limited one. Shortening labels trades information for legibility, and regeneration is a reroll, not a correction, so a diagram with six labels may need several attempts before all six render correctly. There is no stated mechanism for deterministic text placement. The second failure mode is scale. The README warns that when the output goes into a social card, the outer card must give the image enough room, because shrinking the centre image makes the in-image labels unreadable. That makes the skill's output sensitive to a decision made in a different tool. The third is category error, and the README lists it directly: full social card layout belongs to the Social Card Skill, full slide structure to a PPT skill, real photo retouching, portrait photography and long-form poster layout are out of scope, and research figures requiring strict publication-grade data plotting are explicitly not the target. Anyone who needs a chart that will be reviewed by a journal should stop reading here.

Against a general-purpose image model, the difference is the rules file and the QA pass

The obvious alternative is to skip the skill and prompt an image model directly, or to draw diagrams in a tool like Mermaid or Excalidraw. The difference in approach is not the image model, since the skill calls GPT-Image or imagegen or whatever the host agent offers, which is the same model you would call yourself. The difference is everything wrapped around that call: an internal router that picks the diagram type without asking, a prompt pattern library covering loops, flows, hubs, comparisons and hierarchies, a chart extraction checklist that names axes, units, error bars and category order, a reference-gathering rule with three explicit questions, and a QA checklist for in-image text, data, cropping and reference accuracy. A hand-rolled prompt has none of that and will drift between runs. Against a diagram-as-code tool such as Mermaid, the trade is the inverse: Mermaid gives you reproducible, diffable, text-exact labels and no visual styling worth showing to a reader, while this skill gives you a styled illustration whose text is probabilistic. If your labels must be exactly right and reproducible, code-generated diagrams win. If the image has to look designed and carry short labels, this is the category the project is built for.

Maintenance cost, licence status, and what the repository does not tell you

The maintenance burden is low in the ordinary sense: there is nothing to build, no dependencies to patch, no service to run, and updating is a `git pull` in `~/.claude/skills/guizang-material-illustration`. The burden that does exist is prompt drift. Because the skill is a set of Markdown instructions consumed by a model, a change to the underlying image model or to the agent's tool routing can alter results without any commit in this repository, and the absence of releases means there is no changelog to read when output quality shifts. The material supplied does not state a licence for this repository, so the terms under which you may reuse or redistribute SKILL.md, the reference files or the prompt template are unconfirmed. Treat that as unresolved rather than permissive, and check the repository's licence file before shipping the skill inside a commercial product or bundling it into a client deliverable. This is a description of what the material shows, not legal advice. Two other things the material does not settle: which specific image models are known to render Chinese labels reliably, and how many regeneration attempts a typical diagram needs. Neither is quantified anywhere in the README, the FAQ or the roadmap, so any claim about hit rate would be invention.

Editorial conclusion

Adopt it if you produce Chinese-language articles, teaching material or report decks and already run Claude Code or Codex with an image generation tool wired in, because the skill's whole value is the routing rules in SKILL.md and the six files under references/. Do not adopt it if you need publication-grade statistical charts, portrait retouching, or a full card layout, since the README explicitly routes those elsewhere. Before trusting it, open a generated explainer at the size it will actually appear and confirm the Chinese labels are still legible, then check references/qa-checklist.md against the output.

Official sources

1. Issues
2. op7418/guizang-material-illustration on GitHub
3. README