Consistent Social Media Cover Agent Skill

Input an article, go through three rounds of Q&A, and it directly outputs cover prompts that can be used for image generation. No need to understand design, no need to write prompts yourself, the aspect ratio is fixed at 3:4, and facial consistency can be maintained.👇 👹 The entire workflow is as follows: 1. Skill installation and first-time configuration (save a photo of your own face + confirm the image generation model) 2. Give the article content to the agent, which automatically reads and refines it 3. Round 1: Recommend composition styles based on content (10 options to choose from) + automatically come up with candidate titles 4. Round 2: Face reference images + materials such as UI screenshots/product images 5. Round 3: Choose expressions, background tones, fonts, and font colors all at once, or leave it all to the model if you're lazy 6. Output complete prompts: 3:4 composition, safety zone, multi-reference image writing for Image 1 (face) + Image 2 (materials) are all included 7. Take it to Jimeng / Nano Banana / GPT-Image to generate images (if your agent has an image generation API configured, you can have it generate directly in the conversation) 8. Check if the titles have typos, and regenerate any parts you're unhappy with. 🧐 Tips: You must choose an image generation model that supports multiple reference images, otherwise facial consistency cannot be guaranteed. The skill has 8 built-in sets of example prompts reverse-engineered from real trending covers, and the model will automatically use them as references, pushing the image detail to the maximum.

一篇文章丢进去，三轮问答，直接吐出能跑图的封面提示词。不用懂设计，不用自己写 prompt，画幅固定 3:4，人脸还能保持一致👇 👹整个工作流程如下: 1、skill 安装与首次配置（存一张自己的正脸照 + 确认生图模型） 2、把文章内容丢给 agent，自动读完提炼 3、第一轮：根据内容推荐构图风格（共 10 种任选）+ 自动想好候选标题 4、第二轮：人脸参考图 + UI 截图/产品图等素材 5、第三轮：表情、背景色调、字体、字色一次选完，懒得选就全交给模型 6、输出完整提示词：3:4 构图、安全区、图1人脸 + 图2素材的多参考图写法全都带上 7、拿去即梦 / Nano Banana / GPT-Image 跑图（如果你的agent配置了生图API可以直接让它在对话中生成） 8、核对标题有没有错字，不满意哪项改哪项重新生成 🧐Tips: 生图模型一定要选支持多参考图的，不然人脸一致性保不住 skill 内置了 8 组按真实爆款封面逆向还原的示例提示词，模型会自动拿它们做参照，出图具体度直接拉满