← Back to Skills Marketplace
leanerque260

deep-research skill

by wayne260 · GitHub ↗ · v1.0.0 · MIT-0
cross-platform ⚠ suspicious
58
Downloads
0
Stars
0
Active Installs
1
Versions
Install in OpenClaw
/install deep-research-skill-for-business
Description
商业决策级深度调研 Skill V3.1——多步骤编排 + 断点强制校验 + Agent 执行指令。触发条件:用户要求深度调研、行业分析、商业调研、竞品分析等。
README (SKILL.md)

深度调研 Skill V3.1

触发条件

用户在聊天中明确提出调研需求(深度研究、行业分析、商业调研、竞品分析、市场研究等)。

⛔ Agent 入口规则(最先读,最先执行)

⚠️ Agent 收到调研任务后的第一件事不是读 step 文件,而是初始化编排会话。

Agent 必须执行的启动流程

# 1. 初始化会话(必须!)
bash orchestrator.sh "[调研主题]"

# 脚本会:
#   - 创建会话 ID 和状态文件
#   - 打印 s0 步骤内容
#   - 打印 AGENT_INSTRUCTION 块(告诉你在 s0 必须做什么、禁止做什么)
#   - 等待你完成 s0 后用 --complete 推进

Agent 的逐步骤推进规则

规则 说明
必须通过编排脚本推进 每步完成后必须执行 bash orchestrator.sh --complete sX --next sY,不可直接读取下一步文件跳过
不可跨步读文件 你只能知道当前步骤的内容(编排脚本打印什么就看什么),不可提前 read steps/s5.md 等来"预览"
不可跳过 AGENT_INSTRUCTION 每步的 AGENT_INSTRUCTION 块是你的强制执行清单,不是建议
不可跳过断点 s0(范围确认)、s2(搜索前反思)、s9(12项自检)是三个强制断点,任一未完成不可继续

如果编排脚本不可用(兜底)

如因技术原因无法运行 orchestrator.sh(如 bash 不可用 / 文件缺失),则:
1. 明确告知用户:"编排脚本不可用,我将按步骤文件手动执行,但会缺失断点强制校验。"
2. 手动按 s0→s1→...→s9 顺序执行
3. 报告保存后,手动执行 S9 的 12 项自检(见下)
4. 全 ✅ 后才能交付

⛔ 11 条刚性红线(不可违反)

  1. 不可跳步:必须按 s0→s1→s2→...→s9 顺序执行,跳过任一步骤则报告无效
  2. 必出检查表:s0 必须输出 6 要素确认表,s2 必须输出反思校验表,s9 必须输出 12 项自检表
  3. 三源印证:核心数据必须有 3 个及以上独立来源,否则标注"⚠️ 未经充分印证"
  4. 仅标注推测:事实/分析/判断不标注,仅推测内容标注【推测,置信度:X】
  5. 不编造:找不到的信息明确说"该数据暂缺",绝不编造场景或数据
  6. 信源分级:T3 级信源(自媒体/论坛)不得作为核心论点支撑
  7. 纵向叙事体:纵轴必须是故事体(因果链条 + 时代脉络),不是年表流水账
  8. 交汇洞察独立:第 4 章必须是新判断,不是前面内容的缩写版
  9. 缺口诚实:明确指出哪些问题数据不足、哪些结论置信度低
  10. 交付前自检:s9 的 12 项全部打勾才能交付,缺一项都不行。如因故未走编排脚本,报告保存后必须手动执行 S9 12 项自检(bash orchestrator.sh --self-check
  11. 不可绕过编排:Agent 第一步必须 bash orchestrator.sh "主题" 初始化会话,后续每步通过 --complete sX --next sY 推进,不可直接读取 step 文件跳步执行

执行方式

收到调研任务后,必须运行编排脚本来执行:

bash orchestrator.sh "[调研主题]"

编排脚本会在每个步骤:

