social-auto-tool-builder-1.1.0

复用“小红书自动回复项目”实战经验，快速构建新的本地AI自动化工具（含多平台选择器映射模板）

Security Scan

The skill's files and runtime instructions are consistent with its stated purpose (building a local Playwright+Ollama social-media auto-responder) and request no unrelated credentials or installs.

Purpose & Capability

Name/description, SKILL.md, and skill.yaml all describe building a local AI + Playwright auto-responder and the instructions only reference Playwright, local Ollama, packaging to EXE, selector templates and dry-run vs send flows — these are coherent and expected for that purpose.

Instruction Scope

SKILL.md explicitly instructs reading DOM/snippets/screenshots to calibrate selectors, running dry-runs, and then performing sends; it does not instruct the agent to read unrelated system files or exfiltrate secrets. It does reference checking a local Ollama HTTP endpoint (127.0.0.1) and using persistent_context for browser login state, which is appropriate for the stated functionality.

Install Mechanism

This is an instruction-only skill with no install spec and no code files. The SKILL.md suggests using common package installs (pip, playwright install) and EXE packaging tools; nothing is pulled from arbitrary URLs in the skill itself.

Credentials

The skill declares no required environment variables, credentials, or config paths. The only runtime network reference is to localhost:11434 (local Ollama), which matches the stated use of a local model. No unrelated tokens or secrets are requested.

Persistence & Privilege

Skill flags are default (not always:true). It does not request persistent platform-wide privileges or instruct modification of other skills or agent-wide configs. It does suggest storing browser login state (persistent_context) which is expected for an automation tool that must act as a logged-in user.

Assessment

This skill appears coherent with its stated goal of producing a local Playwright + Ollama auto-responder. Before using or following its instructions: 1) Understand that the resulting tool will automate actions as your logged-in browser user — test carefully in dry-run mode and on non-production accounts to avoid policy/ToS violations. 2) The skill assumes you run a local Ollama server (127.0.0.1:11434) and have Playwright/browser binaries installed; ensure those are legitimate and up-to-date. 3) The skill itself asks for no secrets, but the tools you build may require login cookies or credentials stored in the browser profile — keep those local and secure. 4) When packaging an EXE, verify the build scripts and any bundled files yourself to avoid accidentally including sensitive data. If you want a higher-assurance review, provide the actual automation code (auto_responder_production.py, build_exe.ps1, requirements.txt) so those files can be inspected for unsafe behavior (remote endpoints, credential capture, obfuscated operations).

社媒自动化工具构建 Skill v1.1.0

目标

把“新自动化工具”的制作过程标准化，并在 v1.0.0 基础上新增多平台适配能力：

• Python + Playwright（persistent_context）
• 本地 Ollama（意图识别 + 回复生成）
• 规则过滤 + AI 生成混合策略
• 先 dry-run 再真实发送
• 参数化运行 + Windows EXE 打包
• 小红书 / 抖音 / 快手选择器映射模板

---

先决输入（必须确认）

1. 目标平台（xiaohongshu / douyin / kuaishou）
2. 入口 URL（通知页/消息页）
3. 已回复判定规则（例如“作者”）
4. 发送成功信号（toast 或状态文本）
5. 时间窗口与轮询间隔
6. 单轮最多处理条数

