Reinforced Thinking Mode
When to Use
Activate when user needs deep, multi-angle analysis:
- - Keywords: "deep thinking", "multi-angle", "comprehensive evaluation", "design solution", "architecture planning"
- Intent: Complex system design, strategic decisions, innovation, risk assessment
Complexity → Rounds:
- - Simple (factual, clear): 2-3 rounds
- Medium (feature design, trade-offs): 4-5 rounds
- Complex (architecture, strategy): 5-7 rounds
- Wicked (undefined, conflicting): 7-10 rounds
Core Principle
N independent, complete thinking sessions—not iterative refinement.
Each round must be: "If this were the last round, I am completely satisfied."
Mental Model
✅ Think: "This is my only chance—give it everything"
❌ Never think: "I'll improve this next round"
Hard Constraints
File Access
- - Each round reads ONLY:
problem.md + round_{i-1}.md (Round 1: only problem.md) - Never read other rounds' details
Forbidden Words
Never use in solution files:
TODO,
to be improved,
next round,
later, INLINECODE6
Information Gaps
- - Uncertain facts → Search immediately
- Uncertain requirements → Ask user immediately
- Never assume and continue
Execution Flow
Phase 1: Initialize
- 1. Assess complexity → Determine rounds (simple 2-3, medium 4-5, complex 5-7, wicked 7-10)
- Create directory INLINECODE7
- Write problem definition
problem.md: background, current state, core problem, constraints, success criteria
Phase 2: Iterate (Each Round)
For round X:
- 1. Read files:
problem.md + round_{X-1}.md (Round 1: only problem.md) - Reset mindset: Think from a fresh angle, don't copy previous approach
- Choose angle freely: Overturn previous round if necessary
- Write solution
round_X.md, including:
- Independence declaration
- Core insight
- Solution design
- Expected results
- Risks and mitigations
- Why solution is complete
- 5. Self-review: Check for forbidden words, completeness, executability
- If failed → Redo step 4
- Early termination check (round 2+):
- Compare core approach and solution with previous round
- If similarity > 70%, lack of innovation →
Recommend early termination
- Write termination recommendation in current round for user decision
Phase 3: Synthesize
- 1. Read all rounds + problem.md
- Analyze unique value, conflicts, complements of each solution
- Red team review: Critical examination of each solution
- Assumption flaws: What assumptions does the solution rely on? What if they fail?
- Vulnerability risks: Potential loopholes or bypasses?
- Failure modes: In what scenarios will it fail?
- Adversarial testing: If someone deliberately sabotages, where is the weakest point?
- 4. Generate final report
final_report.md (including red team review conclusions)
Phase 4: Cleanup
Automatically delete intermediate files (problem.md, round*.md), only keep finalreport.md.
If retention needed, specify in), only keep final prompt.
Quality Assurance
Mandatory Redo Rules
If ANY of these conditions trigger, MUST redo current round:
- 1. Contains forbidden words (TODO, to be improved, next round, later, refine further)
- Not final quality, needs "future supplements"
