深度研究

Universal

deep-research

by daymade

围绕既定模板拆题、多轮检索、整理证据表并交叉校验引用,产出格式严格、可追溯的研究报告,适合市场调研、竞品分析、政策追踪等高要求写作任务

把网页搜索、信息检索和研究分析串成一条链路,适合需要快速汇总线索、做深度判断的开发与研究工作。

855搜索与获取未扫描2026年3月5日

安装

claude skill add --url github.com/daymade/claude-code-skills/tree/main/deep-research

文档

Deep Research

Create high-fidelity research reports with strict format control, evidence mapping, and multi-pass synthesis.

Quick Start

  1. Clarify the report spec and format contract
  2. Build a research plan and query set
  3. Collect evidence with the deepresearch tool (multi-pass if needed)
  4. Triage sources and build an evidence table
  5. Draft the full report in multiple complete passes (parallel subagents)
  6. UNION merge, enforce format compliance, verify citations
  7. Present draft for human review and iterate

Core Workflow

Copy this checklist and track progress:

code
Deep Research Progress:
- [ ] Step 1: Intake and format contract
- [ ] Step 2: Research plan and query set
- [ ] Step 3: Evidence collection (deepresearch tool)
- [ ] Step 4: Source triage and evidence table
- [ ] Step 5: Outline and section map
- [ ] Step 6: Multi-pass full drafting (parallel subagents)
- [ ] Step 7: UNION merge and format compliance
- [ ] Step 8: Evidence and citation verification
- [ ] Step 9: Present draft for human review and iterate

Step 1: Intake and Format Contract

Establish the report requirements before any research:

  • Confirm audience, purpose, scope, time range, and geography
  • Lock output format: Markdown, DOCX, slides, or user-provided template
  • Capture required sections and exact formatting rules
  • Confirm citation style (footnotes, inline, numbered, APA, etc.)
  • Confirm length targets per section
  • Ask for any existing style guide or sample report

Create a concise report spec file:

code
Report Spec:
- Audience:
- Purpose:
- Scope:
- Time Range:
- Geography:
- Required Sections:
- Section Formatting Rules:
- Citation Style:
- Output Format:
- Length Targets:
- Tone:
- Must-Include Sources:
- Must-Exclude Topics:

If a user provides a template or an example report, treat it as a hard constraint and mirror the structure.

Step 2: Research Plan and Query Set

Define the research strategy before calling tools:

  • Break the main question into 3-7 subquestions
  • Define key entities, keywords, and synonyms
  • Identify primary sources vs secondary sources
  • Define disqualifiers (outdated, low quality, opinion-only)
  • Assemble a query set per section

Use references/research_plan_checklist.md for guidance.

Step 3: Evidence Collection (Deepresearch Tool)

Use the deepresearch tool to collect evidence and citations.

  • Run multiple complete passes if coverage is uncertain
  • Vary query phrasing to reduce blind spots
  • Preserve raw tool output in files for traceability

File structure (recommended):

code
<output_dir>/research/<topic-name>/
  deepresearch_pass1.md
  deepresearch_pass2.md
  deepresearch_pass3.md

If deepresearch is unavailable, rely on user-provided sources only and state limitations explicitly.

Step 4: Source Triage and Evidence Table

Normalize and score sources before drafting:

Evidence table minimum columns:

  • Source ID
  • Title
  • Publisher
  • Date
  • URL or reference
  • Quality tier (A/B/C)
  • Notes

Step 5: Outline and Section Map

Create an outline that enforces the format contract:

Step 6: Multi-Pass Full Drafting (Parallel Subagents)

Avoid single-pass drafting; generate multiple complete reports, then merge.

Preferred Strategy: Parallel Subagents (Complete Draft Each)

Use the Task tool to spawn parallel subagents with isolated context. Each subagent must:

  • Load the report spec, outline, and evidence table
  • Draft the FULL report (all sections)
  • Enforce formatting rules and citation style

Implementation pattern:

code
Task(subagent_type="general-purpose", prompt="Draft complete report ...", run_in_background=false) -> version1.md
Task(subagent_type="general-purpose", prompt="Draft complete report ...", run_in_background=false) -> version2.md
Task(subagent_type="general-purpose", prompt="Draft complete report ...", run_in_background=false) -> version3.md

