给 AI 系统上线写一份 deploy checklist：不是写"要小心"，而是列出那几个一旦漏掉就会出事、又最容易在赶版本时跳过的关口——"回滚脚本已验证""token 成本上限已设""prompt 版本已打 tag""敏感数据已脱敏"。控制在 7 项以内。在家也成立：孩子上学出门清单贴在门口，把"靠妈妈记"换成"孩子自己核对"，顺便把责任和能力一起交还给孩子，这正是清单的社会功能在家庭里的版本。
English Summary
Checklists — they don't help you remember what you don't know; they guard against skipping what you do know under fatigue, pressure, or distraction. They target execution failures on known items, not knowledge gaps. Two types: DO-CONFIRM (act from memory, pause at checkpoints to verify) for experts, and READ-DO (read-then-do) for novices. Keep them short — 5–9 "killer items" per pause point; cramming in everything makes a checklist nobody uses. The hidden lever is social, not mnemonic: a checklist creates a legitimate pause and gives lower-status members permission to flag a missed step. It's a socio-technical intervention, not a memo.
AI Prompts
中文提示词
我要为 [流程/操作，如某系统上线、某次手术准备、某类家务] 设计一份检查清单。请：
① 区分这个场景该用 DO-CONFIRM 还是 READ-DO，并说明理由；
② 只挑出 5–9 个"killer items"——漏掉会致命、且在忙乱中最易被跳过的步骤，删掉所有"锦上添花"项；
③ 标出每个暂停点的位置，以及哪一项是给"地位较低成员"开口纠错用的。
English Prompt
I want to design a checklist for [process/operation]. Please:
1. Decide whether this case calls for DO-CONFIRM or READ-DO, and explain why.
2. Select only 5–9 "killer items" — steps that are fatal if skipped and most likely to be dropped under pressure; cut every nice-to-have.
3. Mark the pause points, and flag which item exists to give a lower-status member permission to catch an error.
非平凡点：① 它真正破解的是群体一致性压力。在正常会议里说"我担心这事会黄"=扫兴、不合群、像在唱衰团队，于是没人说。预先验尸把"找失败原因"变成被正式授权、甚至要比谁找得多的任务——它把质疑从社会惩罚反转成了社会奖励。② 心理机制叫前瞻性后见（prospective hindsight）：把"可能会失败"改写成"已经失败了"，大脑就从抽象概率切换到具体叙事模式。研究显示，这种"事情已成定局"的措辞能让人识别未来结果原因的能力提升约 30%。③ 它是逆向思维（见 Day 1）的团队化、时间化落地版——不止"反过来想"，而是给反过来想配上流程、角色和时间锚点。
经典例子
这一方法由决策研究者 Gary Klein 提出、并经 Kahneman 大力推荐进入主流。做法极简：项目启动会上，主持人说"现在是一年后，这个项目彻底失败了，请各位用 2 分钟独立写下所有可能的死因"，然后轮流朗读。它之所以有效，是因为独立书写避开了从众，而"已经失败"的设定让最资深、最该被听见的悲观声音终于敢出声。
场景 · BigCat
发布一个 AI agent 产品前，召集团队做预先验尸："半年后这产品死了，为什么？"——你会比任何乐观路演都更快听到"幻觉拖垮了信任""token 成本压垮了毛利""用户根本不会写 prompt"这些真问题。个人也能做单人版：重大技术选型前，写一段"这个架构两年后被全员唾弃的悼词"。它甚至与佛学的"念死"同构——观想终局来倒逼当下的优先级，让真正重要的浮上来。
English Summary
Pre-mortem — before launching, assume the project has already failed catastrophically, then have everyone independently write down why. What it really defeats is conformity pressure: in a normal meeting, voicing "I'm worried this will fail" reads as disloyal, so no one does; the pre-mortem reframes finding failure causes as an authorized, even competitive task, flipping dissent from social punishment to social reward. The mechanism is prospective hindsight — rewriting "might fail" as "has failed" shifts the brain from abstract probability to concrete narrative, improving the ability to identify causes of outcomes by roughly 30%. It's the team-and-time-anchored implementation of inversion.
