space-xhs-cover

Xiaohongshu Cover Creation

Turn the user's content into a cover that can be understood at a glance and makes people want to click. By default, use Codex's built-in image_gen, emphasizing title hierarchy, subject recognition, and mobile readability.

This Skill is only responsible for the cover or cover variants:

The body image set of a single note, use space-xhs-image
For image-and-text content that needs to be verbatim accurate, editable, and easy to screenshot in a browser, use xhs-html

Before starting

Must read when creating:

If you need to call the built-in image generation tool, you must also fully read the system imagegen Skill, and follow its invocation, editing, and file-saving rules.

Default settings

When the user does not specify, use the following settings directly, and do not repeatedly ask about conventional options:

Aspect ratio: 3:4 vertical
Output: 1080 × 1440 PNG
Quantity: 1 image
Style: intelligently matched according to the theme
Copy: 1 main title, optional 1 subtitle
Main title: 6–14 Chinese characters, 1–3 lines
Subtitle: no more than 22 Chinese characters
Visual order: main title → subject → subtitle
Whitespace: keep at least about 20% breathing space in the image
Margins: important content at least about 10% from all four edges

When the user says "just do it", "you decide", or "match it to the content", make a reasonable choice directly and generate, without pausing to wait.

Input checklist

Prioritize extracting from the user's existing content:

Theme or body text
Main title; if absent, extract it from the content
Subtitle; can be omitted
Subject: person, product, food, space, or abstract theme
Subject image: optional
Background image: optional, can be multiple
Style or style reference: optional
Font preference, color scheme, decoration, and free-form notes: optional
Number of variants: default 1, can make 3 or 6 images
Account name: add only when the user explicitly requests it

Ask only when information that would change the essence of the task is missing. When there is no image, generate from scratch based on the description; when there is no clear style, match a style first and then create.

Workflow

1. Understand the content

First determine which category this note belongs to:

Tutorial checklist
Experience review
Opinion judgment
Comparison review
Tool recommendation
Workplace or professional knowledge
Product introduction
Lifestyle
Store visits and travel
Personal brand

Extract three core pieces of information:

What the user most wants readers to remember
Which word can best create recognition or interest
What object is most suitable to become the visual subject

2. Organize the title

If the user gives a clear title, keep the original meaning unless the user authorizes optimization.

If the user provides only the body text, offer 3 short title candidates and recommend 1. Titles should be specific and tangible, prioritizing numbers, objects, results, scenarios, or contrasts; do not stack exaggerated words, and do not use empty slogans.

Title line breaks follow:

One line carries only one semantic block
Keywords occupy their own line or use an accent color
Avoid single-character orphan lines
When mixing Chinese and English, keep complete English phrases
Use as little punctuation as possible; do not write long body text on the cover

3. Match style

From the original style library, recommend 4 directions suitable for the current topic, and explain the differences in one sentence each. If the user has specified a style or requested direct production, skip the selection.

Keep the title and subject consistent across the same batch of variants; you may change:

Composition
Color scheme
Font character
Decoration system
Background treatment

Do not count changing just one color as a new variant.

4. Mark image roles

Clearly label the role of each input image:

主体图: Person or product identity must be preserved
背景图: Used for environment, texture, atmosphere, or location
风格参考: Extract only the visual language; do not copy logos, text, or unique assets

Local images must first be checked with view_image before being passed to image_gen。

All target images have local paths: use referenced_image_paths
Target images exist only in the current conversation: use the minimal num_last_images_to_include
The two methods cannot be used at the same time
No input images: omit both

Subject fidelity requirements:

Preserve the person's identity, facial features, hairstyle, clothing, and pose unless the user requests changes
Preserve the product's shape, packaging, logo position, and key structure
The title must not cover the face, hands, product name, or key selling points
The background serves the subject and does not create another visual center

5. Design composition

Prioritize choosing from the following compositions:

Text left, person right: suitable for personal opinions, workplace, and personal branding
Text top, object bottom: suitable for products, food, and spaces
Central subject: suitable for a strong person or a single product
Split-screen comparison: suitable for before-and-after changes and option comparisons
Large background text: suitable for opinion and trend topics
Checklist cards: suitable for steps, tools, and knowledge summaries
Banner labels: suitable for shop visits, travel, and event topics
Collage combination: suitable for recaps, collections, and multi-scene content

Keep only one main composition per cover. Use at most 3 decoration clusters, and avoid cramming the information density of a body page into the cover.

6. Generate cover

Use the prompt recipe to assemble the prompt, and specify clearly:

Use case: ads-marketing
3:4 vertical Xiaohongshu cover
List the text that needs to appear word for word
The text's number of lines, position, hierarchy, and emphasized words
The subject and the purpose of each reference image
Composition, color scheme, font character, and decoration
It must still be clear in a mobile thumbnail
Do not add watermarks, account names, brands, or unrelated text

One image_gen call generates only one variant. When 3 or 6 images are needed, call one by one and write clear differences for each.

7. Check and revise

After each image is generated, a visual check is mandatory:

Whether the main title is correct word for word
Whether there is garbled text, repeated characters, typos, or extra text
Whether it can still be understood within 2 seconds when scaled down
Whether the main title is the first visual
Whether faces, hands, products, and key identifiers are intact
Whether there is enough white space
Whether fonts, colors, and decorations are consistent
Whether there is a safe margin around all sides

If only a localized issue exists, make a targeted edit; do not tear down and redo the entire image. For example:

Only correct the second line of the main title, accurately changing “财务人必备的 Agent” to “财务人的 Agent”, and keep all other people, composition, colors, and decorations unchanged.

When the same text still cannot be generated accurately after two consecutive attempts:

Generate a text-free base image and reserve a clear title area
Tell the user that the text will be overlaid in an editable way
Switch to xhs-html to complete word-for-word accurate text layout

Do not deliver an image with text errors as a finished product.

8. Saving and Delivery

Within the project, save by default to:

09 frames/xhs-cover/YYYY-MM-DD-主题/
├── cover-01.png
├── cover-02.png
└── prompt.md

When only one image is generated, still name it cover-01.png. If a file with the same name already exists, use -v2, and do not overwrite silently.

If the generated image result is in a temporary or system-generated directory, it must be copied to the project directory before delivery. Use:

python3 scripts/check_cover.py path/to/cover-01.png --thumbnail

Check the aspect ratio, dimensions, and file size, and generate a thumbnail. Only adjust to 1080 × 1440 when safe cropping will not damage the title and main subject; prioritize getting the correct aspect ratio during image generation.

Final delivery includes:

The absolute path and clickable link of the finished image
Actual dimensions
The style used
A one-sentence difference between variants
If there are limitations, state them directly

Content and Visual Boundaries

Do not automatically add account names, watermarks, QR codes, or platform identifiers that the user did not request
Do not copy brand signatures, person identities, exclusive characters, or protected elements from reference images
Reference images are only used to extract general layout, color scheme, materials, and font character
Do not fabricate efficacy, benefits, rankings, or professional endorsements
For topics such as finance and healthcare, do not stuff disclaimers into the cover; when needed, remind the user to add them in the body text
Do not download or package images, fonts, and code from external pages

Quick Example

The user says:

Help me make a cover for “5 Agents Every Finance Person Must Have”, white background, orange highlights, with my portrait photo.

Execution:

Check the portrait photo and mark it as the main subject image
Recommend “Professional White Space” or “List Big Text”
Split the main title into “财务人必备 / 5 个 Agent”
Generate a 3:4 cover with a left-text, right-person composition
Check the title, person identity, and thumbnail readability
Save as 1080 × 1440 PNG