如何让 Claude 穿越时空预测项目失败（事前验尸技能）

Ole Lehmann 2026-05-26 8.7 KB 317.4K 248 𝕏 文章列表

文章介绍了一种利用 Claude AI 进行的“事前验尸”技术，该方法通过模拟未来项目已失败的情境，帮助用户识别潜在风险和隐藏假设。作者详述了该技能的工作原理、背后的心理学依据以及具体的应用场景，旨在帮助用户在做出高风险决策前进行压力测试。

Claude 事前验尸决策分析风险控制 AI工作流 Agent技能方法论

# How to make Claude time-travel and tell you how your plan dies (the Pre-Mortem Skill) **作者**: Ole Lehmann **日期**: 2026-05-04T18:49:18.000Z **来源**: [https://x.com/itsolelehmann/status/2051373618727469152](https://x.com/itsolelehmann/status/2051373618727469152) --- ![Image](/doc/2246/image/img_06e34051d3d4.jpg) POV: Claude time-traveled 6 months into the future and told you exactly how your next move failed. It's called a premortem. And it might be the single most useful thing Claude does for high-stakes decisions. Here's the exact skill, and how to run one on your own plan in 5 minutes. P.S. If you want more AI workflows like this one delivered free to your inbox every week, join 37k readers here: aisolo.beehiiv.com/subscribe ## The problem this fixes Claude is a helpful assistant. That's literally the product. That's how it was trained: to be helpful, agreeable, optimistic by default. When you bring it a plan, of course it’ll hype you up and lean toward reasons it'll work. So you walk away feeling confident. You spend the next 6 weeks building it. You move money around to fund it. The weekends you'd normally spend with people you love disappear into "one more push," because this is going to be the one. Then 2 months in, it blows up in your face. The launch hits half its target and refund requests start rolling in by week 2. Or the hire you were so sure about turns out to be the wrong fit, and you've already burned $40k of runway you can't get back. But you can't get the time back. And you can't get the money back. In hindsight, the failure was obvious. You just never stress-tested it, because Claude told you it was solid. But what if you didn’t have to learn “the hard way”? ## Where this comes from The fix is a 30-year-old technique called a premortem. Gary Klein, a research psychologist, invented it in the 80s. He noticed teams kept making decisions everyone privately thought were bad, but nobody would say it out loud. His move was to flip the timeline. You tell the team it's already 6 months later and the project failed. Then everyone describes how it died. That single shift gave people permission to surface the failure modes they'd been swallowing. Daniel Kahneman, the Nobel-winning psychologist behind Thinking Fast and Slow, called it the single most valuable decision-making technique he knew. Researchers at Wharton and Cornell later studied it formally and named it prospective hindsight. Their finding: imagining a future failure makes people significantly better at identifying the actual reasons it would fail than asking "what could go wrong?" does. The technique now runs at the highest-stakes decision tables in the world: - Goldman Sachs uses premortems before major investment committee decisions and IPO underwriting calls. Partners imagine the deal blowing up before they sign. - Google uses them before major product launches. Teams walk through the version of the launch where it failed publicly and reverse-engineer