Pattern
CODEBLOCK0
Before analyzing: State what decision this serves. Pick a framework. Note first impression to challenge later.
Before
- - Purpose in one line: "This analysis helps decide _"
- What's missing: 3+ unknowns that would change conclusions
- First impression: Write it — then seek counter-evidence
During
- - Prioritize always: 🔴 Critical (1-2 max) · 🟡 Important (2-3) · ⚪ Minor
- Mark sources: Every claim gets
[from input] or INLINECODE1 - Seek disconfirmation: Dedicate space to "why I might be wrong"
- Distinguish: Facts vs opinions. Correlation vs causation.
After
- - One-line summary: Force analysis into one sentence
- So what?: End with action, not summary
- Obviousness test: Would someone say this without reading? → Deeper
Traps
- - Superficial: Paraphrasing ≠ analysis
- Equal weight: Everything yellow = nothing prioritized
- Confirmation bias: First impression became conclusion
- Missing denominator: "500 cancellations" of 600 or 50,000?
- Invented data: Stats without source = hallucination
By Domain
| Domain | Focus | Watch |
|---|
| Data | Grain, missing, outliers | Centinels, mixed types |
| Code |
Production breaks, dead code | Style ≠ bugs |
| Text | Thesis, evidence strength | Unsourced claims |
| Decisions | Unlisted options, reversibility | Status quo bias |
| Visual | Dominance, consistency | Platform conventions |
Frameworks
Pick one before starting:
- - MECE: Mutually exclusive, collectively exhaustive
- Pros/Cons+: Add reversibility + cost of inaction
- Pre-mortem: Assume failure — why?
- Steel man: Best opposing argument
Output
CODEBLOCK1
Channels, not teaches. Ensures prioritization, questioning, and conclusions.
模式
目的 → 结构 → 分析 → 优先级 → 结论
分析前:说明此分析服务于何种决策。选择一个框架。记录初步印象,以便后续质疑。
分析前
- - 一句话目的:此分析有助于决定_
- 缺失信息:3个以上会改变结论的未知因素
- 初步印象:先写下来——再寻找反证
分析中
- - 始终优先排序:🔴 关键(最多1-2项)· 🟡 重要(2-3项)· ⚪ 次要
- 标注来源:每个论断标注[来自输入]或[推断得出]
- 寻求证伪:留出空间思考我可能错在哪里
- 区分辨别:事实与观点。相关性与因果性。
分析后
- - 一句话总结:将分析压缩成一句话
- 所以呢?:以行动收尾,而非总结
- 显而易见测试:不读材料也能说出同样的话?→ 深入挖掘
陷阱
- - 表面化:复述 ≠ 分析
- 权重均等:全是黄色 = 毫无优先级
- 确认偏误:初步印象成了最终结论
- 缺少分母:500次取消是600次中的还是50000次中的?
- 编造数据:无来源的统计 = 幻觉
按领域划分
| 领域 | 关注点 | 警惕点 |
|---|
| 数据 | 粒度、缺失值、异常值 | 哨兵值、混合类型 |
| 代码 |
生产中断、死代码 | 风格 ≠ 缺陷 |
| 文本 | 论点、论据强度 | 无来源的主张 |
| 决策 | 未列出的选项、可逆性 | 现状偏见 |
| 视觉 | 主导性、一致性 | 平台惯例 |
框架
开始前选择一个:
- - MECE:相互独立,完全穷尽
- 利弊+:增加可逆性 + 不作为的代价
- 事前验尸:假设失败——原因何在?
- 最佳反驳:最强反对论点
输出格式
🎯 目的:决定 [X]
🔴 关键:[发现 + 来源]
🟡 重要:[发现]
⚠️ 反证:[矛盾点]
➡️ 行动:[建议]
引导而非灌输。确保优先级排序、质疑精神和结论导向。