ResearchMate

v2.1.0

深度写作素材采集 Skill。当用户需要为文章、报告、视频脚本采集素材,或说"帮我找XX相关资料"、"我要写一篇关于XX的文章"、"帮我做选题调研"、"给我采集XX素材"时触发。通过三步交互问询(选题→关键实体→目标字数)+ 多源自动采集 + 四重反幻觉验证(数值/主体/时间/来源)+ 结构化输出,为深度写作提供...

0· 69·0 current·0 all-time

Install

OpenClaw Prompt Flow

Install with OpenClaw

Best for remote or guided setup. Copy the exact prompt, then paste it into OpenClaw for yitao2027/research-mate.

Previewing Install & Setup.
Prompt PreviewInstall & Setup
Install the skill "ResearchMate" (yitao2027/research-mate) from ClawHub.
Skill page: https://clawhub.ai/yitao2027/research-mate
Keep the work scoped to this skill only.
After install, inspect the skill metadata and help me finish setup.
Use only the metadata you can verify from ClawHub; do not invent missing requirements.
Ask before making any broader environment changes.

Command Line

CLI Commands

Use the direct CLI path if you want to install manually and keep every step visible.

OpenClaw CLI

Bare skill slug

openclaw skills install research-mate

ClawHub CLI

Package manager switcher

npx clawhub@latest install research-mate
Security Scan
VirusTotalVirusTotal
Benign
View report →
OpenClawOpenClaw
Benign
high confidence
Purpose & Capability
The name/description (深度写作素材采集) aligns with the SKILL.md instructions: three-step query, multi-source web search, extraction, four-fold verification, scoring, and structured outputs. The declared capabilities (web_search, read_url, create_file usage) are coherent with the stated goal and no unrelated credentials or system accesses are requested.
Instruction Scope
The SKILL.md gives detailed, scoped instructions that stay within the stated purpose: it instructs the agent to run web_search/read_url to fetch publicly available material, perform validation checks, score items, and produce Markdown/CSV/Word/PDF outputs. It does not instruct reading arbitrary local files or environment secrets. Note: the skill assumes the agent has helper tools (web_search, read_url, create_file and possibly document-export capabilities); if those tools are not present or have broader network/IO privileges, actual behavior may differ.
Install Mechanism
This is an instruction-only skill with no install spec and no bundled code, which is the lowest-risk form. However, README.md describes a full Python project (main.py, src/, requirements.txt, clone/install steps). That repository content is not included here—only SKILL.md and README.md were packaged. This mismatch is not necessarily malicious but is an inconsistency: some advertised local features (running a Python binary, pip installs, local exporters) won't work unless the external repo is obtained and run.
Credentials
The skill requests no environment variables, secrets, or config paths. The SKILL.md explicitly states it will not access paid or private databases and will not gather unpublished internal data. The lack of credentials is proportionate to the described public-data collection functionality.
Persistence & Privilege
always is false and the skill is user-invocable; autonomous invocation is allowed by default (disable-model-invocation is false) which is standard for skills. The skill does not request elevated/persistent privileges or modifications to other skills or system-wide settings.
Assessment
This skill appears internally consistent with its stated purpose and contains no requests for credentials or hidden endpoints. Before installing, confirm the host platform provides the tools the SKILL.md expects (web_search, read_url, create_file and any document-export support). Be aware the README advertises a separate Python implementation (main.py, requirements.txt) but that code is not bundled in this skill package — if you need local Python features or Word/PDF exporters, you would have to retrieve and inspect that external repository yourself. Finally, remember that the skill will collect and present third-party content: verify any high-stakes facts against original source documents before publishing.

Like a lobster shell, security has layers — review code before you run it.

latestvk97ds61axfhar16t4mh8syde2h8502yj
69downloads
0stars
1versions
Updated 1w ago
v2.1.0
MIT-0

ResearchMate — 深度写作素材采集 Skill

让 AI 像专业研究员一样「系统采集 + 严格验证」,为你的深度写作提供可信赖的素材库。


触发条件

满足以下任意一项时触发本 Skill:

  • 用户说"帮我采集素材"、"帮我找XX相关资料"、"我要写一篇关于XX的文章"
  • 用户说"帮我做选题调研"、"给我找XX案例/数据/观点"
  • 用户需要为公众号文章、视频脚本、研究报告、商业分析收集内容基础
  • 用户提到"素材不够"、"文章缺乏数据支撑"、"找不到好案例"

不触发场景: 纯问答、日常闲聊、代码调试、不需要素材支撑的简短写作(如写一封邮件、生成一段广告语)


核心能力

能力说明实现方式
🗣️ 交互式问询三步确认真实需求,避免方向跑偏结构化对话
🔍 多源自动采集按选题类型路由到最优来源组合web_search + read_url
🛡️ 四重反幻觉验证数值/主体/时间/来源缺一不可结构化验证规则
📊 质量评分每条素材四维度评分(0-1)+ 综合评分(0-100)加权算法
📦 结构化输出Markdown 素材库 + CSV 评估表create_file

执行流程

Step 1:三步交互问询(必须完整执行,不可跳过)

在采集任何内容之前,依次询问以下三个问题:

❓ 请描述您的选题(想写什么主题的文章):
🔑 关键词(产品名称、企业名称、人物姓名或事件名称,多个用逗号分隔):
📝 您计划写多少字的内容?(例如:3000):

