Giving AI a “Style Menu”: My Hands-on Experience with a Hand-drawn Style Skill

From installing a skill and reworking my autumn photographs to designing posters for a public-interest talk at EGO Bookstore, I wanted to find out more than whether AI could generate images: could I express my intentions more clearly and judge whether it had fulfilled them?

At first, I simply asked AI to help me find a hand-drawn style skill said to offer “278 visual styles.” What we found was the open-source project handraw-style. The repository really does spell its name handraw, and I retain that spelling here; its main prompt entry point is called handdraw-style-prompter.

Once it was installed, I faced hundreds of styles and a large collection of layout patterns. My first questions were simple: what do they actually look like, and how should I use them?

So we kept going beyond the “installation successful” message. We began with “Autumn Reading,” redrew my own photographs as illustrations, created posters for a bookstore talk, and finally tried combining two styles.

This article is a brief account of that practice. It is a record of using the skill, with finished work, shortcomings, and reflections, rather than an advertisement for becoming a designer at the click of a button or a review of every feature.

1. First, Understand What You Installed: It Is Not an Image-generation Model

A Skill is more like a design menu accompanied by a method for consulting it and organizing prompts. A separately connected image-generation model is what actually turns words into pixels.

They have different roles: the skill helps me describe what I want, and the image model tries to draw it. Installing the skill does not also give me a model, an account, a usage allowance, or access to paid services.

The indexes checked in the version installed for this project contained 366 visual styles, 170 layout patterns, and 36 theme colors. These figures describe the version downloaded for this exercise; they are not fixed forever. Some repository documentation still gives older figures such as 365, so version differences matter.

These three kinds of menu should also be kept distinct:

Style determines how an image is drawn: for example, pen hatching, colored-pencil paper texture, or cut-paper collage.

Layout determines how elements are arranged: for example, text above an image, a comparative structure, or information interspersed with open space.

Theme color determines the palette that leads the composition. It does not require every object to use a single color.

2. Installation and Browsing: Look at the Images Before Discussing the Codes

For this exercise, the original author’s repository was installed in WorkBuddy’s user-level skills directory. We kept the full directory structure, including entry files, scripts, indexes, galleries, and reference images, rather than copying only a set of written instructions.

An installation request to the assistant could be phrased as follows:

Please install this skill from https://github.com/yang0/handraw-style.

First check its source and dependencies, and keep the complete resource directory. If a version with the same name already exists, explain that before proceeding; do not overwrite it directly.

After installation, check that the resources are complete and test whether it can output prompts.

This describes what we did in WorkBuddy for this exercise; it is not a universal installation command for every client. Other tools need to follow their own skill-installation procedures. Running the included scripts also involves runtime requirements such as Python. Browsing the galleries and actually calling the scripts do not have the same dependency requirements.

After installation, we checked file completeness and tested prompt output for “style + palette” and “layout + style.” Only after invoking the skill in the conversation and generating images did we further confirm that the workflow was usable. “The files have been copied,” “the assistant can invoke the skill,” and “the model can generate an image” are three things that should be checked separately.

Viewing the samples is much easier: open the three local galleries in the installation directory:

skills/handdraw-style-prompter/gallery/index.html Styles

skills/handdraw-style-prompter/gallery/layouts.html Layouts

skills/handdraw-style-prompter/gallery/colors.html Colors

These paths are relative to the root of the installation package. You can also consult STYLES.md, LAYOUTS.md, and COLORS.md in the original repository. Start by choosing three to five looks you like; there is no need to memorize hundreds of codes.

The codes are catalog references, not a universal language understood by image models. The assistant first looks up a code, expands it into a specific style name, layout description, and color requirements, then passes these to the image model. Sometimes a style reference image is needed too. When moving to another image-generation tool, take the complete prompt and any required images with you, rather than entering only a few codes.

3. The First “Autumn Reading” Image: Keep Asking Questions After It Appears

My first clearly specified attempt was:

Use handraw-style.

Style: FB-006; layout: SC-001; theme color: C-01.

Theme: Autumn Reading. Output only the prompt first; do not generate an image.

Here, FB-006 corresponds to Bendy Absurd Storybook, SC-001 to a social media card with text above the image, and C-01 to Klein Blue. Generating the prompt first lets me see exactly what has been selected, avoiding image generation before the choices have been confirmed.

Next, I asked it to draw from that prompt. The result was a text-and-image card led by blue: “Autumn Reading” at the top, a rabbit reading in an autumn forest below, and book pages, birds, and a little boat.

After seeing the image, I asked another question: I did not upload a sample, so how did you know what this style looked like?