what would have caused it. I took the technique and turned it into a Claude skill that runs the whole thing for you in about 5 minutes. ## How the Pre-Mortem Skill works The skill is more than a clever prompt. Here's what runs under the hood. 1. It gathers context. The skill scans the conversation and your workspace (memory files, project notes, anything you've referenced) for what the plan is, who it's for, and what success looks like. If you've been discussing the plan with Claude already, "premortem this" is enough on its own. 2. It generates the raw failure modes. Claude sets the frame ("it's 6 months from now, this plan is dead, here's why") and lists every genuine reason it could have died. The number is whatever's real for your specific plan. Could be 4, could be 9. 3. It spawns one investigator per failure, in parallel. The part that makes the output feel real. Each failure reason gets its own sub-agent that goes deep on it independently, all running at the same time. Each agent writes a short case study of how that specific failure played out, the underlying assumption behind it, and the early warning signs you'd see. This is my favorite part because you can really visualize how it failed. 4. It pulls everything into a synthesis: - The most likely failure (the one to focus on first) - The most dangerous failure (the one worth insuring against, even if it's less likely) - The single biggest hidden assumption you're making - A revised plan with concrete changes mapped to specific failure modes - A pre-launch checklist of 3-5 things to verify before you ship 5. It hands you a visual report. Plus a markdown transcript if you want to dig back into any single agent's reasoning. Here's an example. You tell Claude you're partnering with another creator on a joint product, 50/50 revenue split, launching in 3 months. The skill comes back with 5 failure modes: 1. Effort imbalance: one of you carries the load while the other gets pulled into another project 2. Audience mismatch: the trust each of your audiences has in you individually doesn't transfer to the joint product 3. Vision drift: 6 weeks in, you disagree on what the product should be and every decision drags 4. Decision deadlock: 50/50 means no tiebreaker, hard calls don't get made 5. No exit clause: one of you wants out in 9 months and there's nothing in writing Each one gets its own deep-dive agent. Here's what the effort imbalance one returns: First 6 weeks are great. By month 3, your partner is heads-down on another project and replies take 48 hours. You're carrying 80% of the work. Launch slips by 6 weeks. When sales come in, the 50/50 split feels brutal. Early warning sign: more than 3 unanswered messages in a week, or a missed deadline you only find out about after the fact. Then in the synthesis, the biggest hidden assumption: equal partnership and equal effort are the same thing. You set up a 50/50 deal expecting the work to balance itself out. The kind of line you'd never write down on your own. The skill drags it into daylight before it costs you 4 months of work. ## How to run one yourself 1. Grab the skill here. Drop it into your skills folder in Claude Code or Cowork. 