收到回答后,输出确认摘要:

✅ 已确认采集需求:
📌 选题:[用户填写]
🔑 关键词:[用户填写]
📊 目标字数:[用户填写] 字
📚 预计采集素材:[目标字数×8~10倍] 字(按 8-10 倍原则)

为什么必须问这三个问题:

  • 选题描述 → 只有了解写作角度,才能精准匹配素材类型
  • 关键实体 → 锁定具体公司/产品/人物,避免泛泛而谈
  • 目标字数 → 智能计算采集量(8-10 倍原则),避免素材不足或过剩

Step 2:按选题类型路由采集源

根据选题描述,自动判断需求类型并激活对应采集源组合:

需求类型激活来源
科技/产品科技媒体(36氪、虎嗅、极客公园)、官方博客、GitHub
财经/商业财报、券商研报、财经媒体(界面新闻、雪球、Wind资讯)
政策/监管政府官网、新华社、人民日报、行业协会公告
学术/研究学术媒体报道、行业白皮书、研究机构报告
人物/事件新闻媒体、采访报道、官方声明
国际动态英文媒体(Reuters、Bloomberg报道摘要)

每个来源使用 web_search 搜索,再用 read_url 提取正文内容。

采集量目标: 目标字数 × 810 倍(例如写 5000 字 → 采集 4000050000 字素材)


Step 3:四重反幻觉验证(每条素材必须通过)

对每一条采集到的数据/观点/案例,执行以下四项检查:

验证维度检查内容失败处理
✅ 数值检查必须有具体数字(增长率、金额、占比等)标记为"低可信度"
✅ 主体检查必须明确公司/产品/机构名称要求补充来源
✅ 时间检查必须有清晰的时间点或时间段标注时效性风险
✅ 来源检查必须标注数据来源(财报/研报/媒体报道)降权或剔除

通过全部四项 → 标记 ✅ 反幻觉检查通过 任意一项失败 → 标记对应风险,不得在输出中作为可信事实引用


Step 4:质量评分

对每条通过验证的素材进行四维度评分:

维度权重说明
可信度30%来源权威性、数据可核实程度
时效性25%发布时间距今的距离
完整性25%信息是否完整、上下文是否充分
交叉验证20%是否有多个来源相互印证

综合评分 = 四维度加权求和 × 100,评级标准:

  • S 级(≥90):直接引用,高优先级
  • A 级(80-89):可用,建议补充交叉验证
  • B 级(70-79):谨慎使用,需补充说明
  • C 级(<70):不建议引用,仅供参考

Step 5:结构化输出

输出两份文件:

① Markdown 素材库(主交付物)

每条素材按以下结构呈现:

## 素材 #001 — [类型] [主题]

### 📝 核心内容摘要
[2-3句话概括核心信息]

### ✅ 验证信息
- **数据来源:** [具体来源名称]
- **发布时间:** [具体日期]
- **反幻觉检查:** ✅/⚠️(数值✓/✗ 主体✓/✗ 时间✓/✗ 来源✓/✗)

### 🔗 引用建议
> [可直接用于文章的引用格式,含来源标注]

### 📎 关联素材
- [与其他素材的关联提示]

② CSV 评估表(辅助交付物)

ID,类型,主题,来源URL,采集时间,可信度,时效性,完整性,交叉验证,综合评分,等级,是否采用,补充建议
001,财务数据,XX公司Q4财报,https://...,2026-04-17,0.92,0.95,0.88,0.85,90,S,☐,

防幻觉铁律(不可违反)

  1. 禁止无来源数据:所有数字、结论、案例必须附带可追溯的原始 URL 和发布时间
  2. 禁止模糊时间:不得写"近年来"、"最近",必须写具体年月
  3. 禁止主体缺失:不得写"某企业"、"有公司",必须写具体名称
  4. 禁止推断性数据:采集的是已发布的事实,不是 AI 推断的结论
  5. 低可信素材必须标注:不得将 C 级素材混入主素材库而不加警示

输出示例

用户输入:帮我采集比亚迪2025年财务表现的素材,准备写5000字分析文章

问询确认后,输出结构:

✅ 已确认采集需求:
📌 选题:比亚迪2025年财务表现分析
🔑 关键词:比亚迪、2025年财报、营收、净利润、毛利率
📊 目标字数:5000 字
📚 预计采集素材:40000-50000 字

正在采集中...
[采集来源1] 比亚迪官方财报披露 → 提取关键数据
[采集来源2] 券商研报摘要 → 提取分析观点
[采集来源3] 财经媒体报道 → 提取市场反应
...

共采集 [N] 条素材,通过四重验证 [M] 条
S级:[x]条 | A级:[y]条 | B级:[z]条 | C级(已剔除):[w]条

能力边界声明

ResearchMate 能做:

  • 系统化采集公开发布的新闻、财报、研报、官方公告等内容
  • 对采集内容进行四重验证和质量评分
  • 生成结构化素材库,方便写作时快速调用

ResearchMate 不能做:

  • 访问付费数据库(Wind、Bloomberg 终端等需订阅的内容)
  • 采集需要登录的内部资料
  • 生成或推断未公开发布的数据(这是幻觉,不是素材)
  • 替代人工判断——最终引用决策由作者负责

Comments

Loading comments...