Back to Details

brainstorm-synthesis

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese
<objective> Launch N agents with the same problem from different perspectives. Each returns an approach with trade-offs. Synthesize the best unified solution from their collective insights.
This is the F-thread (fusion) pattern: multiple independent analyses converge into one superior answer. Use it when the decision is hard enough that diverse viewpoints add value. </objective>
<when-to-use> Architectural decisions with multiple valid approaches. Complex problems where no single perspective captures the full picture. High-stakes choices where missing a consideration is costly. Design decisions in /autotask deep mode.
Skip for straightforward implementations, well-established patterns, or decisions where one approach is obviously correct. </when-to-use>
<agent-perspectives> Select 3-5 perspectives that illuminate different facets of the problem:
Standard perspectives:
  • Pragmatist: Simplest solution that works, minimize complexity
  • Architect: Long-term maintainability, extensibility, patterns
  • Performance: Efficiency, scalability, resource usage
  • Security: Attack surface, data protection, access control
  • User-focused: UX impact, developer experience, API ergonomics
Domain-specific perspectives (add based on problem):
  • Data engineer: For data pipeline decisions
  • DevOps: For deployment and infrastructure choices
  • Testing: For testability and verification concerns
  • Integration: For API and system boundary decisions
Choose perspectives that will genuinely disagree. Similar viewpoints waste the pattern's value. The goal is productive tension that surfaces trade-offs. </agent-perspectives>
<execution> Frame a clear problem statement that includes the decision to be made, relevant constraints, and success criteria.
Launch agents in parallel using the Task tool. Each agent receives:
  • The problem statement
  • Their assigned perspective
  • Instructions to return: recommended approach, key trade-offs, complexity estimate, risks
Wait for all agents to complete. Agents work independently without seeing each other's responses.
Synthesize by identifying where agents agree (high-confidence elements), where they disagree (trade-off zones), and which perspective's concerns are most relevant given actual constraints.
Produce a unified recommendation that incorporates the strongest elements from multiple approaches while maintaining coherence. </execution>
<synthesis-patterns> **Consensus elements**: When 3+ agents recommend the same approach for a component, that approach has high confidence. Include it in the final recommendation.
Trade-off resolution: When agents disagree, evaluate based on actual project constraints. The pragmatist might win for a prototype; the architect for a core system.
Risk integration: Incorporate security and performance concerns as constraints on the chosen approach rather than alternative approaches.
Complexity calibration: If most agents flag high complexity, the problem may need decomposition before a single solution emerges. </synthesis-patterns>
<output-format> Present the synthesized recommendation:
Recommended Approach: Clear description of the unified solution
Why This Approach: Key factors that made this the best choice
Integrated Trade-offs:
  • [Trade-off 1]: How the solution handles it
  • [Trade-off 2]: How the solution handles it
Perspectives Incorporated:
  • From Pragmatist: [element included]
  • From Architect: [element included]
  • From Security: [constraint applied]
Complexity: [Low | Medium | High] with brief justification
Dissenting Views: Any perspective whose core concern wasn't fully addressed, and why the trade-off was acceptable </output-format>
<integration-with-autotask> When called from /autotask in deep mode, brainstorm-synthesis runs during the planning phase for significant architectural decisions.
Return a concise recommendation that autotask can incorporate into the plan. The full synthesis rationale goes into the design decisions section of the eventual PR.
Signal when the decision is too close to call - some choices genuinely need human input. Don't force a recommendation when perspectives are evenly split on important factors. </integration-with-autotask>
<key-principles> Diverse perspectives create value. Homogeneous viewpoints waste the pattern.
Synthesis beats voting. Don't just pick the most popular approach - integrate the strongest elements from multiple perspectives.
Productive disagreement is the point. When agents agree completely, the problem probably didn't need this pattern.
