瀑布流特定画风采图员 (The Vibe Harvester)

自动向下滚动瀑布流网站（如小红书、Pinterest），根据用户设定的审美偏好进行视觉筛选，并自动下载符合要求的参考图。

Security Scan

The skill's stated goal (automatically scroll, visually filter, and download images) is plausible, but its runtime instructions expect GUI automation, a vision model, and explicit anti-bot/captcha handling while the package declares no dependencies, binaries, or credentials — these mismatches and the captcha-bypass behavior are concerning.

Purpose & Capability

The declared purpose (scanning waterfall sites and saving matching images) matches the single declared write permission (~/Downloads/Vibe_Harvester). However the SKILL.md explicitly expects GUI/browser automation (mouse movement, clicks, opening large previews, right-click save) and a 'visual large model' for aesthetic filtering. The skill declares no required binaries, automation tools, or model credentials — an incoherence between what it says it will do and what it requests/install/depends on.

Instruction Scope

Instructions go beyond simple browsing: they direct simulated mouse/scroll behavior, opening/closing pages, interacting with UI elements, and explicitly instruct how to handle '滑动验证码' (slide captchas), login popups, and disabled right-clicks. That is scope-creep: it includes steps that attempt to evade anti-bot measures and interact with authentication UI, which are not justified by the declared metadata and raise legal/ethical concerns.

Install Mechanism

No install spec is provided, yet the runtime behavior described would normally require browser automation tooling (Selenium/Playwright), a browser extension, or a platform-provided UI-driving capability, plus a vision model runtime or API. The absence of any declared install or dependency means it's unclear how the skill will actually perform those actions — this mismatch is a deployment risk and a red flag.

Credentials

At face value the single write permission to a downloads folder is proportionate. But the SKILL.md relies on a '视觉大模型' without declaring any API keys, endpoints, or local model requirements, and instructs interacting with login prompts and captchas (which could lead to credential exposure or unintended clicks). The lack of declared credentials or config for the vision model is inconsistent and suspicious.

Persistence & Privilege

always:false (no forced persistence) which is appropriate. Autonomous invocation is allowed by default; combined with the instructions to autonomously interact with arbitrary websites and bypass anti-bot measures, this increases potential impact. No indication the skill modifies other skills or system settings.

What to consider before installing

This skill instructs an agent to perform automated GUI/browser interactions and to evade anti-bot measures (slide captcha handling, disabled right-click workarounds). Before installing, confirm: (1) how will the skill drive the browser and where does the visual model run (local model, platform capability, or external API)? (2) whether any additional binaries, browser extensions, or API keys are required — ask the author for a dependency list and an install plan; (3) that you are permitted to download the target sites' images (terms-of-service and copyright); and (4) that you are comfortable granting an autonomous agent the ability to interact with websites on your behalf. Avoid installing if you cannot verify how it performs captcha bypasses or if doing so would violate site terms or laws. If you proceed, test in a controlled environment and limit its scope (use a throwaway account and a small download folder).

🎯 核心目标

你需要充当用户的“美术助理”。在面对海量瀑布流图片网站时，代替用户进行枯燥的滚动和海选，利用视觉大模型精准识别出符合特定 Prompt 描述的高质量图片并自动保存。

💡 触发条件

用户打开目标瀑布流网站后，输入类似指令触发：

• “帮我收集极简主义、黑白配色的桌面布置图，存到素材库。”
• “往下滚，看到符合‘赛博朋克风格 UI’的图就帮我下载。”

📋 执行步骤

第一步：接收审美 Prompt

解析用户触发指令中的核心视觉要求（例如主体物、色彩搭配、氛围感、特定材质等），将其作为后续视觉判断的唯一标准。

第二步：拟人化浏览与筛选（核心循环）

1. 视觉扫描： 审视当前屏幕上显示的完整图片卡片。
2. 审美过滤： 针对每一张完整的图，快速进行判断：“这张图是否高度符合用户的审美 Prompt？”
   • 如果符合 -> 执行第三步。
   • 如果不符合 -> 忽略。
3. 拟人化滚动： 模拟鼠标滚轮向下滚动一定距离（避免机械式的匀速翻页被网站识别为爬虫，应加入随机停顿时间）。

第三步：精准下载与规避反爬

当发现符合要求的图片时：

1. 展开大图： 模拟鼠标移动到该图片上方并点击，进入大图预览模式（或新标签页）。
2. 视觉识别保存： 如果页面右键被禁用，识别页面上的“下载”或“保存”按钮进行点击；如果可以右键，模拟 右键 -> 图片另存为。
3. 命名规范： 将图片命名为 [视觉特征描述]_[随机序号].jpg，存入 ~/Downloads/Vibe_Harvester/。
4. 返回列表： 关闭大图或退回上一页，继续执行第二步的浏览循环。

第四步：异常处理

如果屏幕上突然出现“滑动验证码”、“登录弹窗”或“是否继续浏览”的提示，立刻使用视觉识别弹窗位置，并模拟鼠标进行点击关闭或滑动操作，恢复浏览状态。

⚠️ 安全与操作红线

1. 专注任务： 严禁点击任何带有“广告 (Ad)”、“推广”字样的图片或链接，即使它的画风符合要求。
2. 克制下载： 遇到模棱两可的图片一律跳过，宁缺毋滥，避免塞满用户的硬盘。