home / skills / cklxx / elephant.ai / deep-research

deep-research skill

safe

This skill enables deep, traceable research from question framing to actionable recommendations across multiple sources, reducing bias and risk.

npx playbooks add skill cklxx/elephant.ai --skill deep-research

Review the files below or copy the command above to add this skill to your agents.

Files (1)

SKILL.md

3.0 KB

---
name: deep-research
description: 深度调研技能，覆盖问题澄清、检索规划、多源验证、证据跟踪与行动建议。
---

# 深度调研（提问→检索→验证→决策）

## When to use this skill
- 需要对复杂议题做深入、可追溯的研究，输出决策可用的结论与行动建议。  
- 需要跨多源（官方文档/论文/标准/社区/数据）检索并化解信息冲突，降低偏差与过时风险。

## 必备输入
- 目标与决策：要判断/选择/设计什么？成功标准与限制条件（时间、预算、合规、用户范围）。  
- 已知信息：已有假设、来源、未解的问题、相关上下文。  
- 受众与输出要求：高层/技术/运营？期望格式（简报/备忘/表格）与语言。  
- 时间箱与深度：可投入时长、需要覆盖的来源类型、是否需要实验/访谈验证。

## 工作流
1. **问题框定与假设**：用 2–3 句重写问题与成功标准；列出现有假设与关键未知。  
2. **研究计划**：设定时间箱与终止条件（如“找到 5 个独立来源或出现信息收敛”）；列出要覆盖的来源清单（官方文档、标准、论文、数据集、权威博客、案例/复盘、论坛）；为不同来源准备关键词/同义词与查询语法。  
3. **检索与踪迹记录**：记录每轮查询的关键词、运算符与时间；保留未命中的思路，避免重复搜索；对有价值的线索立即做最小笔记（来源、日期、主要结论）。  
4. **可信度与偏差判断**：优先最新和权威/一手资料；标记可能的偏差（营销内容、过旧版本、地域/政策差异）；对关键数据记录口径与采集方法。  
5. **交叉验证与冲突处理**：
   - 寻找独立来源的佐证；对冲突信息写出可能原因（版本、场景、样本、测量方式）。  
   - 若存在高风险假设，设计最小验证（试验/POC/对标数据）并列出所需资源。  
6. **结构化输出**：按“问题 → 发现/证据 → 置信度 → 影响/建议”整理；区分事实/假设/风险；对关键决策给出选项、适用前提、利弊、潜在后果。  
7. **行动与空白**：列出下一步验证/访谈/数据拉取计划（负责人、截止时间）；标记仍未知或待确认的假设，提示需要的支持。

## 输出格式示例
- TL;DR（3–5 条）：结论 + 影响 + 推荐行动 + 置信度（高/中/低）。  
- 证据表：`问题/假设 | 发现 | 来源/时间 | 置信度 | 风险/空白 | 推荐行动`。  
- 决策备选：用表格对比选项、前提、优点、成本/风险、代表来源。  
- 追踪记录：附查询日志与已查看来源列表，便于复现与交接。

## 最终检查清单
- 关键结论均有可追溯来源或被标记为假设；时间与版本已注明。  
- 冲突信息已解释可能原因，并给出进一步验证路径。  
- 输出明确后续行动、负责人与时间，并说明哪些问题仍未知或超出范围。

Overview

This skill performs structured deep research to produce traceable conclusions and actionable recommendations for complex decisions. It guides the process from clarifying the question and assumptions, through multi-source retrieval and cross-validation, to evidence-backed decision options. Outputs emphasize provenance, confidence levels, and next-step validation plans.

How this skill works

The skill rewrites the problem and success criteria in 2–3 sentences, enumerates existing hypotheses and unknowns, and builds a time-boxed research plan listing source types and query terms. It records every search (keywords, operators, timestamps), captures minimal notes for promising leads, evaluates source credibility and bias, and cross-checks findings across independent sources. For conflicts or high-risk assumptions it proposes minimal validation experiments or data checks. Final deliverables are structured summaries linking each conclusion to evidence, confidence, risks, and concrete actions.

When to use it

When you must make a high-stakes or complex decision that needs documented evidence and traceability.
When information is scattered across papers, docs, standards, forums, and data and needs consolidation.
When you need to resolve conflicting claims or assess whether existing assumptions hold.
When you need a reproducible research trail for audits, stakeholders, or cross-team handoff.

Best practices

Start with a tight problem statement and explicit success criteria to limit scope and bias.
Time-box search rounds and set termination rules (e.g., N independent sources or convergence).
Prioritize primary and recent sources; flag marketing, outdated, or region-specific bias.
Log queries and findings immediately to avoid duplicated work and allow reproducibility.
Differentiate facts, assumptions, and risks; attach confidence levels and citation details.

Example use cases

Assessing vendor claims vs. independent benchmarks before procurement.
Evaluating regulatory compliance requirements across jurisdictions for a product launch.
Comparing competing research results and identifying why outcomes diverge.
Preparing an evidence-backed briefing for executives with clear recommended actions.
Designing a minimal POC to validate a critical technical or business assumption.

FAQ

What inputs do I need to start a deep research task?

Provide the decision goal, success criteria and constraints, known assumptions and sources, intended audience and output format, and available time/depth.

How are conflicting sources handled?

Conflicts are explained by likely causes (version, scope, method) and supported by cross-checks; high-risk conflicts trigger proposed minimal validations or data checks.