An agent skill that writes a house style down precisely enough to reuse

A house style packaged as an agent skill

This project is an agent skill that produces editorial illustrations in a particular flat, hand-drawn style: a full-bleed coloured background, an irregular ivory shape carrying the composition, and thick near-black lines drawn over the top. You give it a theme and it returns an image.

The interesting move is what sits between the theme and the picture. The skill's stated job is to turn an abstract subject into a single clear visual metaphor before drawing anything, which is the step most people skip when prompting an image model directly. Asking for an illustration about attention management produces something vague; deciding first that the metaphor is two hands sheltering a small flame produces a picture.

The README is explicit about provenance in a disclaimer near the top. The project describes itself as independently made and unofficial, with no affiliation, authorisation or partnership with the company whose style it emulates, and says it learned observable regularities from public pages rather than reproducing specific published compositions.

The intended uses are named: news illustrations, blog headers, social cards, video thumbnails and conceptual images.

Three layers and eight colours, specified exactly

What makes this more than a prompt fragment is that the style is written down as a specification rather than described in adjectives.

Every image is three layers. An opaque accent colour covers the entire canvas, corner to corner. An irregular ivory carrier shape occupies roughly fifty-five to eighty percent of the frame. Near-black lines sit on top.

The palette is given as eight hex values with names: a near-black for lines, an ivory for the carrier shape, and six accent colours. Stating the exact values rather than describing them as soft or muted is the difference between a style that reproduces consistently and one that drifts every generation.

The line treatment is specified with the same precision: round caps, uneven weight, slight wobble, naive and simplified forms, deliberate mild asymmetry, flat two-dimensional composition with generous whitespace, and legibility when scaled down.

Equally important is the exclusion list. The skill is told to avoid photographic realism, three-dimensional rendering, gradients, shadows and the polished corporate vector look, and to avoid transparent backgrounds and white canvas margins. Negative constraints are what actually hold a generated style in place, because an image model's default pull is toward exactly those things.

Installing it and calling it

Installation is a clone and a copy into the agent's skills directory.

git clone https://github.com/HalfAI1102/anthropic-art.git
cp -R anthropic-art/skill ~/.codex/skills/anthropic-art

The README says to restart the agent or begin a new task afterwards, and offers an alternative for anyone not using git: download the repository archive, copy the skill folder into the agent's skills directory, and rename it to match.

Invocation is explicit rather than automatic. You name the skill in the conversation along with the subject, the aspect ratio and any colour preference, and the README gives several worked examples: a square illustration on collaboration between people and machines, a blog header showing tangled thinking entering a box and a clean line emerging, a globe held in a hand, two hands protecting a small flame, a figure balancing between a sun and a moon.

Those examples are the most useful part of the documentation, because they demonstrate the metaphor-first habit the skill depends on. Each names a concrete scene rather than an abstract topic.

A general prompt template is also included for anyone who wants the rules without the skill, which effectively publishes the specification in a form usable by any image model.

The gap between imitating a style and using it

The disclaimer is appropriately placed and it does not resolve the question a prospective user should ask, which is not about this repository but about what they intend to publish.

A house style is how an organisation is recognised. Producing images in a specific company's visual language and putting them on your own articles, cards and thumbnails invites readers to associate your material with that company, and the more faithfully the skill works the stronger that association becomes. The project's own framing, that it captures observable regularities rather than copying compositions, addresses copying and not recognisability.

That is a judgement for the person publishing rather than a defect in the tool. Personal use, internal decks and experiments raise nothing. Commercial publication under your own brand is where it deserves thought, particularly in a field where the company in question is well known to the audience most likely to see the result.

The repository's licensing makes this sharper rather than softer, because there is no licence file at all. No permission to use, modify or redistribute the skill has been granted, so the reuse question is unresolved on top of the brand question. Anyone considering this commercially has two things to settle and the repository answers neither. This is not legal advice.

Writing your own style specification is the alternative

The alternative is to do exactly what this project did, for your own visual identity: define a layer structure, fix a palette by hex value, specify line treatment, and write the exclusion list.

The difference is whose recognisability you are borrowing. A style specification you write produces images that look like you, become more valuable as they accumulate, and carry no question about association. It costs a designer's afternoon to define and some iteration to tune, which is genuinely more work than copying a folder.

This project's value in that light is as a worked example. It shows the level of specificity that makes a generated style hold: not descriptive adjectives but three named layers, eight hex values, five line properties and an explicit list of what to refuse. Anyone building a house style for generated imagery could follow that structure with their own values and get a comparable result without the association question.

Take the skill as it stands for personal work, for experiments, or where imitating that particular look is genuinely the point. Take its structure, and replace its palette, for anything you intend to publish as your own.

What the repository contains, and its state

The repository is small: the documentation, an examples directory with generated images, and the skill folder itself. There are no releases and the last push was on 2026-07-19.

The examples matter more than the file count suggests, since a style specification is only credible if you can see what it produces, and three sample images sit in the README table.

The documentation is written in Chinese throughout, which a reader should expect when cloning, though the operative content, the hex values, the layer proportions and the prompt template, transfers regardless of the reader's language.

Before using it, two checks in order. Decide what you will publish with the output, since the brand association question is about your usage rather than about the tool. Then note the absence of a licence file, which leaves reuse ungranted, and treat the general prompt template as the portable part if you conclude you want the technique rather than the specific look.

Editorial conclusion

This skill is worth installing for personal work or experiments where that particular flat editorial look is what you want, and its specification is worth reading by anyone building a consistent generated style of their own, because the precision is what makes such a style hold. Think carefully before publishing the output under your own brand, since the project emulates an identifiable company's visual language and the README's disclaimer addresses copying rather than recognisability. Note also that the repository contains no licence file, so reuse is ungranted, and treat the included general prompt template as the portable part if you decide you want the technique rather than the specific look.

Frequently asked questions

Official sources

1. HalfAI1102/anthropic-art on GitHub
2. Issues
3. README