2. Open Claude with a plan in mind. A launch, a hire, a pricing change, a content bet. Anything where being wrong costs you weeks or real money. 3. Type the trigger: "premortem this" The skill takes it from there. If it has enough context from the conversation, it runs. If it doesn't, it asks you focused questions one at a time until it has enough to run a useful premortem. No brief required. 4. Let it run. The skill gathers context, generates the failure modes, spawns the agents in parallel, and saves the HTML report and transcript to your workspace. That's the whole loop. ## What you can use this for Anything where being wrong has a real cost. The skill doesn't care if it's a business decision, a career move, a creative bet, or a personal commitment. Same mechanism every time: name the plan, declare it dead, find out why. Real examples worth running: - A product launch, pricing change, or anything else with money on the line - A hire, partnership, or joint venture you're about to commit to - Career moves like taking a new job, raising money, or going full-time freelance - A repositioning, rebrand, or strategy pivot - Creative bets like starting a podcast, writing a book, or going all-in on a new content channel - Investment decisions, big buys, long-term contracts you're about to sign - Major personal commitments: moving cities, buying a house, taking on a year-long project that'll eat your weekends If you can describe the plan in 3 sentences and the cost of being wrong is more than a few hours, premortem it before you commit. ## Run one tonight Pick the next move you're about to commit to. Premortem it before you go to bed tonight. 5 minutes, and you'll see the failure before you live it. P.S. If you want more AI workflows like this one, I send them to 37k readers every week for free. Plus you get a free Claude Cowork course as a bonus when you join: aisolo.beehiiv.com/subscribe ## 相关链接 - [Ole Lehmann](https://x.com/itsolelehmann) - [@itsolelehmann](https://x.com/itsolelehmann) - [53K](https://x.com/itsolelehmann/status/2051373618727469152/analytics) - [aisolo.beehiiv.com/subscribe](https://aisolo.beehiiv.com/subscribe) - [aisolo.beehiiv.com/subscribe](https://aisolo.beehiiv.com/subscribe) - [Upgrade to Premium](https://x.com/i/premium_sign_up) - [2:49 AM · May 5, 2026](https://x.com/itsolelehmann/status/2051373618727469152) - [53.3K Views](https://x.com/itsolelehmann/status/2051373618727469152/analytics) - [View quotes](https://x.com/itsolelehmann/status/2051373618727469152/quotes) --- *导出时间: 2026/5/26 12:24:23* --- ## 中文翻译 # 如何让 Claude 时空穿越并告诉你你的计划为何失败（事前验尸技能） **作者**: Ole Lehmann **日期**: 2026-05-04T18:49:18.000Z **来源**: [https://x.com/itsolelehmann/status/2051373618727469152](https://x.com/itsolelehmann/status/2051373618727469152) --- ![Image](/doc/2246/image/img_06e34051d3d4.jpg) POV（视角）：Claude 时空穿越到了 6 个月后的未来，并准确地告诉你你的下一步行动是如何失败的。这叫做“事前验尸”。这可能是 Claude 对高风险决策所做的最有用的一件事。这就是这项确切的技能，以及如何在 5 分钟内为你自己的计划运行一次。 P.S. 如果你想每周收到更多像这样的免费 AI 工作流，请加入这里订阅，和 37,000 名读者一起：aisolo.beehiiv.com/subscribe ## 这个技能解决的问题 Claude 是一个乐于助人的助手。这字面上就是它的产品。这就是它的训练方式：乐于助人、随和、默认乐观。当你把一个计划带给它时，它自然会为你打气，并倾向于它会成功的理由。所以你会带着自信离开。你在接下来的 6 周里构建它。你调动资金来资助它。那些通常与你所爱之人共度的周末消失在“再推一把”的努力中，因为这一次定会成功。