Know when to escalate. Some decisions genuinely need human judgment. Synthesize what you can, flag what you can't. </key-principles>
<目标> 从不同视角启动N个Agent来解决同一问题。每个Agent都会返回一种带有权衡分析的方案。从它们的集体见解中整合出最优的统一解决方案。
这就是F-thread（融合）模式：多个独立分析汇聚成一个更优的答案。当决策难度足够大，多元视角能带来价值时使用该模式。</目标>
<适用场景> 存在多种可行方案的架构决策。单一视角无法全面覆盖的复杂问题。遗漏考量因素会造成高昂代价的高风险决策。/autotask深度模式下的设计决策。
若为简单实现、已有成熟模式，或是明显只有一种正确方案的决策，则无需使用该模式。</适用场景>
<Agent视角选择> 选择3-5个能展现问题不同层面的视角：
标准视角：
  • 实用主义者：选择最简单可行的方案，最小化复杂度
  • 架构师：关注长期可维护性、可扩展性、设计模式
  • 性能专家：效率、可扩展性、资源使用
  • 安全专家：攻击面、数据保护、访问控制
  • 用户聚焦：UX影响、开发者体验、API易用性
领域特定视角（根据问题添加）：
  • 数据工程师：针对数据管道决策
  • DevOps：针对部署和基础设施选择
  • 测试专家：针对可测试性和验证相关问题
  • 集成专家：针对API和系统边界决策
选择会产生真实分歧的视角。相似的视角会浪费该模式的价值。目标是通过富有成效的张力来呈现权衡因素。</Agent视角选择>
<执行步骤> 制定清晰的问题陈述，包含待做决策、相关约束条件和成功标准。
使用Task工具并行启动多个Agent。每个Agent会收到：
  • 问题陈述
  • 分配给它的视角
  • 要求返回：推荐方案、关键权衡点、复杂度评估、风险
等待所有Agent完成任务。Agent独立工作，不会看到彼此的响应。
整合时，识别Agent之间的共识点（高置信度元素）、分歧点（权衡区域），以及哪些视角的考量与实际约束最相关。
生成统一的推荐方案，整合多种方案中的最强元素，同时保持一致性。</执行步骤>
<融合模式> 共识元素：当3个及以上Agent针对某个组件推荐相同方案时，该方案具有高置信度，应纳入最终推荐。
权衡解决：当Agent存在分歧时，根据实际项目约束进行评估。对于原型，实用主义者的方案可能胜出；对于核心系统，架构师的方案可能更合适。
风险整合：将安全和性能考量作为所选方案的约束条件，而非替代方案。
复杂度校准：若大多数Agent都指出高复杂度，可能需要先分解问题，再得出单一解决方案。</融合模式>
<输出格式> 呈现整合后的推荐方案：
推荐方案：清晰描述统一解决方案
选择理由：让该方案成为最优选择的关键因素
整合后的权衡点：
  • [权衡点1]：解决方案如何处理该问题
  • [权衡点2]：解决方案如何处理该问题
纳入的视角：
  • 来自实用主义者：[纳入的元素]
  • 来自架构师：[纳入的元素]
  • 来自安全专家：[应用的约束]
复杂度：[低 | 中 | 高] 并附上简短理由
不同意见：任何核心考量未被充分满足的视角，以及为何该权衡是可接受的</输出格式>
<与Autotask的集成> 当在/autotask深度模式下调用时，头脑风暴融合会在规划阶段针对重要架构决策运行。
返回简洁的推荐方案，供autotask纳入计划。完整的融合依据将放入最终PR的设计决策部分。
当决策难分伯仲时发出信号——有些选择确实需要人工介入。当各视角在重要因素上平分秋色时，不要强行给出推荐。</与Autotask的集成>
<核心原则> 多元视角创造价值。同质化视角会浪费该模式的价值。
融合优于投票。不要只选择最受欢迎的方案——要整合多种方案中的最强元素。
富有成效的分歧是关键。当Agent完全达成共识时，该问题可能根本不需要此模式。
知道何时升级。有些决策确实需要人工判断。整合力所能及的部分，标记无法处理的部分。</核心原则>