CIA 红队四连问

AI被训练成讨好你,点子在发布前从没被真正攻击过。破解办法是把Claude设成有动机的敌人,用CIA红队四种技术在30分钟内主动把坏点子杀死,省下白烧的六个月。

作者
@heynavtoor
标签
决策研究
NO.013收录 2026-06-30

提示词 1:关键假设检查

You are now a CIA Red Team analyst. Do not evaluate whether my
idea is good. Your only job is to audit the assumptions it is
built on.
1. List every assumption my plan depends on. Not just the
obvious ones. The hidden ones I probably have not noticed.
Give me at least 10.
2. Classify each one into three tiers:
- LOAD-BEARING: If wrong, the entire plan fails.
- IMPORTANT: If wrong, the plan is weakened but survives.
- MINOR: If wrong, barely affects the outcome.
3. For each LOAD-BEARING assumption, answer: What specific
evidence would prove it wrong? If I cannot point to that
evidence, I am operating on faith, not analysis.

提示词 2:事前验尸

It is now 18 months from today. The idea I shared with you has
failed catastrophically. Not "did okay." Failed. Burned.
Embarrassing.
You are writing the honest post-mortem. Walk me through exactly
what went wrong, in chronological order, step by step.
Cover these stages:
- Month 1-3: The early warning signs we ignored
- Month 4-9: The decisions that made it worse
- Month 10-15: The point of no return
- Month 16-18: The collapse and what it cost
Be specific. Name the exact mistakes. Do not be vague.
End with one sentence: "The root cause was ___."

提示词 3:敌对竞争者

You are now a competitor with $100 million in funding, world-class
talent, and a personal motivation to crush the idea I just shared.
You have 90 days. You have unlimited budget. You hate me.
Write a 90-day attack plan to make my idea irrelevant.
Cover:
- Days 1-30: How you study, copy, and reposition
- Days 31-60: How you launch a better version
- Days 61-90: How you starve me of customers, attention, or talent
- What I am uniquely vulnerable to that I probably do not see
Be specific. Name tactics, not vague strategy.
End with one sentence: "The weakness that lets me win is ___."

提示词 4:1 星差评

You are now a customer who tried my idea and hated it. You spent
real money. You spent real time. You feel cheated.
Write the 1-star review that goes viral on Twitter and gets 10,000
likes. Be specific about what disappointed you. Be funny. Be brutal.
Use the voice of someone who is angry and articulate.
Then write 3 follow-up tweets from people quoting your review and
adding their own complaints.
End with one sentence: "The single thing that made me feel cheated was ___."