Teaching AI Agents to Draw the ‘Middle Image’

The Hype Moment: A Screenshot Worth Packaging

In early July 2026, a single social-media post demonstrated what happens when aesthetic taste meets agentic infrastructure. The user behind the account guizang.ai had been using Claude Code to generate 3:4 images for an article. The result—clean white backgrounds, restrained three-dimensional materials, and crisp Chinese labels—was striking enough to rack up nearly sixty thousand views on X within days [10]. Rather than letting the prompt dissolve into the usual ephemera of model outputs, the author formalized it into a reusable artifact: a Claude Code / Codex Skill named guizang-material-illustration.

This is the repository’s origin story in miniature. It did not emerge from a research lab or a design-systems consortium. It was reverse-engineered from a happy accident and then packaged for repetition. In the current agent-tooling landscape, that trajectory is becoming familiar. A lone practitioner discovers a reliable aesthetic groove inside a large image model, wraps it in structured instructions, and open-sources it as a “Skill”—a directory of markdown files that teach an agent how to behave. The hype is not about a new model architecture. It is about the sudden realization that visual taste, once the province of human designers, can be version-controlled and injected into an agent’s workflow.

What a Skill Actually Is

To understand why this repository matters, one must first understand what it is not. It is not a fine-tuned checkpoint, not a LoRA, not a rendering engine, and not a design tool with a GUI. It contains no trainable weights, no React components, and no installable binary. The entire artifact is a folder of markdown documents—SKILL.md, references/*.md, assets/prompt-template.md—that encode a workflow: understand the material, classify the diagram type, gather references if the concept is obscure, compress the text into short labels, write a generation prompt, call GPT-Image or imagegen, and then QA the result.

In other words, it is glue code written in prose. Its technical payload is a set of constraints and checklists that keep an image-generation agent from drifting into generic decoration. When a user asks for “a mechanism diagram for this product spec,” the Skill intervenes before the model starts drawing, forcing it to decide whether the request is a flowchart, a cycle, a hierarchy, or a scientific mechanism. It then mandates short, object-pointing Chinese labels rather than abstract noun clouds. The value lies entirely in curation and routing logic—the boring parts that usually determine whether an AI-generated image is usable or landfill.

The Aesthetic: Swiss Editorial, Soft 3D, Chinese Labels

The visual system the Skill enforces has a distinct personality. The README describes it as “restrained Swiss editorial composition, soft 3D materials, clear spatial relationships, and a few highlight colors,” defaulting to an IKB blue with extensions into lemon yellow, safety orange, and graphite black. The goal is not photorealism and not flat infographic minimalism. It occupies a middle ground: dimensional enough to feel tactile, clean enough to read at thumbnail size, and structured enough to function as explanation.

This addresses what the project calls the “middle image” problem. Social cards, blog posts, and work reports do not need a masterpiece; they need a central diagram that translates an idea into spatial relationships. The Skill therefore treats illustration as semantic infrastructure. It supports concept breakdowns, flow diagrams, feedback loops, comparative paths, scientific mechanisms, and even humanities metaphors. Crucially, it insists that explanatory images should carry text inside the frame—short Chinese labels, arrows, and data annotations—rather than relying on captions that most readers never read.

That insistence on embedded Chinese text is both a feature and an admission of fragility. Current image models still hallucinate glyphs or render them as gibberish. The Skill mitigates this by enforcing brevity (three to five characters per label) and by including a QA checklist that prioritizes re-generation over post-hoc patching. It is a workaround for a model limitation, elevated into a design constraint.

Two Hard Problems: Charts and References

Beneath the aesthetic wrapper, the Skill distinguishes two workflows that separate it from a simple prompt template.

The first is chart beautification. When the input is a screenshot of an ugly dashboard or a raw data table, the agent is instructed to extract semantic properties—chart type, title, axis units, category order, outliers, and the conclusion the data supports—and then redraw the chart in the material style. The explicit rule is: do not reskin the original layout; reinterpret it. The resulting chart may be smaller on the canvas, surrounded by micro-scenes or iconography that emphasize the takeaway. This is closer to editorial information design than to template-based visualization.

The second is reference gathering. For obscure concepts—PKCE flows, Andon cords, Zettelkasten structures, Panopticon architecture, or specific brand logos—the Skill tells the agent to search for factual and visual references before generating. The reference phase answers three questions only: what is this thing, which structural details must not be wrong, and what visual cues allow instant recognition. After that, the facts are translated into the unified Guizang style. It is a knowledge-to-image pipeline, acknowledging that accurate explanation requires research before rendering.

The Companion Economy

guizang-material-illustration does not pretend to be a full publishing stack. Its README is explicit: it generates the central illustration, not the surrounding layout. For complete social-media cards, it hands off to a sibling project, guizang-social-card-skill, which handles aspect ratios, titles, body copy, and platform-specific sizing [1]. The social-card sibling already carries nearly five thousand stars, suggesting that a small ecosystem of interoperable agent skills is forming around content production.

This modularization is telling. Instead of monolithic “AI design tools” that try to own the entire pipeline, the author has decomposed the work into discrete taste-packages: one for imagery, one for layout. Users are expected to chain them in conversation—generate the center image here, pass it to the social-card skill there, then export for Xiaohongshu or WeChat. It is a vision of creative work as agent orchestration, where human judgment moves upstream into specification and curation while execution is handled by instructed agents.

Limits and Honest Boundaries

The repository’s candor about its boundaries is refreshing. It will not generate full slide decks, strict scientific-publishable figures, photorealistic portraits, or long-form poster typography. It inherits every limitation of the underlying image model. If GPT-Image decides that “Swiss editorial” means something else on a given Tuesday, the Skill cannot override the weights. The QA loop and reference checks are compensatory mechanisms, not guarantees.

Moreover, the project is fundamentally a workflow wrapper. Its innovation is not algorithmic; it is organizational. It encodes a specific cultural-aesthetic preference—Chinese-labeled, soft-3D, editorial diagrams—into a reusable agent instruction set. That is valuable, but it is also fragile. Aesthetic trends in AI-generated content saturate quickly. The name “Material Illustration” inevitably evokes Google’s Material Design [5], a system that became the default visual language for billions of screens. Whether the Guizang style can avoid similar saturation as it propagates through agent workflows is an open question.

The Boring Part Is the Point

The long-term significance of this repository may have little to do with its specific shade of IKB blue. It lies in the precedent: a practitioner discovered a reproducible visual dialect inside a general-purpose model, documented it as a versioned skill, and released it for other agents to adopt. That is a shift from prompt engineering—ephemeral text wrangling—to skill engineering, where domain-specific taste and judgment are maintained as infrastructure.

As agents expand from coding assistants into content operations, marketing, and education, they will need thousands of such taste-packages. guizang-material-illustration is an early, well-documented example of what that layer looks like for visual communication. It is not revolutionary. It is simply the moment when someone decided that the “middle image” deserved a repeatable standard—and taught an agent how to draw it.