deep-research

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Deep Research

深度研究

Produce Gemini Deep Research-quality output: rich artifacts with embedded screenshots, mermaid diagrams, comparison tables, and narrative synthesis. Tuned for developer decisions — framework selection, architecture patterns, dependency evaluation, competitive analysis.
生成具备Gemini深度研究级别的成果:包含嵌入式截图、Mermaid图表、对比表格及叙事性综合分析的丰富产出物。专为开发者决策场景优化——如框架选型、架构模式、依赖评估、竞品分析等。

When to Use This Skill

何时使用该Skill

  • "Research the current state of X"
  • "Compare Framework A vs Framework B"
  • "What are the best approaches for..."
  • "Deep dive into..."
  • Any request where the answer requires synthesizing information from many sources
Do NOT use for: quick factual lookups, single-source answers, or "find me a CSS button" (use
design-lookup
instead).
  • 「调研X领域的当前现状」
  • 「对比框架A与框架B」
  • 「解决...问题的最佳方案有哪些」
  • 「深入探索...」
  • 任何需要整合多信息源内容的请求
请勿用于:快速事实查询、单信息源即可解答的问题,或「帮我找一个CSS按钮」(请使用
design-lookup
替代)。

Input Protocol — Before Any Search

输入协议 —— 搜索前准备

  1. Decompose the topic into 3-5 research axes.
    • Example: "Compare Next.js vs Remix" → Performance, DX, Ecosystem, Deployment, Community
  2. Identify the decision context — what is the user actually deciding?
    • Framework choice? Architecture pattern? Build vs buy? Migration risk?
  3. Draft a research plan — present 3-5 axes with planned queries to the user.
    • Save it as an artifact (e.g.,
      research_plan.md
      ).
    • Proceed on approval, or refine if the user redirects scope.
  1. 主题拆解:将研究主题拆解为3-5个研究维度。
    • 示例:「对比Next.js与Remix」→ 性能、DX(开发者体验)、生态系统、部署、社区
  2. 明确决策场景:用户实际要做什么决策?
    • 框架选型?架构模式选择?自研还是采购?迁移风险评估?
  3. 制定研究计划:向用户展示3-5个研究维度及对应的拟用查询词。
    • 将计划保存为产出物(如
      research_plan.md
      )。
    • 获得用户批准后再推进,若用户调整范围则相应优化计划。

Phase 1: Breadth Scan

第一阶段:广度扫描

Goal: Map the landscape. Find what exists before reading anything.
  1. Run 5-8 parallel searches across different axes. Use at least two tools:
    • tavily_search
      — broad topic queries
    • search_web
      — alternate search perspective
    • tavily_research
      — delegate an entire sub-question (powerful for "state of X" queries)
  2. Dev-specific breadth:
    • search_code
      or
      search_repositories
      — find relevant GitHub repos
    • Search npm trends, bundle sizes, download counts when evaluating packages
    • Search for migration stories: "migrating from X to Y" experience reports
  3. Collect 15-25 candidate URLs, not 5. Score each by authority tier (see references/research-heuristics.md).
  4. Do not stop at snippets. Snippets are for candidate selection only.
Output: Candidate source list with tier ratings. Present to user if interactive, or proceed if autonomous.
目标:梳理研究领域全貌,先明确有哪些信息源再深入阅读。
  1. 针对不同维度运行5-8个并行搜索,至少使用两种工具:
    • tavily_search
      —— 宽泛主题查询
    • search_web
      —— 补充搜索视角
    • tavily_research
      —— 委托处理整个子问题(对「X领域现状」类查询尤为高效)
  2. 开发者专属广度调研
    • search_code
      search_repositories
      —— 查找相关GitHub仓库
    • 评估包时,搜索npm趋势、包体积、下载量数据
    • 搜索迁移实践案例:「从X迁移到Y」的经验分享
  3. 收集15-25个候选URL,而非5个。根据权威等级为每个URL评分(参考references/research-heuristics.md)。
  4. 不要仅停留在搜索摘要:摘要仅用于候选信息源筛选。
产出:带权威等级评分的候选信息源列表。若为交互式场景则展示给用户,若为自主模式则直接推进。