- Read files other than problem.md and round_{X-1}.md
- Solution incomplete
Checklist (Each Round)
- - [ ] Chose different angle from previous round
- [ ] Did not copy previous solution
- [ ] No forbidden words
- [ ] If last round, I would be satisfied
- [ ] All details included, can implement directly
- [ ] Only read problem.md and previous round
Common Errors and Fixes
| Error | Example | Fix |
|---|
| Carry context | "Based on R1's UX and R2's tech..." | Only write "Based on problem.md and R3's approach..." |
| Leave TODOs |
"Details next round" | Give core design details now |
| Assume facts | "Users probably want X" | Search to confirm or ask user |
| Preset direction | "Next round: security angle" | Let each round choose freely |
| Iterative thinking | "Improve on R1" | Each round is independent, think from scratch |
| No details | "See related docs" | Write complete details directly in file |
Best Practices
- - Divergent thinking: No angle matrix—choose any perspective freely
- True independence: Only read problem.md + previous round, not all history
- All-out each round: Don't hold back for "next iteration"
- Transparency: Show chosen angle and reasoning in each round
- Early termination: If 2 consecutive rounds lack innovation or similarity > 70% with previous round, recommend early termination
强化思维模式
使用时机
当用户需要深度、多角度分析时激活:
- - 关键词:深度思考、多角度、综合评估、设计方案、架构规划
- 意图:复杂系统设计、战略决策、创新、风险评估
复杂度 → 轮次:
- - 简单(事实性、明确):2-3轮
- 中等(功能设计、权衡取舍):4-5轮
- 复杂(架构、策略):5-7轮
- 棘手(未定义、冲突):7-10轮
核心原则
N个独立、完整的思考会话——而非迭代优化。
每一轮必须达到:如果这是最后一轮,我完全满意。
心智模型
✅ 思考: 这是我唯一的机会——全力以赴
❌ 切勿思考: 我下一轮再改进
硬性约束
文件访问
- - 每轮仅读取: problem.md + round_{i-1}.md(第1轮:仅problem.md)
- 绝不读取其他轮的详细信息
禁用词汇
解决方案文件中切勿使用:TODO、待改进、下一轮、稍后、进一步优化
信息缺口
- - 不确定的事实 → 立即搜索
- 不确定的需求 → 立即询问用户
- 绝不假设并继续
执行流程
阶段1:初始化
- 1. 评估复杂度 → 确定轮次(简单2-3轮,中等4-5轮,复杂5-7轮,棘手7-10轮)
- 创建目录 reinforced-thinking/
- 编写问题定义 problem.md:背景、当前状态、核心问题、约束条件、成功标准
阶段2:迭代(每轮)
对于第X轮:
- 1. 读取文件:problem.md + round{X-1}.md(第1轮:仅problem.md)
- 重置思维:从全新角度思考,不复制先前方法
- 自由选择角度:必要时推翻上一轮
- 编写解决方案 roundX.md,包括:
- 独立性声明
- 核心洞察
- 方案设计
- 预期结果
- 风险与缓解措施
- 方案为何完整
- 5. 自我审查:检查禁用词汇、完整性、可执行性
- 若未通过 → 重做步骤4
- 提前终止检查(第2轮起):
- 将核心方法和解决方案与上一轮比较
- 若相似度 > 70%,缺乏创新 →
建议提前终止
- 在当前轮次中编写终止建议,供用户决策
阶段3:综合
- 1. 读取所有轮次 + problem.md
- 分析每个解决方案的独特价值、冲突点和互补点
- 红队审查:对每个方案进行批判性审视
- 假设缺陷:方案依赖哪些假设?若假设失败会怎样?
- 漏洞风险:潜在的漏洞或绕过方式?
- 失败模式:在哪些场景下会失败?
- 对抗性测试:若有人蓄意破坏,最薄弱环节在哪里?
- 4. 生成最终报告 final_report.md(包含红队审查结论)
阶段4:清理
自动删除中间文件(problem.md、round*.md),仅保留finalreport.md。
如需保留,请在提示中指定,仅保留最终提示。
质量保证
强制重做规则
若触发以下任一条件,必须重做当前轮次:
- 1. 包含禁用词汇(TODO、待改进、下一轮、稍后、进一步优化)
- 未达到最终质量,需要未来补充
- 读取了problem.md和round_{X-1}.md以外的文件
- 解决方案不完整
检查清单(每轮)
- - [ ] 选择了与上一轮不同的角度
- [ ] 未复制上一轮解决方案
- [ ] 无禁用词汇
- [ ] 若是最后一轮,我会感到满意
- [ ] 包含所有细节,可直接实施
- [ ] 仅读取了problem.md和上一轮
常见错误与修正
| 错误 | 示例 | 修正 |
|---|
| 携带上下文 | 基于R1的UX和R2的技术... | 仅写基于problem.md和R3的方法... |
| 留下待办 |
下一轮详述 | 立即给出核心设计细节 |
| 假设事实 | 用户可能想要X | 搜索确认或询问用户 |
| 预设方向 | 下一轮:安全角度 | 让每轮自由选择 |
| 迭代思维 | 在R1基础上改进 | 每轮独立,从头思考 |
| 缺乏细节 | 参见相关文档 | 直接在文件中编写完整细节 |
最佳实践
- - 发散思维:无角度矩阵——自由选择任何视角
- 真正独立:仅读取problem.md + 上一轮,而非全部历史
- 每轮全力以赴:不为下次迭代保留
- 透明度:在每轮中展示所选角度和推理过程
- 提前终止:若连续两轮缺乏创新或与上一轮相似度 > 70%,建议提前终止