The guizang illustration skill is entirely markdown, and two of the files it cites are not in the tree

The deliverable is markdown, one YAML file and a prompt template

There is no code in this repository. The primary language is not even recorded, because there is nothing to record one for. What ships is a main instruction file, the readme you are reading, a handoff document holding facts, structure, test cases and verification method, a product document holding positioning, scenarios, boundaries and roadmap, one agent display configuration, one reusable image prompt template, and a set of reference documents. The whole thing runs because an agent reads the instructions and calls whatever image generation it has. The published structure is this:

guizang-material-illustration/
├── SKILL.md                         # Skill 主文件：触发条件、工作流、交付规则
├── README.md                        # 本文件
├── HANDOFF.md                       # 交接文档：事实、结构、测试案例、验证方式
├── PRODUCT.md                       # 产品文档：定位、场景、边界、roadmap
├── agents/
│   └── openai.yaml                  # Codex / OpenAI Skill 展示配置
├── assets/
│   └── prompt-template.md           # 可复用图像提示词模板
└── references/
    ├── visual-style.md              # 歸藏材质插画风格、比例、安全区、主题色
    ├── prompt-patterns.md           # 循环、流程、Hub、对比、层级等提示结构
    ├── chart-beautify.md            # 图表语义抽取、数据优先重画、图标参考

The formatting will tell you it was written by hand rather than generated. The tree ends with a trailing comma after the last entry and a blank line before it closes, and it lists three reference files.

Two reference documents the page tells you to improve are not listed

The workflow text sends you to the reference directory for the visual style, chart and reference-search rules, and the contributing section names two files specifically as places where a useful change would be welcome. One governs the boundaries of reference searching, the other governs the rules for checking in-image text and data. Neither appears in the tree above, and neither is listed as a file you can read. So the two documents the project most wants improved are the two it does not show you, and a reader auditing how the skill checks its own output cannot see the checklist it says it follows. The same gap appears once more at the end of the page, where contributors are told to put tests and demos in a directory that does not exist in the repository listing, with an instruction not to leave one-off output in the skill root. That instruction is sound housekeeping; the directory it names is simply absent.

Every install path is a Claude directory in a project that names several agents

The opening line describes the skill as adapting to agent environments such as Claude Code and Codex, and the trigger phrases cover a wide range of requests. The installation instructions then ignore all of that. There are three documented routes and all three land in the same place. The one-line route runs a skills installer command against the repository URL. The prose route is four numbered steps for an agent with shell access: create the directory if it is missing, clone the repository into it, verify that three specific entries exist, and report back. The manual route is the clone command on its own. In all three the destination is inside a directory named for one specific agent, so a user on a different agent is told to install into a path that agent does not read. The only cross-agent artefact in the repository is a display configuration, and its own comment describes it as presentation metadata rather than anything functional.

Updating means an unpinned pull inside the installed directory

The update story is one paragraph. If the skill is already installed, you send the agent a request to enter the installed directory, run a pull, and report the current commit. There is no version check, no tag comparison and no rollback instruction, and the repository publishes no releases at all, so there is nothing on the remote to pin an install to. The installed skill is therefore a git working copy tracking whatever branch it was cloned from, and the only version identity available is a commit hash the agent reads out loud on request. This is a reasonable trade for a prompt package with no build step, and it is also the shape that produces the usual surprise: a pull that changes the visual style mid-project, with the previous instructions recoverable only from that project's own history.

The agent picks the diagram type, and you never choose a mode

The workflow has eight steps and two of them are about control. Step one reads the material and works out which relationship genuinely needs drawing. Step two decides the type internally, whether that is a flow, a mechanism, a chart, a humanities scene or an educational explanation, and the design principles restate this as a rule: the agent judges the type itself and only asks when something essential is missing. Step four compresses the copy to a single sentence plus three to five short labels, which is the real budget constraint on any diagram. Step six delegates generation to whatever the agent has, naming two image models as examples rather than requiring one. Step seven is the interesting one: if the labels, data, cropping, legends or reference cues are wrong, the instruction is to regenerate, and the readme's own question-and-answer section repeats that advice and explicitly discourages pasting text over the image afterwards as a workaround.

Data accuracy is listed as a principle, ahead of any style rule

The chart path is the most specified part of the skill, and it is specified as extraction rather than restyling. From a screenshot or a set of numbers, the agent is told to pull out the chart type, the title and conclusion, both axes with their units and ticks, the category order, the values with percentages and error bars, and which maximum, minimum, bottleneck or anomaly deserves emphasis. It then produces a new chart that may be smaller, with a small scene beside it, icons, explanatory labels and a visual focus. The explicit instruction is that the goal is not to put new clothes on the original screenshot, and the original layout is deliberately not reproduced. Among the design principles, the one about data comes before the ones about style, and its wording is that data must not be invented: get the values, axes and units right first, then talk about style. That ordering is the right one for a tool that accepts a screenshot as input.

The boundary is teaching material, and publication-grade figures are refused

The unsuitable list is short and specific, and one entry does more work than the others. It rules out full social card layout, pointing you at the companion skill; full presentation structure design, pointing you at a different skill; real photo retouching and portrait photography; long-form poster layout; and scientific figures that need publication-grade data accuracy. That last exclusion sits oddly beside the supported type list, which includes science mechanism diagrams covering force, electricity, magnetism, ecology, biological structure and chemical reaction, and beside a worked example of a lever diagram for a primary school science text with the fulcrum, effort point, resistance point and lever arm labelled. So the line is not science versus humanities. It is explanation for a reader versus a figure for a referee, and the skill is on the right side of it. The humanities claim follows the same logic, with the instruction not to fabricate a real historical scene, using imagery and structure with few labels instead.

Theme colours are named after pigment standards, and the safe area is part of the prompt

The style is described as restrained editorial composition with soft three-dimensional materials, clear spatial relations and a small number of highlight colours, and the default theme colour is named after a specific blue pigment standard, with four alternatives offered as directions rather than as exact values. Because the output is destined for a fixed frame, the prompt is required to state more than subject matter: labels, data, ratio, safe area, visual style and reference cues all have to be specified when the generation call is written. Quality assurance is treated as a delivery step rather than an optional one, checking the in-image text, the data, the cropping, the accuracy of any reference facts and legibility at the size the image will actually be seen. That last check connects to the handoff: the companion card skill takes over titles, body text, theme colour and export ratios, and the readme warns that if the illustration is the point of the card, the outer layout has to give it enough area or the in-image labels stop being readable.

Editorial conclusion

This suits someone publishing in Chinese who needs the one diagram in the middle of a card, deck or article to actually explain something, since the skill refuses to produce wordless decoration and treats short in-image labels as the point rather than a limitation. It is a poor fit for publication-grade scientific figures, for real photography, or for anyone who needs the installed version pinned, because updates are a raw pull in the installed directory and the repository publishes no releases. Before you install it, read the unsuitable list rather than the feature list, and expect the agent to choose the diagram type itself, since not making you pick a mode is an explicit design decision.

Frequently asked questions

Official sources

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