← 返回 Skills 市场
e2e5g

蜂兵虾将

作者 e2e5g · GitHub ↗ · v1.4.0
cross-platform ⚠ suspicious
323
总下载
0
收藏
0
当前安装
1
版本数
在 OpenClaw 中安装
/install bingbing-xiajiang
功能描述
蜂兵虾将——你的AI牛马团队,替你干活,帮你赚钱。 全行业热点监控+内容创作+趋势洞察+自动执行,4个AI智能体分工协作: 信息守护者(全网采集)、内容军师(创作策略)、趋势预言家(走势预判)、工作记账本(自动记录)。 适用于:金融、医疗、教育、零售、科技、制造业、餐饮、服务业、汽车、房产等**全行业**。 核心...
使用说明 (SKILL.md)

蜂兵虾将 V1.4

🎯 最终版本:用户自适应 - 学习用户偏好,动态调整交互

V1.4 核心升级

新功能 说明
用户画像 记录用户交互偏好
自适应确认 根据跳过率调整确认频率
个性化输出 根据偏好调整报告风格
预测服务 主动预测用户下一步需求

完整执行流程

用户输入
    │
    ▼
┌─────────────────────────────────────┐
│  查用户画像 ←─────────────────────┐ │
│  • 了解用户偏好                   │ │
│  • 获取历史交互模式               │ │
└─────────────────────────────────────┘ │
    │                                    │
    ▼                                    │
意图识别 → 智能路由 ←───────────────────┘
    │                    (参考用户偏好)
    ▼
主动感知
    │
    ▼
┌─────────────────────────────────────┐
│           模块执行                   │
│  (串行/并行/跳过)                   │
└─────────────────────────────────────┘
    │
    ▼
┌─────────────────────────────────────┐
│           反思机制                   │
│  评估 → 优化 → 重试                │
└─────────────────────────────────────┘
    │
    ▼
记录行为 → 更新画像 ─────────────────→ (回到用户画像)
    │
    ▼
用户确认 → 继续/退出

用户画像详解

学习数据

interface UserProfile {
  user_id: string;
  
  // 交互习惯
  confirmation_habit: {
    total_decisions: number;
    skip_count: number;
    skip_rate: number;          // 跳过率
    avg_decision_time_ms: number;
  };
  
  // 输出偏好
  output_preference: {
    detailed_count: number;
    concise_count: number;
    preferred_style: 'detailed' | 'concise' | 'balanced';
  };
  
  // 推荐接受
  recommendation: {
    total: number;
    accepted: number;
    acceptance_rate: number;
  };
  
  // 执行偏好
  execution: {
    parallel_count: number;
    serial_count: number;
    preferred_mode: 'parallel' | 'serial';
  };
  
  // 模块偏好
  module_preference: {
    module_sequence_history: string[];
    common_paths: string[];
  };
  
  updated_at: string;
}

自适应策略

用户特征 系统调整
跳过率 > 60% 减少确认步骤
跳过率 \x3C 30% 保持完整确认
偏好详细 输出更多解释
偏好精简 输出关键要点
推荐接受 > 70% 多推荐
推荐接受 \x3C 30% 少推荐
偏好并行 优先并行执行
偏好串行 保持串行执行

V1.4 交互示例

完整交互流程

用户:帮我分析新能源汽车行业趋势
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

【行业分析】
识别意图:信息获取 + 内容创作
目标行业:新能源汽车

【用户画像】
┌─────────────────────────────────────┐
│  画像:user_001                     │
│  • 跳过率:75% → 简化确认           │
│  • 输出偏好:详细(80%)           │
│  • 推荐接受:90% → 多推荐           │
│  • 执行偏好:并行                   │
└─────────────────────────────────────┘

【自适应决策】
✓ 减少确认步骤
✓ 输出详细报告
✓✓ 尝试并行 多提供推荐


━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

【执行】模块1 → 模块2(并行)

【模块1 - 反思评估】
✓ 完整性:90% | 质量:8.2/10 | 可用性:92%

【模块2 - 反思评估】
✓ 完整性:88% | 质量:8.0/10 | 可用性:90%

【模块执行完成】

【自适应确认】
✓ 跳过非必要确认(跳过率75%)

摘要:
- 行业趋势:3个
- 创作方案:2套
- 预估时间:25分钟

继续到模块3/4?
1. 继续到模块3
2. 继续到模块4
3. 查看完整报告
4. 结束

请回复:
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

预测性服务

用户:看看金融行业新闻

【行业分析】
识别意图:信息获取
目标行业:金融

【预测服务】
根据您的历史行为:
━━━━━━━━━━━━━━━━━━━━━━━━━━━
• 87% 概率:您会继续到模块2(创作)
• 60% 概率:您会查看详细报告
• 常用路径:模块1 → 模块2
━━━━━━━━━━━━━━━━━━━━━━━━━━━

【预执行】
已在后台准备模块2内容(如果继续)

开始执行模块1...
━━━━━━━━━━━━━━━━━━━━━━━━━━━

注意:以上示例适用于任何行业(金融、医疗、教育、零售、科技、新能源汽车、餐饮等)