Phase 2: Deep Read

第二阶段:深度阅读

Goal: Extract actual content — implementation details, code examples, benchmarks, data.
  1. Select the top 8-12 sources from Phase 1 (prioritize S and A tier).
  2. Full extraction — get the complete page content:
    • tavily_extract
      or
      read_url_content
      for text-heavy pages
    • tavily_crawl
      to follow documentation multi-page structures
    • browser_subagent
      to screenshot key pages (UIs, dashboards, architecture diagrams)
    • get_file_contents
      (GitHub MCP) to read actual source code from repos
  3. Analyze each source:
    • Extract specific claims, numbers, patterns, code examples
    • Note the authority tier and any bias (is this the framework's own marketing?)
    • Tag findings by research axis
  4. Self-correction: If a source is fluff (marketing-only, thin tutorial, SEO filler):
    • Discard it
    • Run a refined follow-up search with more specific terms
    • Try adding: "benchmark", "technical deep dive", "lessons learned", "postmortem"
Output: Annotated source notes organized by axis.
目标:提取实际内容——实现细节、代码示例、基准测试数据等。
  1. 从第一阶段的候选中筛选出排名前8-12的信息源(优先选择S级和A级)。
  2. 完整内容提取:获取页面全部内容:
    • 针对文本密集型页面,使用
      tavily_extract
      read_url_content
    • 使用
      tavily_crawl
      处理多页文档结构
    • 使用
      browser_subagent
      对关键页面(UI、仪表盘、架构图)截图
    • 使用
      get_file_contents
      (GitHub MCP)读取仓库中的实际源代码
  3. 逐个分析信息源
    • 提取具体结论、数据、模式、代码示例
    • 记录信息源的权威等级及可能存在的 bias(是否为框架官方营销内容?)
    • 按研究维度为研究发现打标签
  4. 自我修正:若信息源内容空洞(仅营销话术、浅显教程、SEO填充内容):
    • 丢弃该信息源
    • 使用更精准的关键词进行补充搜索
    • 尝试添加:「benchmark」「technical deep dive」「lessons learned」「postmortem」等关键词
产出:按研究维度整理的带注释信息源笔记。

Phase 3: Synthesis

第三阶段:内容整合

Goal: Build the research briefing artifact. This is the main deliverable.
  1. Choose the report template from references/report-templates.md:
    • Comprehensive Brief — for landscape/state-of-the-art research
    • Comparison Brief — for head-to-head evaluations
  2. Write the report as a rich markdown artifact:
    • Narrative prose in the executive summary — not bullets, not lists. Write as if briefing a tech lead.
    • Comparison tables with real data extracted from sources
    • Mermaid diagrams for architecture, decision trees, ecosystem maps
    • Embedded screenshots captured via
      browser_subagent
      during Phase 2
    • Code examples pulled from actual repos or docs
    • Use
      generate_image
      for custom visualizations when no screenshot captures the concept
  3. Cite every claim — link to the source URL inline. Use the format:
    [Source Name](URL)
    .
  4. Gap analysis — explicitly call out:
    • What couldn't be determined and why
    • Conflicting information between sources
    • Areas where only low-tier sources were found
Output: The research artifact (e.g.,
research_report.md
).
目标:生成研究简报成果,这是核心交付物。
  1. references/report-templates.md选择合适的报告模板
    • 综合简报 —— 适用于领域全景/前沿技术调研
    • 对比简报 —— 适用于一对一评估
  2. 撰写富格式Markdown报告
    • 执行摘要采用叙事性 prose,而非项目符号列表。以向技术负责人汇报的口吻撰写。
    • 加入从信息源提取的真实数据制作对比表格
    • 使用Mermaid图表展示架构、决策树、生态系统图谱
    • 嵌入第二阶段通过
      browser_subagent
      捕获的截图
    • 从实际仓库或文档中提取代码示例
    • 若截图无法呈现概念,使用
      generate_image
      生成自定义可视化内容
  3. 为所有结论添加引用 —— 内联链接至信息源URL。使用格式:
    [信息源名称](URL)
  4. 差距分析:明确指出:
    • 无法确定的内容及原因
    • 不同信息源之间的矛盾信息
    • 仅能找到低等级信息源的领域
产出:研究成果物(如
research_report.md
)。