然后，在 2 个月后，它在你的面前炸裂了。发布只达到了目标的一半，到了第 2 周，退款请求开始滚滚而来。或者，那个你曾如此确信的招聘结果证明是错误的人选，而你已经烧掉了 4 万美元的跑道资金，无法挽回。但你无法追回时间。你也无法追回金钱。事后诸葛亮，失败是显而易见的。你只是从未对其进行压力测试，因为 Claude 告诉你它很可靠。但如果你不必“吃一堑长一智”呢？ ## 这个方法的来源补救措施是一种名为“事前验尸”的 30 年前的技术。加里·克莱恩，一位研究心理学家，在 80 年代发明了它。他注意到团队不断做出每个人都私下认为糟糕的决定，但没有人会大声说出来。他的做法是翻转时间线。你告诉团队现在是 6 个月后，项目已经失败了。然后每个人都描述它是如何死的。这单一的变化给了人们许可，去吐露他们一直吞在肚子里的失败模式。丹尼尔·卡尼曼，《思考，快与慢》背后的诺贝尔奖心理学家，称这是他所知道的唯一最有价值的决策技术。沃顿和康奈尔大学的研究人员后来对其进行了正式研究，并将其命名为“预期性后见之明”。他们的发现：想象未来的失败，使人们比询问“可能会出什么错？”更能显著识别出导致失败的真正原因。这项技术现在运行于世界上风险最高的决策桌旁： - 高盛在主要投资委员会决策和 IPO 承销电话会议之前使用事前验尸。合作伙伴在签字前想象交易告吹。 - 谷歌在主要产品发布之前使用它们。团队演练那个公开失败的发布版本，并逆向推导导致其失败的原因。我采用了这项技术，并将其变成了一个 Claude 技能，可以在大约 5 分钟内为你自动运行整个过程。 ## 事前验尸技能是如何工作的该技能不仅仅是一个聪明的提示。以下是它底层运行的内容。 1. 它收集背景信息。该技能扫描对话和你的工作区（内存文件、项目笔记、你引用的任何内容），以了解计划是什么、为谁服务、成功是什么样子。如果你已经在和 Claude 讨论该计划，“premortem this”（对这个做验尸）就足够了。 2. 它生成原始的失败模式。 Claude 设定框架（“现在是 6 个月后，这个计划死了，这就是原因”）并列出每一个可能导致其死亡的真正原因。数量取决于你的具体计划。可能是 4 个，也可能是 9 个。 3. 它为每个失败模式并行生成一个调查员。这是让输出感觉真实的关键部分。每个失败原因都有自己的子代理，独立深入挖掘，所有代理同时运行。每个代理编写一个简短的案例研究，说明该特定失败是如何发生的、其背后的潜在假设，以及你会看到的早期预警信号。这是我最喜欢的部分，因为你真的可以想象它是如何失败的。 4. 它将所有内容汇总成一个综合报告： - 最可能的失败（首先要关注的那个） - 最危险的失败（那个值得投保的，即使可能性较小） - 你做出的最大隐藏假设 - 修改后的计划，其中包含针对特定失败模式的具体变更 - 一个包含 3-5 项内容的发布前检查清单，以便在发货前进行核实 5. 它为你提供一份可视化报告。如果你想要深入挖掘任何单个代理的推理，还会提供一份 markdown 文字记录。这是一个例子。你告诉 Claude 你正与另一位创作者合作推出联合产品，五五分账，3 个月后发布。该技能返回了 5 种失败模式： 1. **付出失衡**：其中一方承担了重任，而另一方被拉入了另一个项目 2. **受众不匹配**：你的受众各自对你个人的信任不会转移到联合产品上 3. **愿景漂移**：6 周后，你们对产品应该是什么产生分歧，每个决策都变得拖沓 4. **决策僵局**：50/50 意味着没有决胜票，艰难的决策无法做出 5. **没有退出条款**：其中一方在 9 个月后想退出，但没有任何书面协议每一个都有其深入挖掘的代理。这是关于“付出失衡”的那个返回的结果：前 6 周很棒。到了第 3 个月，你的合作伙伴正埋头于另一个项目，回复需要 48 小时。你承担了 80% 的工作。发布推迟了 6 周。当销售额进来时，50/50 的分成感觉很残酷。早期预警信号：一周内超过 3 条未回复的消息，或者一个错过了的截止日期，你事后才知道。然后在综合分析中，最大的隐藏假设：平等的合作伙伴关系和平等的付出是一回事。你建立了一个 50/50 的交易，期望工作会自动平衡。这是一条你自己永远不会写下来的话。该技能在它花费你 4 个月的工作之前，将其拖到了光天化日之下。 ## 如何自己运行一次 1. 在这里获取该技能。将其放入你的 Claude Code 或 Cowork 的技能文件夹中。 2. 打开 Claude，心中有一个计划。发布、招聘、定价变更、内容押注。任何一旦出错会让你损失数周或真金白银的事情。 3. 输入触发词：“premortem this” 该技能会从那里接手。如果它从对话中获得了足够的背景信息，它就会运行。如果没有，它会一次向你问一个重点问题，直到它有足够的信息来运行一个有用的事前验尸。不需要简报。 4. 让它运行。该技能收集背景信息，生成失败模式，并行生成代理，并将 HTML 报告和文字记录保存到你的工作区。这就是整个循环。 ## 你可以用它来做什么任何一旦出错有真实成本的事情。该技能不在乎这是商业决策、职业变动、创意押注还是个人承诺。每次都是同样的机制：命名计划，宣布它已死，找出原因。值得运行的真实例子： - 产品发布、定价变更，或任何其他涉及资金的事情 - 你即将承诺的招聘、合作伙伴关系或合资企业 - 职业变动，如接受新工作、筹集资金或成为全职自由职业者 - 重新定位、品牌重塑或战略转型 - 创意押注，如开始播客、写书或全力投入一个新的内容渠道 - 投资决策、大宗购买、你即将签署的长期合同 - 重大个人承诺：搬家、买房、接手一个会吃掉你周末的一年期项目如果你能用 3 句话描述计划，且出错的成本超过几个小时，请在承诺前对其进行事前验尸。 ## 今晚就运行一次选择你即将承诺的下一步行动。今晚睡觉前对其进行事前验尸。 5 分钟，你将在经历失败之前预见它。 P.S. 如果你想获得更多像这样的 AI 工作流，我每周都会免费将它们发送给 37,000 名读者。此外，当你加入时，你会免费获得一门 Claude Cowork 课程作为奖励：aisolo.beehiiv.com/subscribe ## 相关链接 - [Ole Lehmann](https://x.com/itsolelehmann) - [@itsolelehmann](https://x.com/itsolelehmann) - [53K](https://x.com/itsolelehmann/status/2051373618727469152/analytics) - [aisolo.beehiiv.com/subscribe](https://aisolo.beehiiv.com/subscribe) - [aisolo.beehiiv.com/subscribe](https://aisolo.beehiiv.com/subscribe) - [升级到 Premium](https://x.com/i/premium_sign_up) - [2026年5月5日上午 2:49](https://x.com/itsolelehmann/status/2051373618727469152) - [53.3K 观看](https://x.com/itsolelehmann/status/2051373618727469152/analytics) - [查看引用](https://x.com/itsolelehmann/status/2051373618727469152/quotes) --- *导出时间: 2026/5/26 12:24:23*