开发工具CLIAI 技能/插件测试人工智能AI 评测【免费下载链接】SuperClaude_FrameworkA configuration framework that enhances Claude Code with specialized commands, cognitive personas, and development methodologies.项目地址https://gitcode.com/gh_mirrors/su/SuperClaude_Framework点击查看免费下载本篇技术指南以 SuperClaude Framework 中的 deep_research_workflows.md 为核心骨架系统讲解/sc:research深度研究命令的三种规划策略planning-only、intent-planning、unified、十类典型实战场景、三大高级工作流模式与质量校验体系。读者读完将掌握如何根据查询复杂度选择策略、如何编排多跳multi-hop搜索与工具路由Tavily / Playwright / Sequential Thinking / Serena、如何用置信度与自反思机制驱动研究中途重规划最终输出带证据链与引用的高质量研究报告。一、理解深度研究命令的整体设计在深入示例之前先建立对/sc:research命令的整体认知。该命令定义在 research.md核心定位是Deep web research with adaptive planning and intelligent search即具备自适应规划与智能搜索能力的深度网络研究。1.1 命令形态与触发场景/sc:research [query] [--depth quick|standard|deep|exhaustive] [--strategy planning|intent|unified]典型触发场景包括超出模型知识截止日期knowledge cutoff的研究问题复杂研究问题时事与实时信息学术或技术研究需求市场分析与竞争情报。1.2 六阶段行为流命令文档将一次研究拆分为六个阶段各阶段标注了建议精力占比阶段精力占比关键动作Understand理解5-10%评估查询复杂度与歧义、识别所需信息类型、定义成功标准Plan规划10-15%按复杂度选择规划策略、识别并行化机会、生成问题分解、建立里程碑TodoWrite任务分解5%建立自适应任务层级规模按查询复杂度伸缩至 3-15 个任务设立依赖关系Execute执行50-60%并行优先搜索、智能提取路由、多跳探索、证据与置信度跟踪Track跟踪持续监控进度、更新置信度、记录成功模式、识别信息缺口Validate验证10-15%验证证据链、核查来源可信度、消解矛盾、确保完整性1.3 自适应深度Adaptive Depth等级--depth参数决定研究的深度档位对应资源投入与输出复杂度其完整定义在 RESEARCH_CONFIG.md 的research_depth_profiles中档位最大来源数最大跳数迭代次数时间上限置信度目标提取方式quick10112 分钟0.6仅 Tavilystandard20325 分钟0.7选择性提取deep40438 分钟0.8综合提取exhaustive505510 分钟0.9全来源提取1.4 MCP 工具协同分工深度研究依赖四个 MCP 服务器协同工作其配置与角色如下MCP 服务器角色配置来源Tavily主搜索与提取引擎tavily.json通过TAVILY_API_KEY环境变量认证Sequential Thinking复杂推理与综合引擎sequential.jsonPlaywrightJavaScript 重内容提取与交互playwright.jsonSerena研究会话持久化与记忆后端serena.json提取路由规则定义于 RESEARCH_CONFIG.md 的extraction_routing决定了何时用哪个工具静态 HTML 走 TavilyJavaScript 渲染/动态内容/需要认证/需截图时走 Playwright技术文档/API 参考走 Context7本地文件与通用知识走原生工具。1.5 输出边界命令有明确的CRITICAL BOUNDARIES/sc:research只产出研究报告保存为claudedocs/research_[topic]_[timestamp].md不做任何实现、不写代码、不做架构决策。研究完成后如需架构设计应调用/sc:design如需编码应调用/sc:implement。二、三种规划策略的选型逻辑深度研究的自适应规划由--strategy参数驱动三种策略在 RESEARCH_CONFIG.md 的planning_strategies中有精确配置在 deep-research-agent.md 中有行为描述策略澄清提问用户确认执行时机适用场景指标planning-only否否立即执行查询清晰具体、技术文档类请求、范围明确、无歧义intent-planning是最多 3 问否澄清后执行存在歧义术语、主题宽泛、多种可能解释、用户专业水平未知unified可选是展示计划确认后执行复杂多面查询、用户协作有益、预期迭代细化、高 stakes 研究三、十个实战示例详解以下十个示例完整继承自 deep_research_workflows.md覆盖从明确问题到多域并行的全谱系场景。示例 1Planning-Only 策略清晰问题直接执行场景清晰研究问题 Latest TensorFlow 3.0 featuresTensorFlow 3.0 最新特性。/sc:research Latest TensorFlow 3.0 features --strategy planning-only --depth standard工作流1. Planning (Immediate): - Decompose: Official docs, changelog, tutorials - No user clarification needed 2. Execution: - Hop 1: Official TensorFlow documentation - Hop 2: Recent tutorials and examples - Confidence: 0.85 achieved 3. Synthesis: - Features list with examples - Migration guide references - Performance comparisons要点查询无歧义时直接分解为官方文档 → 变更日志 → 教程三个来源域跳过澄清环节两跳即达成 0.85 置信度最终合成特性清单 迁移指南 性能对比。示例 2Intent-Planning 策略歧义请求先澄清场景歧义请求 AI safetyAI 安全。/sc:research AI safety --strategy intent-planning --depth deep工作流1. Intent Clarification: Questions: - Are you interested in technical AI alignment, policy/governance, or current events? - Whats your background level (researcher, developer, general interest)? - Any specific AI systems or risks of concern? 