全面自检报告

✅ 版本一致性检查

检查项 状态 说明
V1.0 核心保留 串行流程、用户确认、数据传递、灵活退出
V1.1 增量引入 意图识别、智能路由、并行执行
V1.2 增量引入 反思机制、自动重试、优化尝试
V1.3 增量引入 主动感知、增量更新、模式复用
V1.4 增量引入 用户画像、自适应、预测服务

✅ 功能完整性检查

模块 功能 状态
模块1 信息采集+过滤+评估+分级
模块2 趋势分析+爆款分析+创作方案+发布策略
模块3 状态分析+成长洞察+AI信件
模块4 工具推荐+工作流记录+模板生成+效率报告

✅ 记忆系统检查

功能 状态
五层架构(L0-L4)
场景化配置
智能检索
遗忘机制
反馈闭环
增量更新

✅ 逻辑一致性检查

检查项 状态
模块执行顺序(1→2/3/4)
数据传递(后续模块可访问前置输出)
用户确认点(每模块后)
灵活退出(任何时候可退出)
记忆流转(L0→L2→L3→L4)
反思触发(每模块后)
自适应依赖(需要画像数据)

✅ 向后兼容性检查

配置项 默认值 可关闭
intent_recognition true
smart_routing true
parallel_execution true
reflection true
proactive_memory true
user_adaptation true
user_confirmation true ✗ (必须开启)
data_passing true ✗ (必须开启)
flexible_exit true ✗ (必须开启)

完整配置

const MULTI_AGENT_SYSTEM_V1_4 = {
  // 版本
  version: "1.4",
  release_date: "2026-02-25",
  
  // V1.4 功能
  user_adaptation: {
    enabled: true,
    profile_tracking: true,
    adaptive_confirmation: true,
    personalized_output: true,
    predictive_service: true
  },
  
  // V1.3 功能
  proactive_memory: {
    enabled: true,
    incremental_update: true,
    cache_ttl_hours: 168,
    reuse_bonus: 0.2
  },
  
  // V1.2 功能
  reflection: {
    enabled: true,
    auto_retry: true,
    max_retries: 3,
    dimensions: ['completeness', 'quality', 'usability']
  },
  
  // V1.1 功能
  routing: {
    intent_recognition: true,
    smart_routing: true,
    parallel_execution: true,
    patterns: ['serial', 'parallel', 'skip', '精简']
  },
  
  // V1 核心(不可关闭)
  core: {
    user_confirmation: true,
    data_passing: true,
    flexible_exit: true
  },
  
  // 模块配置
  modules: {
    module1: {
      name: "信息守护者",
      layer: "L0",
      retention: "1小时"
    },
    module2: {
      name: "内容趋势优化系统",
      layer: "L2",
      retention: "7天"
    },
    module3: {
      name: "状态洞察模块",
      layer: "L3-L4",
      retention: "90天"
    },
    module4: {
      name: "工作流沉淀系统",
      layer: "L3",
      retention: "永久"
    }
  },
  
  // 记忆系统
  memory: {
    enabled: true,
    layers: ['L0', 'L1', 'L2', 'L3', 'L4'],
    scenarios: ['duty', 'sentiment', 'workflow', 'goal', 'general']
  }
};

执行模式汇总

模式 V1.0 V1.1 V1.2 V1.3 V1.4
串行执行
意图识别 -
智能路由 -
并行执行 -
反思机制 - -
主动感知 - - -
用户自适应 - - - -