  1. 打印步骤内容(目标 / 操作 / 校验标准)
  2. 打印 AGENT_INSTRUCTION 块(当前步骤的强制执行清单 + 禁止事项)
  3. 等待你完成步骤后用 --complete sX --next sY 推进

不要自己读完所有步骤文件后一次性写报告。编排脚本的存在就是为了防止你这么做。


S9 交付自检

无论是否走编排脚本,报告完成后的最终自检命令:

bash orchestrator.sh --self-check

脚本会自动检查 6 项(文件完整性 / 结构 / 推测标注 / 引用 / 模糊表述),然后列出 10 项手动核验清单。

所有 16 项必须全 ✅ 才能交付。


文件结构

deep-research-v3/
├── README.md             ← 人类阅读的说明文档
├── SKILL.md              ← Agent 阅读的 Skill 定义
├── orchestrator.sh       ← 编排脚本(入口)+ Agent 指令 + 自检
├── steps/                ← 10 个步骤文件(s0-s9)
├── templates/            ← 固定模板
└── rules/                ← 信源分级 + 写作风格
Usage Guidance
This skill looks purpose-aligned for deep research. Before installing, review the full orchestrator script, make sure bash and jq or python3 are available, choose where reports should be saved, and avoid confidential research topics unless you trust the search and webpage-fetch tools the agent will use.
Capability Analysis
Type: OpenClaw Skill Name: deep-research-skill-for-business Version: 1.0.0 The skill bundle implements a highly structured multi-step research workflow, but the orchestration script (orchestrator.sh) contains a shell injection vulnerability. Specifically, the 'init_session' function uses an unquoted heredoc (<<EOF) to write the user-supplied research topic into a JSON file, which allows for arbitrary command execution if the topic contains shell expansion sequences like $(command). While the instructions in SKILL.md and the various step files (s0-s9) appear benign and focused on research quality, the lack of input sanitization in the bash script poses a security risk.
Capability Assessment
Purpose & Capability
The purpose is coherent: the files define a staged business research workflow with source grading, search, scraping, writing, and self-checking. The use of external search/fetch tools is expected for deep research but worth noticing for confidential topics.
Instruction Scope
The skill uses strong mandatory workflow instructions and checkpoints. This is aligned with its stated quality-control purpose, but it intentionally constrains how the agent proceeds.
Install Mechanism
There is no install spec, but the skill requires running a local bash script and depends on bash plus jq or python3. The registry metadata does not fully declare those runtime binaries.
Credentials
The skill asks the agent to use current environment search and webpage-fetch tools and to write reports locally. No credentials, tokens, privileged system access, or broad local indexing are shown.
Persistence & Privilege
The orchestrator creates local state and report directories, which is proportionate for a research workflow but means topics and generated reports may persist after use.
How to Use
  1. Make sure OpenClaw is installed (local or Docker)
  2. Run the install command in chat: /install deep-research-skill-for-business
  3. After installation, invoke the skill by name or use /deep-research-skill-for-business
  4. Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.0.0
新建skill
Metadata
Slug deep-research-skill-for-business
Version 1.0.0
License MIT-0
All-time Installs 0
Active Installs 0
Total Versions 1
Frequently Asked Questions

What is deep-research skill?

商业决策级深度调研 Skill V3.1——多步骤编排 + 断点强制校验 + Agent 执行指令。触发条件:用户要求深度调研、行业分析、商业调研、竞品分析等。 It is an AI Agent Skill for Claude Code / OpenClaw, with 58 downloads so far.

How do I install deep-research skill?

Run "/install deep-research-skill-for-business" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is deep-research skill free?

Yes, deep-research skill is completely free, licensed under MIT-0. You can download, install and use it at no cost.

Which platforms does deep-research skill support?

deep-research skill is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created deep-research skill?

It is built and maintained by wayne260 (@leanerque260); the current version is v1.0.0.

💬 Comments