2. User Response: - Technical alignment for LLMs, researcher level 3. Refined Planning: - Focus on alignment techniques - Academic sources priority - Include recent papers 4. Multi-Hop Execution: - Hop 1: Recent alignment papers - Hop 2: Key researchers and labs - Hop 3: Emerging techniques - Hop 4: Open problems 5. Self-Reflection: - Coverage: Complete ✓ - Depth: Adequate ✓ - Confidence: 0.82 ✓要点AI safety 一词可指技术对齐、政策治理或时事必须先澄清。用户回答面向 LLM 的技术对齐、研究者水平后研究转向学术源优先四跳探索覆盖论文、学者、新兴技术与开放问题最后自反思确认覆盖完整、深度充分、置信度 0.82。示例 3Unified 策略 中途重规划复杂协作式研究场景复杂研究 Build AI startup competitive analysisAI 创业公司竞争分析。/sc:research Build AI startup competitive analysis --strategy unified --hops 5工作流1. Initial Plan Presentation: Proposed Research Areas: - Current AI startup landscape - Funding and valuations - Technology differentiators - Market positioning - Growth strategies Does this cover your needs? Any specific competitors or aspects to focus on? 2. User Adjustment: Focus on code generation tools, include pricing and technical capabilities 3. Revised Multi-Hop Research: - Hop 1: List of code generation startups - Hop 2: Technical capabilities comparison - Hop 3: Pricing and business models - Hop 4: Customer reviews and adoption - Hop 5: Investment and growth metrics 4. Mid-Research Replanning: - Low confidence on technical details (0.55) - Switch to Playwright for interactive demos - Add GitHub repository analysis 5. Quality Gate Check: - Technical coverage: Improved to 0.78 ✓ - Pricing data: Complete 0.90 ✓ - Competitive matrix: Generated ✓要点unified 策略先呈现初始研究计划并征求用户反馈用户将焦点收敛到代码生成工具 定价 技术能力后重排五跳。研究途中发现技术细节置信度仅 0.55低于 RESEARCH_CONFIG.md 中confidence_scoring.minimum_threshold的 0.6触发重规划切换到 Playwright 抓取交互式演示、补充 GitHub 仓库分析最终技术覆盖提升至 0.78、定价数据 0.90。示例 4基于案例的研究学习Memory 复用场景与历史研究相似的问题 Rust async runtime comparisonRust 异步运行时对比。/sc:research Rust async runtime comparison --memory enabled工作流1. Case Retrieval: Found Similar Case: - Go concurrency patterns research - Successful pattern: Technical benchmarks code examples community feedback 2. Adapted Strategy: - Use similar structure for Rust - Focus on: Tokio, async-std, smol - Include benchmarks and examples 3. Execution with Known Patterns: - Skip broad searches - Direct to technical sources - Use proven extraction methods 4. New Learning Captured: - Rust community prefers different metrics than Go - Crates.io provides useful statistics - Discord communities have valuable discussions 5. Memory Update: - Store successful Rust research patterns - Note language-specific source preferences - Save for future Rust queries要点--memory enabled激活案例检索case-based reasoning。系统回忆到成功的 Go concurrency patterns 案例复用其基准测试 代码示例 社区反馈结构跳过宽泛搜索直达技术源。研究结束捕获新学习Rust 社区偏好不同指标、Crates.io 统计价值、Discord 讨论价值并写回记忆供未来 Rust 查询复用。案例的持久化结构定义在 RESEARCH_CONFIG.md 的case_schema中。示例 5自反思精化循环置信度驱动重规划场景演进式研究 Quantum computing for optimization用于优化的量子计算。