参考文档

安全使用建议
This package appears to implement what it claims (web monitoring, content strategy, multi-agent coordination and a persistent memory system), but several things warrant caution: - Incoherence to review: the registry lists no install steps, yet the archive contains package.json, dist/, demo scripts and install.sh; the docs also include an external CDN tarball URL. Before running anything, inspect install.sh and package.json (and any postinstall scripts) to ensure no unexpected downloads or commands run. Prefer running in a disposable container or VM. - Privacy & persistence: the UnifiedMemorySystem writes persistent files under memory/<skillName> (L0-L4, shared, logs). The system stores user profiles, interaction histories and outputs. If you will process sensitive data, disable persistent memory or run with restricted filesystem access and clear the memory directory regularly. - Network behavior: agent prompts instruct scraping many social platforms. The code references use of platform-provided tools (web_search, extract_content_from_websites). Verify whether these rely on your agent platform's network capabilities or whether the package will itself perform HTTP requests. Search the code (install.sh and runtime scripts) for curl/wget/http(s). If the package pulls additional code from the CDN or other remote hosts, treat it as high-risk. - Least privilege: run demo/installation only after auditing code, and if possible, run it in a sandbox/container with no access to production credentials. Ensure no unexpected environment variables or local config files are read/written outside the memory/<skillName> directory. - If you need to keep confidentiality: do not enable persistent memory, or restrict access to the memory directory; disable any proactive/background fetching features until you confirm their behavior. If you want, I can scan the install.sh and package.json (and any scripts referenced there) for network calls, eval/exec usage, or automatic downloads and summarize anything suspicious.
功能分析
Type: OpenClaw Skill Name: bingbing-xiajiang Version: 1.4.0 The skill bundle implements an 'AI Collaboration System' with memory, signal recognition, workflow, and goal tracking. The `SKILL.md` and `AGENT_PROMPTS.md` define the AI's role and tools, focusing on information processing and self-improvement, without any prompt injection attempts to subvert the agent. The `install.sh` script performs standard Node.js project setup, and the JavaScript code primarily handles local file system operations within its designated `memory` directory for persistence. No evidence of data exfiltration, malicious execution, persistence outside the skill's scope, or other harmful behaviors was found across any of the analyzed files.
能力评估
Purpose & Capability
The name/description (multi-agent hotspot monitoring, analysis, content creation, and memory) matches the included prompts, examples, and Node code: agents for collection, cleaning, pattern recognition, decisioning, and a UnifiedMemorySystem that persists L0-L4 memories. There are no requested environment variables or unrelated binaries, so capability alignment is reasonable.
Instruction Scope
The SKILL.md and AGENT_PROMPTS explicitly direct broad web data collection (Weibo, Zhihu, Douyin, B站, 微信公众号, 36kr, 虎嗅, etc.), creation of persistent user profiles, and continual adaptation/prediction. The runtime instructions assume access to web_search and extract_content_from_websites tools (agent platform services) and instruct persistent recording of user interaction history and templates. This collects and stores behavioral and possibly sensitive content; the instructions mandate some global behaviors (user_confirmation/data_passing 'must be on') which increases privacy surface. The skill's instructions also leave wide discretion to 'proactively prepare' content in background — a broad behavior that may run without fine-grained constraints.
Install Mechanism
Registry metadata lists no install spec (instruction-only), but the package includes many runnable Node artifacts (package.json, dist/, demo.js) and an install.sh, plus an external CDN tarball URL in UPGRADE_REPORT.md. That mismatch is an incoherence: the package appears intended to be installed/run, yet the registry declares no install steps. The presence of an external CDN download link in docs is a higher-risk artifact you should review before running any install script (install.sh) or npm install.
Credentials
The skill requests no environment variables or API tokens in metadata. That is plausible if it relies on the platform's web_search/extraction tools, but the system performs broad data collection and persistent profiling. It will create files/directories under memory/<skillName> and persist L1-L4 data (including user behavior and decision history). Even without declared secrets, this creates privacy and data-retention risks: any sensitive input (user messages, industry data, possible credentials captured during conversation) could be stored. There is also an undocumented CDN URL in docs (external host) — not declared as a runtime dependency but present in repo.
Persistence & Privilege
The skill does not set always:true and does not request elevated platform privileges. However, its code (UnifiedMemorySystem) explicitly creates directories and writes persistent files (L0..L4, shared, logs) under a 'memory' base directory. That is expected for a memory system, but it means the skill will have persistent on-disk state that may contain user data. Confirm where that directory is stored and set appropriate filesystem permissions or sandboxing.
如何使用
  1. 确保已安装 OpenClaw(本地或 Docker 部署)
  2. 在对话框中输入安装命令:/install bingbing-xiajiang
  3. 安装完成后,直接呼叫该 Skill 的名称或使用 /bingbing-xiajiang 触发
  4. 根据 Skill 的参数说明提供必要输入,即可获得结构化输出
版本历史
v1.4.0
蜂兵虾将 v1.4.0 - 引入用户自适应系统,自动学习用户偏好,动态调整交互方式和报告风格 - 支持用户画像、跳过率分析、个性化输出、执行模式智能切换 - 新增预测服务,主动预判并准备用户下一步需求 - 强化确认流程:根据用户行为自动减少或保持确认步骤 - 保持全行业支持,流程结构和核心AI能力完整向后兼容
元数据
Slug bingbing-xiajiang
版本 1.4.0
许可证
累计安装 0
当前安装数 0
历史版本数 1
常见问题

蜂兵虾将 是什么?

蜂兵虾将——你的AI牛马团队,替你干活,帮你赚钱。 全行业热点监控+内容创作+趋势洞察+自动执行,4个AI智能体分工协作: 信息守护者(全网采集)、内容军师(创作策略)、趋势预言家(走势预判)、工作记账本(自动记录)。 适用于:金融、医疗、教育、零售、科技、制造业、餐饮、服务业、汽车、房产等**全行业**。 核心... 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件,目前累计下载 323 次。

如何安装 蜂兵虾将?

在 OpenClaw 或 Claude Code 对话框中运行命令「/install bingbing-xiajiang」即可一键安装,无需额外配置。

蜂兵虾将 是免费的吗?

是的,蜂兵虾将 完全免费(开源免费),可自由下载、安装和使用。

蜂兵虾将 支持哪些平台?

蜂兵虾将 跨平台运行,可在任意部署了 OpenClaw / Claude Code 的环境中使用(cross-platform)。

谁开发了 蜂兵虾将?

由 e2e5g(@e2e5g)开发并维护,当前版本 v1.4.0。

💬 留言讨论