gbro-cover-design asks three questions, then hands you a prompt to run elsewhere

The deliverable is a prompt you run somewhere else

The first line of the page settles the scope. You hand the agent an article, it reads the content, asks you three rounds of questions, and then outputs a cover prompt you can take straight to an image model. Titles are suggested for you, the composition logic is already built in, and the frame is fixed at 3:4 vertical. No image is generated here, which is why there is no API key to configure: the page states the first-run setup needs none, because the skill only produces prompts. That also fixes where the responsibility sits. Layout decisions, style choice and the title are the skill's output; the model you run afterwards decides what the rest looks like, and you decide whether the prompt is any good. For anyone whose posting pipeline already has a separate image step, that split is the feature. For anyone expecting the skill to produce a file, there is nothing to collect.

Three rounds, and the last one delegates whatever you skip

The questioning is compressed into three rounds, and that compression is one of the stated differences from the project it is based on. Round one is composition style, offered with a recommendation derived from the article, plus the cover title. Round two is the face reference photo and any extra materials such as product screenshots. Round three lists expression, background tone, font and text colour all at once, with the instruction that anything you do not mention is left to the model. That last clause is the one to notice, because it is the difference between a skill that asks and one that asks and then accepts an answer. There is no default palette, no house font and no list of blocked combinations in the visible page, so the visual consistency of a series comes from what you type in round three and from the model, not from the templates. Three rounds is also short enough that the interaction does not turn into an interview.

First run writes two things into the skill directory

The setup step is two items and both land inside the cloned folder rather than in a settings file elsewhere. The first is a face reference: prepare a clear front-facing photo and save it as `assets/my-face.png` in the skill directory, which becomes the default image 1, or choose to upload a new one each time instead. The second is a check on your image model: it has to accept multiple reference images, with Seedream 4.0, Nano Banana and GPT-Image named as models that do. Without that capability the face cannot be kept consistent, so the page frames it as a precondition rather than a preference. The answers are recorded in `config.md` in the skill directory and are not asked again. That is convenient and it has a location to remember: the configuration and the face photo live next to the code you cloned, so re-cloning or deleting the skill directory resets both, and a face photo is not the sort of file you want in three repositories at once.

Ten styles, and image 1 is always your face

The style list is ten entries, each with a stated use, which is more useful than a gallery. Dark gradient puts the person in the centre with the large text behind it and is described as the strongest impact. Solid flat is a clean person plus prop against a plain background. Product hero is the first choice when there is a UI screenshot or product shot, with the screenshot as the subject. Contrast cards are for before and after or good and bad content. Minimal whitespace leaves the title as the only focus. Poster collage is for when there are many materials, layering them for depth. Person offset puts the figure to one side and gives the other side entirely to the title. Back view, for motivational and inspirational content, turns the person away from the camera. Partial appearance shows only hands or half a face so the product is the protagonist. Front-facing goes for direct eye contact. Reference images have a fixed order: image 1 is your own face, where the skill is expected to keep the facial features consistent, and image 2 onwards is any material you want on the cover.

A single file is not the skill

Installation is a clone into the skills directory:

git clone https://github.com/pyang5166/gbro-cover-design.git \
  ~/.claude/skills/gbro-cover-design

The warning under it is the part worth reading. The style templates and the example prompt library live in the `references/` directory, SKILL.md by itself is incomplete, so the clone has to be complete. That matches the file layout at the top level, which is `.gitignore`, `LICENSE`, `README.md`, `SKILL.md`, `assets/` and `references/`. Inside `references/` each style has its own full prompt template named `style-XX-*.md`, one file per style, and `examples.md` holds eight complete example prompts to work from. So the prompt templates are loaded on demand rather than inlined into SKILL.md, which keeps the always-loaded file small and means a partial copy produces an agent that asks all the right questions and then has nothing to answer them with. The page says the skill triggers automatically once the article is sent to the agent.

A fork with its changes listed, and a licence the page cannot state

This is an improvement on another project rather than a new one, and the page says which: it is based on oh-my-cover-design, credited to feitangyuan and licensed MIT. The differences it claims are specific enough to check against the original. The questions were compressed to three rounds. The ten style templates are loaded on demand. A built-in example prompt library provides the few-shot material. The 3:4 frame and a safe-area definition are fixed rather than chosen per post. First-use configuration guidance was added, which is the two-step setup described earlier. The licence section repeats that this is MIT, derived from that project, with the original copyright information kept in the LICENSE file. What it cannot say is what the repository page itself reports, because the licence field on the page reads NOASSERTION. So the in-tree statement and the page-level record disagree, and the file to rely on is the one in the clone.

The results table is empty and the README has no English version

Two things about the page itself are worth a reader's attention. The results section is a table of eight cells in two rows, and every cell is empty: the examples were images, and nothing else is left in their place. So the visual promise of ten styles rests on the prose descriptions of each one rather than on anything you can look at before installing. The documentation is Chinese, and the top-level listing contains only `README.md`, so there is no English version beside it even though the skill targets agents that work in either language. The repository's primary language field is empty as well, which fits a page that is mostly prose and templates. Two smaller signals: the project is scoped to WeChat official accounts and Xiaohongshu covers rather than generic art direction, and its audience is stated as anyone whose agent supports custom skills, with Claude Code and Codex named, so the host matters less than the ability to load a skill file.

Editorial conclusion

gbro-cover-design fits someone publishing several posts a week on a Chinese account who wants the framing decided for them and the wording left to a model. It does not fit anyone expecting a finished image, a title they can publish unchanged, or a workflow that runs unattended, since the questions are the product and you supply the model. Four things to check before the first run: whether your image model accepts multiple reference images, because that single requirement decides whether a face can stay consistent at all; whether you are comfortable with the other fields being left to the model, since anything the third round does not name is delegated rather than defaulted; where the configuration is stored, since the answers land in `config.md` inside the skill directory and your face photo in `assets/`, both of which sit next to a cloned copy of someone else's repository; and whether the clone is complete, because SKILL.md on its own is not the skill. The repository is written in Chinese with no English version in the tree, it is an MIT fork of oh-my-cover-design with a named list of changes, the licence recorded on the repository page is NOASSERTION, and the last commit is dated July 15, 2026.

Frequently asked questions

Official sources

1. Issues
2. pyang5166/gbro-cover-design on GitHub
3. README