/sc:research Quantum computing for optimization --confidence 0.8 --depth exhaustive工作流1. Initial Research Phase: - Academic papers collected - Basic concepts understood - Confidence: 0.65 (below threshold) 2. Self-Reflection Analysis: Gaps Identified: - Practical implementations missing - No industry use cases - Mathematical details unclear 3. Replanning Decision: - Add industry reports - Include video tutorials for math - Search for code implementations 4. Enhanced Research: - Hop 1→2: Papers → Authors → Implementations - Hop 3→4: Companies → Case studies - Hop 5: Tutorial videos for complex math 5. Quality Achievement: - Confidence raised to 0.82 ✓ - Comprehensive coverage achieved - Multiple perspectives included要点首轮研究置信度 0.65低于--confidence 0.8目标阈值。自反思识别三类缺口缺实践实现、缺行业用例、数学细节不清据此重规划补充行业报告、数学视频教程与代码实现检索通过实体扩展论文 → 作者 → 实现与领域深入公司 → 案例将置信度提升至 0.82。RESEARCH_CONFIG.md 的self_reflection配置了触发条件置信度低于阈值、检测到矛盾、时间消耗达 80%、用户介入与动作质量评估、缺口识别、重规划、策略调整。示例 6技术文档研究Playwright 处理 JS 重页面场景研究最新 Next.js 14 App Router 特性。/sc:research Next.js 14 App Router complete guide --depth deep --scrape selective --screenshots工作流1. Tavily Search: - Find official docs, tutorials, blog posts - Identify JavaScript-heavy documentation sites 2. URL Analysis: - Next.js docs → JavaScript rendering required - Blog posts → Static content, Tavily sufficient - Video tutorials → Need transcript extraction 3. Playwright Navigation: - Navigate to official documentation - Handle interactive code examples - Capture screenshots of UI components 4. Dynamic Extraction: - Extract code samples - Capture interactive demos - Document routing patterns 5. Synthesis: - Combine official docs with community tutorials - Create comprehensive guide with visuals - Include code examples and best practices要点先由 Tavily 完成发现再对 URL 做复杂度评估官方文档需要 JavaScript 渲染走 Playwright、博客为静态内容Tavily 足够、视频教程需提取字幕。--scrape selective意味着按需提取而非全量抓取--screenshots开启截图以保留 UI 组件视觉证据。示例 7竞争情报 可视化文档场景分析 AI 写作助手工具的定价与功能。/sc:research AI writing assistant tools pricing features 2024 --scrape all --screenshots --interactive工作流1. Market Discovery: - Tavily finds: Jasper, Copy.ai, Writesonic, etc. - Identify pricing pages and feature lists 2. Complexity Assessment: - Dynamic pricing calculators detected - Interactive feature comparisons found - Login-gated content identified 3. Playwright Extraction: - Navigate to each pricing page - Interact with pricing sliders - Capture screenshots of pricing tiers 4. Feature Analysis: - Extract feature matrices - Compare capabilities - Document limitations 5. Report Generation: - Competitive positioning matrix - Visual pricing comparison - Feature gap analysis - Strategic recommendations要点竞争情报场景常遇到动态定价计算器、交互式对比组件与登录门槛。--interactive允许 Playwright 操作价格滑块等交互元素--scrape all全量提取--screenshots记录各定价层级截图最终产出定位矩阵、可视化定价对比与功能差距分析。示例 8学术研究认证与访问策略场景研究最新机器学习论文 transformer architecture improvements 2024。/sc:research transformer architecture improvements 2024 --depth exhaustive --auth --scrape auto工作流1. Academic Search: - Tavily finds papers on arXiv, IEEE, ACM - Identify open vs. gated content 2. Access Strategy: - arXiv: Direct access, no auth needed - IEEE: Institutional access required - ACM: Mixed access levels 3. Extraction Approach: - Public papers: Tavily extraction - Gated content: Playwright with auth - PDFs: Download and process 4. Citation Network: - Follow reference chains - Identify key contributors - Map research lineage 5. Literature Synthesis: - Chronological development - Key innovations identified - Future directions mapped - Comprehensive bibliography要点--auth启用认证处理需遵守 RESEARCH_CONFIG.md 中extraction_settings.authentication_handling: ethical的伦理边界。