Phase 4: Iteration

第四阶段:迭代优化

Goal: Fill gaps identified in Phase 3.
  1. Review the gap analysis section of your report.
  2. For each fillable gap:
    • Run 1-2 targeted searches with refined queries
    • Extract and read the results
    • Update the report artifact in-place
  3. Max 3 total iterations (Phase 1-3 = round 1, then up to 2 more targeted rounds).
  4. After final iteration, mark remaining gaps as "Unresolved" with explanation.
目标:填补第三阶段中发现的信息空白。
  1. 审阅报告中的差距分析部分。
  2. 针对每个可填补的空白:
    • 运行1-2次精准搜索
    • 提取并阅读结果
    • 直接更新报告成果物
  3. 最多进行3轮迭代(第一至第三阶段为第1轮,之后可进行最多2轮针对性补充)。
  4. 最终迭代后,将剩余空白标记为「未解决」并说明原因。

Tool Strategy

工具策略

PurposePrimaryFallback
Topic discovery
tavily_search
search_web
Delegated deep research
tavily_research
Manual multi-search
Full page extraction
tavily_extract
read_url_content
Multi-page docs
tavily_crawl
tavily_map
+ manual
Visual evidence
browser_subagent
(screenshot)
generate_image
GitHub analysis
search_code
,
get_file_contents
read_url_content
on raw GitHub
Architecture diagramsMermaid in markdown
generate_image
Data visualizationMarkdown tables
generate_image
for charts
用途首选工具备选工具
主题探索
tavily_search
search_web
委托式深度研究
tavily_research
手动多工具搜索
整页内容提取
tavily_extract
read_url_content
多页文档处理
tavily_crawl
tavily_map
+ 手动处理
视觉证据收集
browser_subagent
(截图)
generate_image
GitHub分析
search_code
,
get_file_contents
读取GitHub原始页面的
read_url_content
架构图制作Markdown中的Mermaid
generate_image
数据可视化Markdown表格
generate_image
生成图表

Quality Gates

质量校验清单

Before delivering the report, verify:
  • Source diversity — at least 1 S-tier and 2 A-tier sources cited (or explicitly flagged as unavailable)
  • Visual richness — at least 1 screenshot/image AND 1 diagram/table embedded
  • Narrative quality — executive summary reads as prose, not bullet points
  • Citation completeness — every factual claim links to a source
  • Gap transparency — gaps and conflicts are explicitly documented
  • Actionable output — recommendations section exists with ranked, specific advice
交付报告前,请验证:
  • 信息源多样性 —— 至少引用1个S级和2个A级信息源(若无法获取则需明确标记)
  • 视觉丰富度 —— 至少嵌入1张截图/图片 及 1个图表/表格
  • 叙事质量 —— 执行摘要为 prose 格式,而非项目符号列表
  • 引用完整性 —— 所有事实性结论均链接至信息源
  • 差距透明度 —— 明确记录信息空白及矛盾内容
  • 可行动性 —— 包含建议部分,提供分级的具体建议

Anti-Patterns

反模式

  • Snippet-only research — stopping at search result descriptions without full extraction
  • Text-wall reports — no visuals, no tables, no diagrams. The whole point is richness.
  • Source-by-source organization — findings must be grouped thematically by research axis, not by URL
  • Single-tool reliance — use at least 2 different search/extraction tools for source diversity
  • Uncited claims — every substantive finding must link to its source
  • Marketing echo — repeating a framework's own marketing claims without independent verification
  • Premature stopping — delivering after 3-5 sources when the topic warrants 15+
  • 仅依赖搜索摘要的研究:未提取完整内容就停留在搜索结果描述
  • 纯文本报告:无视觉元素、表格或图表。本技能的核心就是产出丰富的成果
  • 按信息源组织内容:研究发现需按研究维度主题分组,而非按URL分组
  • 单一工具依赖:至少使用2种不同的搜索/提取工具以保证信息源多样性
  • 无引用结论:所有实质性研究发现均需链接至其信息源
  • 营销话术重复:未经过独立验证就直接引用框架官方营销内容
  • 提前终止研究:当主题需要15+信息源时,仅调研3-5个就交付成果

References

参考资料

  • Source authority scoring and query patterns: references/research-heuristics.md
  • Report structure templates: references/report-templates.md
  • 信息源权威评分及查询模式references/research-heuristics.md
  • 报告结构模板references/report-templates.md