wechat-cover

WeChat Official Account cover image creation

📋 Overview

Based on viral article data collected daily across the web, obtain the visual elements of viral covers in the same niche, use AI to analyze and summarize high-conversion visual patterns, and generate cover design proposals that fit the article content and match the platform's traffic aesthetics (including cases + image generation prompts). Execute only in the main Agent; do not dispatch to sub-Agents.

🎯 Trigger Rules

Scenario	Handling Method
User provides content topic/niche	Execute task flow
User requests generation of a cover proposal	Execute task flow
User uploads copy or images for analysis	Execute task flow
User asks about skill features/usage	Introduce skill only
User does not provide a clear creative topic	Introduce skill only

📝 Terminology Standards

Uniformly use "viral cover"; prohibited: explosive cover→viral cover, explosive data→viral data, scrape data→obtain/query data, scrape covers→obtain covers.

🔒 Data Sources

All viral data must come from the interface; web search, use of other data sources, or fabricated data are prohibited. When data is insufficient, clearly inform the user.

🔑 Authentication

Get API Key

Please go to RedFox Hub to get the API KEY

Configure API Key

Option 1: Taking OpenClaw as an example, add REDFOX_API_KEY to ~/.openclaw/openclaw.json:

{ "env": { "REDFOX_API_KEY": "ak_xxxx..." } }

Option 2: Terminal configuration

export REDFOX_API_KEY="ak_xxxx..."

⚙️ Prerequisites

• Dependencies:requests（pip install requests）
• Credentials: environment variable REDFOX_API_KEY(see the Authentication section for details)
• Data range: includes only data from yesterday to 30 days ago; defaults to querying the last 30 days

🔄 Operation Flow

Step 1: Intent Parsing

Structurally parse the user's needs and extract 4 dimensions:

Dimension	Description	Example
Content topic	Prioritize domain terms	Skincare, fashion, food, workplace
Content type	Content format	Practical tips, reviews, recommendations, tutorials
Style preference	Visual/tone	Professional, click-attracting, emotional, minimalist
Core keywords	Search keywords	Autumn/winter, whitening, skincare, dry skin

Extract 2-5 core keywords (comma-separated), prioritize niche domain terms, and supplement with scenario terms based on style preference.

Step 2: Viral Data Query

# 有赛道关键词
python3 scripts/fetch_explosive_covers.py --keyword <关键词1,关键词2,关键词3>
# 无赛道关键词
python3 scripts/fetch_explosive_covers.py --keyword ""

• At most 5 keywords, total length ≤200 characters
• Pass keywords into the interface all at once; do not split them into looped calls
• Time parameters are not passed by default (the interface decides the default range on its own); when the user specifies "the last N days", startDate = today - (N-1) days

Step 3: Intelligent matching and filtering

Data volume	Processing
≥20 items	Use directly
<20 items	Expand keywords (e.g.: whitening→whitening, skincare, whitening and freckle removal), query again, merge data

Filtering rule: sort by read count from high to low, take the top 20. Keyword expansion at most 1-2 times, with 2-3 similar words each time.

Step 4: AI image analysis and feature extraction

1. Parse the JSON data returned by the script (including the articles article array, as well as the latestHotArticles、hotTopics、relatedSearches auxiliary data)
2. Extract the imageUrl field and merge into the cover candidate pool——Use the original value directly, without any modification(the new interface field name is imageUrl, the old interface is coverUrl, the script has already handled compatibility)
3. Sort by interaction volume from high to low and take the top 20
4. Analyze cover images one by one(use native image understanding capability, directly access imageUrl)

7-dimension analysis：

Dimension	Analysis content
Text content	Whether there is text, content, font size, position, color, prominence
Main subject elements	Main visual elements such as people/products/scenes/patterns, etc.
Color characteristics	Main color tone, color scheme, contrast, warm/cool tones
Composition and layout	Image structure, element distribution, subject position, use of blank space
Visual style	Style types such as real-shot/illustration/collage/minimalist/flashy, etc.
Click appeal	Pain point/curiosity/aesthetics/practicality——why it attracts clicks
Size detection	Width x height, whether it is qualified (≥10px)

Image filtering rules：

• Blank images (solid color blocks, no content): mark as "needs filtering", exclude from all subsequent steps
• Small-size images (width or height <10px): filter all out
• Second check before output: ensure no blank/small-size images are mixed in

Classification summary: group covers of similar styles into the same category, and summarize feature patterns (filter blank images before classification).

Step 5: Output HTML analysis report and design plan

Use references/report_template.html(JS-rendered template), only inject JSON data, and it is forbidden to temporarily generate complete HTML at output time.

1. Copy the template to the output file:cp references/report_template.html ./爆款封面分析报告_{关键词}.html
2. Assemble the analysis results into a JSON object (for the format see references/report_template.md）
3. Use SearchReplace in theoutput file(not the original template file) to replace the __REPORT_DATA__ default empty value with the actual JSON
4. Execute open ./爆款封面分析报告_{关键词}.html open in the browser

JSON structure(only fill in the following fields, no need to generate any HTML):

{
  "keyword": "用户关键词",
  "analysisCount": 20,
  "styles": [{ "name": "", "count": 0, "coreVisual": "", "features": "", "covers": [] }],
  "plans": [{ "name": "", "coreVisual": "", "case": { "imageUrl": "", "url": "", "title": "", "author": "", "reads": "" }, "prompt": "" }]
}

Requirement	Description
Template	references/report_template.html(JS-rendered, styles/anti-hotlinking are built in)
Data format	See references/report_template.md JSON data structure section
Cover display	Classified by style type, up to 5 per category
Case reference	imageUrl original value + title (url link) + author + read count only
Image generation prompt	Strictly follow the 2.35:1 landscape aspect ratio

Step 6: User selects a plan and generates the cover image

⚠️ Ask the user first, and wait for their reply before generating—do not generate the cover image before the user makes a selection!

Phase 1(Execute immediately after outputting the report): The end of the HTML template already contains the plan selection area; stop after displaying the report and wait for the user's reply.

Phase 2(Execute after the user replies):

1. Confirm the plan (1/2/3) selected by the user and its image generation prompt
2. If the user uploaded an image, analyze it using native image understanding capabilities
3. Use native image generation capabilities to generate the cover image (2.35:1, 900x383 pixels)
4. Display the image directly, no HTML wrapping needed

Step 7: Self-check before output

Check item	Content
Style type	Each category has: name, frequency, core visual, key features, cover example
Three plans	Each has a style name, core visual, case reference (non-duplicate), image generation prompt

• Cover image links use the original value of the imageUrl field | Image filtering | No blank images, no small-size images | | User interaction | Plan selection has been asked and a reply is awaited |

⚠️ Notes

• Cover image container ratio 2.35:1 (900x383), use object-fit: cover + object-position: center adaptive cropping
• Add anti-hotlinking meta tag to HTML <meta name="referrer" content="no-referrer">
• Case references keep only read count (clicksCount), remove likes and recommendations
• Cover image uses <a> tag wrapped to be clickable and jump to the large image; plan titles use the url field link
• Image analysis uses native image understanding capabilities, directly accessing cover URLs, analyzing one by one