The answer needs to be clear: this image was based mainly on text expanded after consulting the library, rather than drawn while looking at the library’s sample. A style name can activate visual associations the model already has, but that does not establish that the style has been reproduced accurately.

This led me to a habit: ask not only which style was used, but also which materials were actually supplied. Written descriptions, style samples, layout samples, and color swatches are different levels of input. A requirement appearing in a prompt does not mean that the result has met it.

4. Redrawing Photographs: Separate Content References from Style References

For the second step, I used my own photograph of autumn trees and a building and asked for a redrawing in the library’s FB-027 style. This time, two images were actually supplied: the original photograph and the corresponding style sample from the library.

A content reference answers “what to draw”; a style reference answers “how to draw it.” The photograph’s upward-looking viewpoint, slanting tree trunks, autumn leaves, and building on the right provided the content. The sample’s pen hatching, ink shadows, paper ground, and palette provided the means of expression.

This distinction matters. Otherwise, the model may bring in the sample’s people, clothing, and setting as well, turning the photographic redrawing into a story borrowed from another picture. For this exercise, I asked it to extract only the sample’s visual language, without copying its subjects.

Interestingly, I had initially thought of it as a “classic storybook” style, but the specific catalog entry turned out to be Edwardian Vintage Ink Fashion Editorial. An impression of a category is not the same as an entry’s definition. The actual sample found in the library is more reliable than imagining a style from its name.

What I wanted was more than a layer of paper texture placed over a photograph. I wanted the structure of the original to be extracted and its lines, color shapes, and materials reorganized. This is style transfer or redrawing. Even a prompt that stresses “redraw from scratch,” however, cannot guarantee that the model will not retain too much photographic texture. The finished work still needs to be examined.

5. The Bookstore Posters: Make Real Details Memorable

The next task was more specific: create a poster for the public-interest talk “In the Age of AI, What Is Worth Asking Further?” at Yigou EGO Bookstore.

The details most worth keeping came from the bookstore itself, rather than generic robots, chips, or glowing brains: a black cat, a wicker chair, a green cushion, a full-height wall of books, reading steps, a pale blue boat, and a cargo bicycle. These came from the two photographs I supplied. An additional video could not be read successfully, so it cannot be counted among the materials used.

After a colored-pencil attempt, we developed two more directions: FA-003 pen and light watercolor, which felt lighter and quieter; and FF-005 cut-paper collage, with layered paper and more distinct shapes. They used the same event information but created different visual impressions.

View the full-size pen-and-watercolor poster (PNG) · View the full-size cut-paper collage poster (PNG)

For an event poster, one thing matters even more than visual appeal: the information must be usable. We checked the time, address, and WeChat QR code against the attachments and clarified “5W” as “5Why” on that basis.

The QR code was extracted from an attachment and composited afterward, rather than redrawn by the model. Both the original-size result and a version reduced to 900 pixels wide were successfully decoded, with contents matching the original code. But successful software decoding does not establish that the WeChat add-contact process has been completed, nor that a printed copy will necessarily scan well. Those last two checks have not yet been performed and should still be carried out before official use.

This suggests a practical approach: let the image model handle illustration and, when necessary, use post-production compositing for critical items such as names, times, addresses, and QR codes. A complete poster does not have to be generated in a single pass.

6. Advanced Style Fusion: More Parameters Do Not Guarantee a Better Result

I then tried a more complex set of requirements: SC-014 whitespace-led layout, FB-012 for the character style, FB-034 for the setting, C-34 Cream White, a vertical 9:16 format, and a mood of “energy and humor.”

Style fusion does not mean blending every element in equal measure. The idea here was to place a minimalist character with half-closed eyes and deadpan humor in a bookstore setting that was detailed in some areas and simply drawn in others. The visual languages would differ while sharing coherent space and lighting.

After generation and layout, the event information was complete, and the cream-white palette and open space were reasonably clear, but both the setting and the character remained rather realistic. The cat had not fully become the minimalist, deadpan character I had imagined, and the contrast between the two visual languages was not pronounced enough.

View the full-size style-fusion draft (PNG)

This does not mean that every requirement was ineffective. It reminds me that piling up more style names does not automatically give more precise control. When a model is pulled toward the realism of the original photograph, it may be better to check the character design or the main visual on its own before making the full poster, rather than rerunning the entire long prompt every time.

Layout requirements must also be coordinated with the density of the copy. There is a real trade-off between “extensive whitespace” and “a complete biography, lecture outline, and long address, all in large type.” A phrase such as “integrated text and image” cannot make that spatial constraint disappear.

