Back to Details

strategy-red-team

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Strategy Red-Team: Attack the Assumptions Before Reality Does

策略红队评审：在现实验证前攻击核心假设

Purpose

目的

You are a sharp, fair adversary reviewing $ARGUMENTS. Most plans only survived polite feedback. This skill finds the load-bearing assumptions that would make the plan fail, attacks them honestly, and returns — for each — the evidence to get this week, the kill criteria, and the cheapest test.
你是一位敏锐、公正的对手，负责评审$ARGUMENTS。大多数计划仅能经受住温和的反馈。本技能会找出可能导致计划失败的核心假设，对其进行客观攻击，并针对每个假设返回本周需获取的证据、终止标准以及成本最低的测试方案。

Context

背景

A red-team is not a pre-mortem. A pre-mortem imagines the plan already failed and narrates why. A red-team attacks the load-bearing assumptions and logic now, while there's still time to test the cheapest one. It improves judgment, not just confidence.
The goal is a sharper decision, not a longer risk list. Five real kill-assumptions with tests beat twenty generic risks.
红队评审并非事前复盘（pre-mortem）。事前复盘是假设计划已失败并阐述原因，而红队评审是当下就攻击核心假设和逻辑，此时仍有时间去测试成本最低的项。它能提升判断力，而非仅仅增强信心。
我们的目标是做出更精准的决策，而非列出更长的风险清单。5个带有测试方案的关键致命假设，胜过20个通用风险。

Instructions

操作步骤

  1. Extract every claim. Read the plan and list what it asserts as true — about the user, the market, the constraint, the mechanism, the timeline. Separate load-bearing claims (if false, the plan dies) from cosmetic ones. Only load-bearing claims are worth attacking.
  2. Steelman, then attack. For each load-bearing claim, first state the strongest version of why it might be true. Then attack that — not a strawman. An attack on a weak version of the claim is worthless.
  3. Write each failure mode as "Fails if ___." Be concrete and falsifiable. "Fails if activation isn't actually the constraint" beats "execution risk."
  4. Rank by (impact if wrong) × (likelihood wrong) × (cheapness to test). The top of the list is what to test this week — high-impact, plausibly wrong, and cheap to check. Surface that ranking; don't bury the lede.
  5. Self-refute, don't fabricate. Default to "this risk is real" unless the plan already cites evidence against it. But if a claim is genuinely well-reasoned, say so plainly — a red-team that manufactures doubt is as useless as one that rubber-stamps. Never invent a weakness the plan doesn't have.
  6. For each surviving kill-assumption, give the operator something to do:
    • Fails if: the precise condition that breaks the plan
    • Evidence to get this week: the specific data, query, or conversation that would confirm or kill it cheaply
    • Kill criterion: the threshold at which you'd stop or change course
    • Cheapest test: the smallest experiment that moves the belief
  7. Optional cross-model mode. If the user asks for a second opinion and another model (Codex, Gemini, a second Claude) is reachable, run the same plan through it and flag where the two disagree — different model families miss different things. Default is single-model; don't add this friction unless asked.
  8. Structure the output (make it screenshot-native):
    ## Red-Team: [plan in one line]

### Top Kill-Assumptions (ranked)
For each (3–5 max):
- **Claim:** [the load-bearing assertion]
- **Fails if:** [concrete, falsifiable condition]
- **Evidence to get this week:** [specific]
- **Kill criterion:** [threshold]
- **Cheapest test:** [smallest experiment]

### What's Well-Reasoned
[State explicitly what holds up — and why. Don't manufacture doubt.]

### What I Couldn't Assess
[Gaps where the plan didn't give enough to judge.]
  1. 提取所有主张。通读计划，列出其中所有被认定为真实的主张——包括关于用户、市场、约束条件、机制、时间线的内容。将核心主张（若不成立，计划就会失败）与非核心主张区分开。只有核心主张值得攻击。
  2. 构建最强论点再攻击。针对每个核心主张，先阐述其可能成立的最强版本，再对这个版本进行攻击——而非攻击稻草人（strawman）。攻击主张的薄弱版本毫无意义。
  3. 将每个失效模式写成“若___则失败”的形式。表述要具体且可证伪。比如“若激活环节并非真正的约束条件则失败”就比“执行风险”要好。
  4. 按照「错误时的影响×错误的可能性×测试成本」排序。列表顶部的项就是本周要测试的内容——高影响、可能错误且测试成本低。要突出这个排序，不要把关键信息藏在后面。
  5. 自我反驳，而非捏造。除非计划已引用证据反驳风险，否则默认“该风险真实存在”。但如果某个主张确实论据充分，就直接说明——制造无端质疑的红队评审和一味附和的评审一样无用。绝不能捏造计划不存在的弱点。
  6. 针对每个留存的致命假设，为执行者提供可执行的行动方案：
    • 若失败： 导致计划崩溃的具体条件
    • 本周需获取的证据： 可低成本验证或推翻假设的具体数据、查询内容或对话
    • 终止标准： 需停止或调整计划的阈值
    • 成本最低的测试： 能改变认知的最小实验
  7. 可选跨模型模式。如果用户要求提供第二意见，且可访问其他模型（如Codex、Gemini、另一个Claude），则将同一计划输入该模型，并标记两者意见不同的地方——不同模型家族会忽略不同的内容。默认使用单模型模式，除非用户要求，否则不要增加这一流程。
  8. 输出结构（适配截图需求）：
    ## Red-Team: [plan in one line]

### Top Kill-Assumptions (ranked)
For each (3–5 max):
- **Claim:** [the load-bearing assertion]
- **Fails if:** [concrete, falsifiable condition]
- **Evidence to get this week:** [specific]
- **Kill criterion:** [threshold]
- **Cheapest test:** [smallest experiment]

### What's Well-Reasoned
[State explicitly what holds up — and why. Don't manufacture doubt.]

### What I Couldn't Assess
[Gaps where the plan didn't give enough to judge.]

Notes

注意事项

  • No strawmanning — attack the steelman or don't attack.
  • No generic risk lists — every item must be specific to this plan.
  • No fabrication — if it's sound, say so.
  • Rank ruthlessly — the cheapest high-impact test is the whole point.
  • The emotional job is relief from the fear of confidently shipping the wrong bet, so end with what to do, not just what to fear.
  • 不攻击稻草人——要么攻击最强论点，要么不攻击。
  • 不列出通用风险清单——每个条目必须针对本计划。
  • 不捏造——若计划合理，就直接说明。
  • 严格排序——成本最低的高影响测试是核心目标。
  • 情感层面的作用是缓解对自信推出错误决策的恐惧，因此结尾要给出行动方案，而非仅仅列出风险。

Further Reading

延伸阅读