访问策略按源类型分流arXiv 直接访问、IEEE 需机构认证、ACM 混合级别。提取阶段按内容类型路由公开论文走 Tavily、受限内容走 Playwright、PDF 下载处理并通过引用网络追踪研究谱系。示例 9实时市场数据研究场景收集加密货币市场分析BTC、ETH2024。/sc:research cryptocurrency market analysis BTC ETH 2024 --scrape all --interactive --screenshots工作流1. Market Discovery: - Find: CoinMarketCap, CoinGecko, TradingView - Identify real-time data sources 2. Dynamic Content Handling: - Playwright loads live charts - Capture price movements - Extract volume data 3. Interactive Analysis: - Interact with chart timeframes - Toggle technical indicators - Capture different views 4. Data Synthesis: - Current market conditions - Technical analysis - Sentiment indicators - Visual documentation 5. Report Output: - Market snapshot with charts - Technical analysis summary - Trading volume trends - Risk assessment要点实时行情页面均为动态内容必须由 Playwright 加载实时图表、切换时间周期与技术指标截图保留不同视图。最终报告包含带图表的市场快照、技术分析摘要、成交量趋势与风险评估。示例 10多域研究 并行执行场景综合研究 AI in healthcare 2024AI 医疗应用。/sc:research AI in healthcare applications 2024 --depth exhaustive --hops 5 --parallel工作流1. Domain Decomposition: Parallel Searches: - Medical AI applications - Regulatory landscape - Market analysis - Technical implementations - Ethical considerations 2. Multi-Hop Exploration: Each Domain: - Hop 1: Broad landscape - Hop 2: Key players - Hop 3: Case studies - Hop 4: Challenges - Hop 5: Future trends 3. Cross-Domain Synthesis: - Medical ↔ Technical connections - Regulatory ↔ Market impacts - Ethical ↔ Implementation constraints 4. Quality Assessment: - Coverage: All domains addressed - Depth: Sufficient detail per domain - Integration: Cross-domain insights - Confidence: 0.87 achieved 5. Comprehensive Report: - Executive summary - Domain-specific sections - Integrated analysis - Strategic recommendations - Visual evidence要点--parallel开启并行执行。先按域分解医疗应用、监管、市场、技术实现、伦理五个子域并行搜索每个子域内部再执行五跳递进全局 → 关键玩家 → 案例 → 挑战 → 未来趋势最后跨域综合医疗与技术、监管与市场、伦理与实现之间的交叉影响达成 0.87 置信度。并行机制与 parallel.py 及 RESEARCH_CONFIG.md 中parallel_execution_rules的强制并行默认值一致。四、三大高级工作流模式模式 1迭代深化Iterative Deepening多轮收敛式研究适用于从广到深的渐进探索Round_1: - Broad search for landscape - Identify key areas Round_2: - Deep dive into key areas - Extract detailed information Round_3: - Fill specific gaps - Resolve contradictions Round_4: - Final validation - Quality assurance模式 2来源三角验证Source Triangulation通过多层级来源交叉验证提升可信度Primary_Sources: - Official documentation - Academic papers Secondary_Sources: - Industry reports - Expert analysis Tertiary_Sources: - Community discussions - User experiences Synthesis: - Cross-validate findings - Identify consensus - Note disagreements来源可信度打分可参考 RESEARCH_CONFIG.md 的source_credibility四级矩阵学术期刊/政府出版物/官方文档/同行评审论文为 Tier 10.9-1.0成熟媒体/行业报告/专家博客/技术论坛为 Tier 20.7-0.9社区资源/用户文档/已认证社交媒体/Wikipedia 为 Tier 30.5-0.7用户论坛/未认证社交媒体/个人博客/评论区为 Tier 40.3-0.5。