7. Do Not Mistake a Change for a Requirement Met

In another attempt, I paired an autumn photograph with an abstract painting as a reference, hoping that the leaves and building would dissolve into layers of blue-green and magenta, with rubbed marks, glazing, and dissolving forms. This time, the reference was an abstract painting I supplied, rather than another numbered sample from the library.

The result did change the colors, but it retained clearly defined leaves, branches, and window grids. It was closer to a painterly treatment of a photograph and still fell short of the reference painting’s degree of abstraction.

Through these exercises, I gradually separated evaluation into four questions:

Is the content right? Have the key identifying elements of the photograph or place been retained?

Has the style been achieved? Do the lines, materials, and organization of forms match the intention, rather than merely changing the colors?

Is the information accurate? Are there errors in names, numbers, addresses, or wording?

Is the file usable? Do its proportions, clarity, QR code, and intended context of use fit together?

A Canvas File is useful for displaying work together in a workspace. For everyday viewing and sharing, I still prefer saving PNG images. Being able to open a file is part of delivery quality too. Likewise, a clear digital preview cannot replace a physical print proof; dimensions and bleed requirements still need to be confirmed with the producer.

8. If You Want to Try It, Start with These Three Requests

If you are unfamiliar with the codes, first ask the assistant for recommendations:

Use handraw-style to recommend three distinctly different visual styles for my reading event.

Explain the mood each would suit. Propose the options first; do not generate images.

Once you see a style you like, specify the task:

I choose this style. Please use my photographs to create an event poster.

Use the photographs as content references. Use the style sample only for lines, brushwork, and materials.

First confirm the actual codes and images you will use, then generate one image so we can see the result.

When you receive the result, go beyond “make it look nicer” and request changes that can be checked:

Keep the positions of the bookstore’s wall of books and black cat.

Simplify the cat further, reduce the fur detail, and emphasize the humor of its half-closed eyes.

Keep the verified text and the real QR code; do not regenerate them.

What attracts me most to this resource library is the language it provides for turning vague aesthetic preferences into choices we can discuss, rather than the number of styles. I may not know the precise professional term for a particular kind of brushwork, but I can look at samples, choose a direction, compare results, and keep asking questions.

In this process, AI helps me expand the possibilities of expression. But deciding what to retain, where a result falls short, and whether to use it still requires human judgment. Aesthetic judgment and a human in the loop are part of creation itself, rather than extra steps added after generation.

As images become easier to obtain, perhaps what deserves practice is precisely this: express our intentions clearly, look carefully at the results, and know when one more question is needed.

About the Author

Tree Ye | Professor Quwei Ye

An education researcher, professor, and AI instructional system designer. He previously served as vice president of Peking University Founder Education Group and president of Xi’an Technology and Business College. He holds a master’s degree in artificial intelligence and expert systems from the National University of Defense Technology and was a visiting scholar at Connecticut State University in the United States.

His longstanding interests include AI questioning power, learning innovation, and personal growth. He developed the “Six-Cycle Questioning Framework” (Record · Reflect · Question · Act · Review · Transform) and hopes to bring ideas from educational research into conversations about reading and everyday life.

Contact

Website: www.treeye.cn

WeChat: tree_ye1961

Email: shuye_1961@qq.com

Sources and Notes

1. Open-source author: yang0. Original project: https://github.com/yang0/handraw-style . Installation methods, resource categories, and dependency notes refer to README.md, DEPENDENCIES.md, and the skill documentation in the package installed for this exercise.

2. Visual catalogs: STYLES.md, LAYOUTS.md, and COLORS.md in the original repository. The counts in this article were checked against the local indexes of the version downloaded on October 25, 2026, corresponding to source commit `e15a50eb68718bd541f6d8e0a8307d73246d839d`. They are not a statement about the latest upstream version.

3. This article draws on actual operations carried out on October 25–26, 2026. The photographs were supplied by the author. Figure 1 contains project reference samples; the remaining images were generated with AI or laid out in post-production for this exercise. AI-generation labels in the images have been retained. The evaluations describe observations from this exercise, rather than a uniform conclusion about all models and every style.

4. This article records only functions and results that were verified. Not every style was tested, the underlying image-generation model version was not confirmed, and the skill’s descriptions of model compatibility do not replace hands-on testing for this exercise. No fixed commitment is made about credits or costs.

5. The project’s statement that it is available for use does not mean that all reference images, photographs, fonts, or generated results automatically share the same license. Before public or commercial use, check the relevant terms for each asset and service separately. The posters contain information about an existing event and the bookstore’s QR code; confirm their applicability again before publication.

Prepared by Tree Ye from his own practice, with AI assistance in organizing the material and arranging text and images. Completed on October 27, 2026.

Read the Chinese edition