默认值（信息不全时）：

• interval-minutes=5
• recent-hours=1
• max-replies=0
• 先 --once dry-run

---

多平台选择器映射模板（核心）

在主程序中维护一个映射表（先填模板，后按真实 DOM 覆盖）：

SELECTOR_MAP = {
    "xiaohongshu": {
        "url": "https://www.xiaohongshu.com/notification",
        "comment_item": ".interaction-item, .comment-item",
        "comment_text": ".interaction-content, .comment-text, .content",
        "reply_trigger": "button:has(svg.reply-icon), svg.reply-icon, .reply-btn",
        "input": "p#content-textarea.content-input, [contenteditable='true'].content-input, textarea.comment-input",
        "submit": "button.submit, button:has-text('发送')",
        "success": "text=评论成功",
        "replied_marker": "作者"
    },
    "douyin": {
        "url": "<TO_FILL_DOUYIN_URL>",
        "comment_item": "<TO_FILL>",
        "comment_text": "<TO_FILL>",
        "reply_trigger": "<TO_FILL>",
        "input": "<TO_FILL>",
        "submit": "<TO_FILL>",
        "success": "<TO_FILL_SUCCESS_TEXT>",
        "replied_marker": "<TO_FILL_MARKER>"
    },
    "kuaishou": {
        "url": "<TO_FILL_KUAISHOU_URL>",
        "comment_item": "<TO_FILL>",
        "comment_text": "<TO_FILL>",
        "reply_trigger": "<TO_FILL>",
        "input": "<TO_FILL>",
        "submit": "<TO_FILL>",
        "success": "<TO_FILL_SUCCESS_TEXT>",
        "replied_marker": "<TO_FILL_MARKER>"
    }
}

适配规则：

• 先读平台映射，再执行统一流程。
• 每个关键 selector 至少保留 1 个 fallback。
• 任何平台都先跑 dry-run 列候选，再启用真实发送。

---

标准流程（6 Phase）

Phase 1：骨架搭建

• 主程序拆分：AI 客户端、自动化模块、参数解析。
• 支持参数：
  • --platform
  • --interval-minutes
  • --recent-hours
  • --max-replies
  • --once
  • --interactive

Phase 2：平台映射接入

• 将平台 selector 映射注入 Automation。
• URL、评论列表、输入框、发送按钮、成功信号全部来自映射。

Phase 3：真实 DOM 校准

• 让用户提供截图/DOM 片段，增量修正映射。
• 每次只改一个动作链路（打开弹窗、输入、发送）。

Phase 4：安全与准确

• 已回复过滤（marker）
• 时间窗口过滤（recent-hours）
• 去重（(text, age)）
• 2-5 秒随机延迟

Phase 5：验证

• dry-run：列出候选不发送
• 用户确认后真实发送
• 必须等待成功信号

Phase 6：交付

• EXE 打包（含 Playwright 浏览器路径兼容）
• QUICK_START + FAQ
• Git 提交与推送

---

质量闸门（必须通过）

1. python -m py_compile auto_responder_production.py
2. --once 模式可运行
3. dry-run 候选与人工预期一致
4. 至少 1 条真实发送成功信号
5. EXE 参数模式可运行
6. EXE 交互模式可运行

---

Windows 命令模板

# 安装依赖
pip install -r requirements.txt
python -m playwright install chromium

# 本地模型检查
python -c "import requests;print(requests.get('http://127.0.0.1:11434/api/tags',timeout=5).status_code)"

# 单轮 dry-run
python auto_responder_production.py --platform xiaohongshu --once --recent-hours 1 --max-replies 3

# 构建 EXE
powershell -ExecutionPolicy Bypass -File .\build_exe.ps1

# EXE 参数模式
.\dist\auto_responder.exe --platform xiaohongshu --interval-minutes 5 --recent-hours 2 --max-replies 3

# EXE 交互模式
.\dist\auto_responder.exe --interactive

---

可复用 Prompt（新项目直接用）

按 social-auto-tool-builder v1.1.0 的流程，帮我做一个【平台名】自动化工具。
要求：
1) Python + Playwright + 本地 Ollama
2) 用 persistent_context 保存登录状态
3) 先 dry-run 列候选，再真实发送
4) 支持 --platform --interval-minutes --recent-hours --max-replies --once --interactive
5) 用多平台选择器映射模板实现，允许按我提供的DOM增量修正
6) 最后打包EXE并给 QUICK_START

---

输出物检查清单

• skill.yaml 与 SKILL.md 完整
• 选择器映射模板已建立
• 平台参数化设计明确
• 验证与交付步骤可执行