Write drafts to files, not conversation context:

code
<output_dir>/intermediate/<topic-name>/version1.md
<output_dir>/intermediate/<topic-name>/version2.md
<output_dir>/intermediate/<topic-name>/version3.md

Step 7: UNION Merge and Format Compliance

Merge using UNION, never remove content without evidence-based justification:

  • Keep all unique findings from all versions
  • Consolidate duplicates while preserving the most detailed phrasing
  • Ensure every claim in the merged draft has a cited source
  • Enforce the exact section order, headings, and formatting
  • Re-run formatting rules from references/formatting_rules.md

Step 8: Evidence and Citation Verification

Verify traceability:

  • Every numeric claim has at least one source
  • Every recommendation references supporting evidence
  • No orphan claims without citations
  • Dates and time ranges are consistent
  • Conflicts are explicitly called out with both sources

Use references/completeness_review_checklist.md.

Step 9: Present Draft for Human Review and Iterate

Present the draft as a reviewable version:

  • Emphasize that format compliance and factual accuracy need human review
  • Accept edits to format, structure, and scope
  • If the user provides another AI output, cross-compare and UNION merge

Output Requirements

  • Match the requested language and tone
  • Preserve technical terms in English
  • Respect the report spec and formatting rules
  • Include a references section or bibliography

Reference Files

FileWhen to Load
research_report_template.mdBuild outline and draft structure
formatting_rules.mdEnforce section formatting and citation rules
source_quality_rubric.mdScore and triage sources
research_plan_checklist.mdBuild research plan and query set
completeness_review_checklist.mdReview for coverage, citations, and compliance

Anti-Patterns

  • Single-pass drafting without parallel complete passes
  • Splitting passes by section instead of full report drafts
  • Ignoring the format contract or user template
  • Claims without citations or evidence table mapping
  • Mixing conflicting dates without calling out discrepancies
  • Copying external AI output without verification
  • Deleting intermediate drafts or raw research outputs

相关 Skills

SEO审计工具

by amdf01-debug

热门

搜索与获取
未扫描4.2k

接口规范

by alexxxiong

热门

API 规范管理工具 - 跨项目 API 文档的初始化、更新、查询与搜索。Triggers: 'API文档', 'API规范', '接口文档', '路由解析', 'apispec', 'API lookup', 'API search'.

搜索与获取
未扫描4.2k

多视角辩论

by caius-kong

热门

Multi-perspective structured debate for complex topics. Spawn parallel subagents with opposing stances, cross-inject arguments for rebuttal, then synthesize via neutral judge into a consensus report with recommendations and scenario matrix. Use when: (1) user asks for deep comparison, pros/cons, or X vs Y analysis, (2) user asks for multi-angle research on a controversial or complex topic, (3) user explicitly requests debate, dialectical analysis, or adversarial research. NOT for: simple factual lookups, single-perspective deep research (use academic-deep-research), or quick opinion questions.

搜索与获取
未扫描4.2k

相关 MCP 服务

网页抓取

编辑精选

by Anthropic

热门

Fetch 是 MCP 官方参考服务器,让 AI 能抓取网页并转为 Markdown 格式。

这个服务器解决了 AI 直接处理网页内容时格式混乱的问题,适合需要让 Claude 分析在线文档或新闻的开发者。不过作为参考实现,它缺乏生产级的安全配置,你得自己处理反爬虫和隐私风险。

搜索与获取
83.9k

Brave 搜索

编辑精选

by Anthropic

热门

Brave Search 是让 Claude 直接调用 Brave 搜索 API 获取实时网络信息的 MCP 服务器。

如果你想让 AI 助手帮你搜索最新资讯或技术文档,这个工具能绕过传统搜索的限制,直接返回结构化数据。特别适合需要实时信息的开发者,比如查 API 更新或竞品动态。不过它依赖 Brave 的 API 配额,高频使用可能受限。

搜索与获取
83.9k

by Anthropic

热门

Puppeteer 是让 Claude 自动操作浏览器进行网页抓取和测试的 MCP 服务器。

这个服务器解决了手动编写 Puppeteer 脚本的繁琐问题,适合需要自动化网页交互的开发者,比如抓取动态内容或做端到端测试。不过,作为参考实现,它可能缺少生产级的安全防护,建议在可控环境中使用。

搜索与获取
83.9k

评论