gbro-cover-design: a prompt generator for 3:4 WeChat and Xiaohongshu covers

A prompt writer for 3:4 vertical covers, not an image generator

The README states the tool takes an article, reads it, asks three rounds of questions, and returns one prompt you can run in an image model. The frame is fixed at 3:4 vertical, which is the shape WeChat and Xiaohongshu covers use. Titles are proposed by the skill rather than by you. The people this is aimed at are writers and operators who publish on those two platforms and who can describe an article but cannot write an image prompt. The README says plainly that the skill only produces prompts and needs no API key, so it does not call an image service on your behalf. You still need a separate model to turn the text into a picture.

Three rounds of questions, then ten style templates loaded on demand

The interaction is the whole mechanism. Round one asks for the composition style and the cover title, and the skill recommends a style after reading your article. Round two asks for the face reference image and any extra assets such as product screenshots. Round three lists expression, background tone, typeface and text colour in one pass; anything you leave unset is left to the model. That compression to three rounds is one of the listed changes from oh-my-cover-design, which this repository is based on.

The ten styles are described only by what they suit. The dark gradient style puts the person centred with large type pressed behind them. The flat colour style combines a person, a prop and a plain background. The product hero style is the first choice when you have a UI screenshot or product image, and the screenshot takes the main area. Other entries cover before-and-after cards, minimal white space, poster collage, a person placed to one side, a back view, a partial appearance such as a hand or half a face, and direct eye contact. Each style has a full prompt template under references/style-XX-*.md, and references/examples.md holds eight complete example prompts used as few-shot references. The loading is on demand, so the agent does not pull all ten templates into context for every request.

Installing gbro-cover-design and getting a first prompt

The README gives one clone command, and it places the skill inside the Claude Code skills directory. The repository supports Claude Code, Codex, and any AI agent that accepts custom skills, so the path below is the Claude Code one shown in the documentation.

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

The README adds a warning that matters: the style templates and the example prompt library live in the references/ directory, and SKILL.md on its own is incomplete. Clone the whole repository.

On first trigger the skill walks you through configuration. It asks for a clear frontal face photo saved as assets/my-face.png inside the skill directory, used as the default first reference image, and it asks which image model you use, checking that the model accepts multiple reference images. The README names Jimeng/Seedream 4.0, Nano Banana and GPT-Image as examples. The result is written to config.md in the skill directory so the questions are not repeated.

After that, send your article text to the agent. The skill triggers, asks its three rounds, and prints the prompt. The README's own framing is that you need no design knowledge and do not write the prompt yourself.

Face consistency depends on your image model, not on this skill

The first reference image is your own face photo, and the skill writes the prompt so the facial features stay consistent. That is a text instruction. Whether the output actually resembles you depends on whether the image model reads multiple reference images at once. The README states this directly: if the model does not support multi-reference input, face consistency cannot be maintained. So the failure mode is not in the skill, it is downstream, and it will look like a bad likeness rather than an error message.

There are other boundaries. The frame is fixed at 3:4 vertical, so a square or 16:9 cover is out of scope. The README does not document rollback, versioning or a way to pin a specific prompt revision, and no releases were retrieved for this repository. The last push was on 2026-07-15. The licence field reports NOASSERTION while the README says MIT, which is worth checking against the LICENSE file before you depend on the terms.

How it differs from oh-my-cover-design

This project is based on feitangyuan's oh-my-cover-design, which the README identifies as MIT licensed, and the differences are listed rather than implied. The question flow is compressed to three rounds. The ten style templates are loaded on demand instead of being held in context. A built-in example prompt library supplies few-shot references. The 3:4 frame and a safe-area definition are fixed rather than left open. First-use configuration guidance is new.

The practical difference is where the design decisions sit. The upstream project is the starting point; this one narrows it to a single aspect ratio and a fixed set of compositions, and it adds a setup step that records your face photo and model choice in config.md. If you want a different ratio or your own style set, the narrowing works against you, and the upstream repository is the more sensible base.

Who should adopt it, and what to check first

The fit is narrow and clear. You publish covers on WeChat or Xiaohongshu, you write articles and not prompts, and you already run Claude Code or Codex with custom skills enabled. The ten styles cover the common cover layouts, and the three-round flow keeps the interaction short.

What to verify before you commit: that a full clone put the style-XX files and examples.md under references/, since SKILL.md alone is documented as incomplete; that your image model accepts a second reference image, because the face-consistency step has nothing to work with otherwise; and that the LICENSE file matches the MIT statement in the README, given the NOASSERTION licence field. The repository had no releases retrieved and its last push was on 2026-07-15, so treat the current state as the state you get.

Editorial conclusion

Adopt gbro-cover-design if you publish 3:4 vertical covers for WeChat or Xiaohongshu, you already run Claude Code or Codex, and your image model accepts multiple reference images. Skip it if you need a square or 16:9 cover, or if your model takes only one image, because the face-consistency step has nothing to work with. Before relying on it, clone the whole repository rather than copying SKILL.md alone, confirm that references/ contains the style-XX files and examples.md, and confirm your model accepts a second reference image.

Frequently asked questions

Official sources

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