模式 3时间维度分析Temporal Analysis沿时间轴组织研究用于趋势与演进类问题Historical_Context: - Past developments - Evolution timeline Current_State: - Present situation - Recent changes Future_Projections: - Trends analysis - Expert predictions Synthesis: - Development trajectory - Inflection points - Future scenarios五、性能优化技巧查询优化Query Optimization以具体术语起步Start with specific terms尽早使用领域过滤器Use domain filters early批量合并相似搜索Batch similar searches缓存中间结果Cache intermediate results复用成功模式Reuse successful patterns。缓存策略在 RESEARCH_CONFIG.md 中有具体配置Tavily 搜索结果缓存 1 小时、Playwright 提取缓存 24 小时、Sequential 分析缓存 1 小时、案例模式始终复用。提取效率Extraction Efficiency先评估内容复杂度Assess complexity first按来源选择合适的工具Use appropriate tool per source尽可能并行化Parallelize when possible设置合理超时Set reasonable timeouts默认每页 15 秒优雅处理错误Handle errors gracefully如超时则跳过源、导航失败标记不可达后继续。综合策略Synthesis Strategy尽早组织发现Organize findings early快速识别模式Identify patterns quickly系统化消解冲突Resolve conflicts systematically渐进构建叙事Build narrative progressively保持证据链完整Maintain evidence chains。六、质量验证清单Quality Validation Checklist规划阶段Planning Phase目标定义清晰Clear objectives defined选择了合适的策略Appropriate strategy selected资源估算正确Resources estimated correctly成功标准已确立Success criteria established执行阶段Execution Phase所有规划搜索已完成All planned searches completed提取方法恰当Extraction methods appropriate多跳链逻辑合理Multi-hop chains logical置信度分数已计算Confidence scores calculated综合阶段Synthesis Phase所有发现已整合All findings integrated矛盾已消解Contradictions resolved证据链完整Evidence chains complete叙事连贯Narrative coherent交付阶段Delivery Phase格式适配受众Format appropriate for audience引用完整准确Citations complete and accurate包含视觉证据Visual evidence included置信度水平透明Confidence levels transparent七、配套资源与延伸阅读命令规范research.md——六阶段行为流、深度档位、输出标准与 CRITICAL BOUNDARIES全局配置RESEARCH_CONFIG.md——规划策略、跳配置、置信度打分relevance 权重 0.5 completeness 权重 0.5、自反思、提取路由、案例学习 Schema、错误处理与监控指标Agent 行为定义deep-research-agent.md——四类多跳推理模式实体扩展、时间递进、概念深化、因果链、自反思机制与证据管理规范Skill 协议SKILL.md——研究协议五步法与标准输出格式Key Findings / Analysis / Confidence / SourcesMCP 配置tavily.json、playwright.json、sequential.json、serena.json。结语把深度研究变成可复用方法论SuperClaude Framework 的深度研究工作流本质上是把科研工作者 调查记者的思维方式固化为可执行协议用 planning-only 处理清晰问题、用 intent-planning 消解歧义、用 unified 承载协作式复杂研究用多跳推理与并行执行控制搜索深度和成本用置信度分数与自反思驱动中途重规划用案例记忆让每次研究都比上次更高效。理解这套工作流的关键是把握策略选型 → 深度档位 → 工具路由 → 质量门禁这条决策链——它决定了从一条/sc:research命令到一份带证据链研究报告的全过程质量。赞分享开发工具CLIAI 技能/插件测试人工智能AI 评测【免费下载链接】SuperClaude_FrameworkA configuration framework that enhances Claude Code with specialized commands, cognitive personas, and development methodologies.项目地址https://gitcode.com/gh_mirrors/su/SuperClaude_Framework点击查看免费下载相关推荐SuperClaude Framework 的 /sc:research 深度研究命令自适应规划、多跳推理与证据合成实战指南SuperClaude Framework 的 /sc:research 深度研究命令自适应规划、多跳推理与证据合成实战指南 导读 /sc:researc开发工具CLIAI 技能/插件测试人工智能AI 评测SuperClaude Framework Deep Research Agent 实战指南自适应规划、多跳推理与证据驱动的研究工作流SuperClaude Framework Deep Research Agent 实战指南自适应规划、多跳推理与证据驱动的研究工作流 导读 SuperCla开发工具CLIAI 技能/插件测试人工智能AI 评测SuperClaude Framework 深度研究模式Deep Research Mode实战指南证据驱动与多跳推理的系统化调研方法论SuperClaude Framework 深度研究模式Deep Research Mode实战指南证据驱动与多跳推理的系统化调研方法论 导读 本文围绕开发工具CLIAI 技能/插件测试人工智能AI 评测创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考