AI Prompts
中文提示词
我即将推进 [项目/决策/计划]，目标是 [描述]。请帮我做一次预先验尸：
① 设定"一年后它已经彻底失败"，列出 6–8 个最可能的"死因"，按"杀伤力 × 发生概率"排序；
② 指出其中哪些是团队此刻因为"不想扫兴"而不会主动说出口的；
③ 针对排名前 2 的死因，各给一个现在就能采取的预防动作。
English Prompt
I'm about to pursue [project/decision/plan], aiming to [describe]. Run a pre-mortem:
1. Assume it's one year later and the plan has failed completely; list 6–8 likely "causes of death," ranked by impact × probability.
2. Identify which of these the team would avoid voicing now to not seem like a downer.
3. For the top 2 causes, give one preventive action I can take right now for each.
Red Teaming — a group formally tasked with attacking your plan, system, or conclusion. Unlike a one-shot pre-mortem, it's a standing, adversarial role. Its real value isn't the specific flaw found but institutionalizing adversarial thinking: it turns "challenging authority" from an act of personal courage into an authorized job, erasing the social cost of dissent. It differs from a devil's advocate (a temporary, often performative contrarian) by requiring structural independence and a genuine incentive to win. The fatal failure mode is co-optation — a red team that reports to or seeks to please the blue team is worthless. Highly current in the LLM era via AI red teaming for safety and alignment.
AI Prompts
中文提示词
请你扮演我的红队，任务是击溃以下方案，而不是改良它：[描述方案/结论]。请：
① 找出 3 个最致命的隐含假设，并说明各自在什么条件下会崩；
② 设计一个最可能让它失败的对抗场景（攻击者视角）；
③ 诚实评估：如果我反驳你，你最强的反击是什么？最后说明哪个攻击我无法轻易化解。
English Prompt
Act as my red team. Your job is to defeat the following, not improve it: [describe plan/conclusion]. Please:
1. Surface the 3 most fatal hidden assumptions and the conditions under which each collapses.
2. Design the adversarial scenario most likely to make it fail (attacker's view).
3. Honestly assess: if I push back, what's your strongest counter? End by naming the one attack I cannot easily neutralize.
Decision Journal — at the moment of a major decision, record four things: what you expect to happen, your core reasoning, your emotional state, and your confidence level (e.g., "70% sure"). Later, compare against the real outcome. It defeats two hidden killers: hindsight bias ("I knew it all along") and outcome bias (judging decision quality by results). Without it, your brain rewrites the memory and you only learn the emotion of outcomes, never decision quality. Core distinction: a good decision ≠ a good outcome — under uncertainty you must judge process, not result. The journal is your only data source for separating skill from luck, and a training ground for Bayesian calibration: check whether things you marked "70% sure" actually happened ~70% of the time.
AI Prompts
中文提示词
这是我决策当时记录的日记条目：预期 [X]，理由 [Y]，情绪 [Z]，置信度 [N%]。现在的真实结果是 [描述]。请帮我复盘：
① 分离实力与运气——结果的好坏有多少来自我的理由正确，多少来自偶然？
② 我标的置信度是否校准（过度自信还是过度保守）？
③ 我的情绪状态是否系统性影响了这次判断，未来该设什么触发器提醒自己？
English Prompt
Here is the journal entry from when I decided: expectation [X], reasoning [Y], emotion [Z], confidence [N%]. The actual outcome is now [describe]. Please review:
1. Separate skill from luck — how much of the outcome came from correct reasoning vs chance?
2. Was my stated confidence calibrated (over- or under-confident)?
3. Did my emotional state systematically bias this decision, and what trigger should I set to flag it next time?