How to make Claude time-travel and tell you how your plan dies

A pre-mortem assumes the plan has already failed, then works backward to explain how it died. Unlike a standard risk review that asks 'what could go wrong?', the past-tense framing triggers prospective hindsight: research from Gary Klein (1989) shows this makes people roughly 30% better at identifying the actual reasons for failure. The cognitive shift matters because it bypasses the emotional investment teams have in plans they've already started building.

You upload a single SKILL.md file to Claude, after which Claude is available as a structured adversarial auditor on demand. You feed it your plan, and it reconstructs the context, generates ranked failure modes by likelihood and severity, identifies the hidden assumption everyone treated as settled, and returns a revised plan with owners, early-warning metrics, and kill criteria. The article says the full cycle runs in about 60 seconds from a well-formed input.

The full Linas Beliūnas piece on Substack is marked as paid content, requiring a subscription to access the complete SKILL.md file and all five worked playbooks. The aisolo.beehiiv.com and generativeai.pub versions are separately available and cover the core technique, the cognitive science behind it, and practical prompt framing without a paywall.

The article identifies five primary playbooks: product launches, fundraising rounds, hiring decisions, strategy pivots, and investments. The common thread is irreversibility or high cost, decisions where optimism bias is most dangerous because you cannot easily unwind them. The technique is less necessary for low-stakes, easily reversible choices where moving fast outweighs the cost of a wrong turn.

Claude's RLHF training causes it to prefer agreeable responses, especially when a user signals ownership by presenting their own plan. Asking 'is this good?' reliably produces validation with cosmetic caveats. The pre-mortem prompt sidesteps this entirely by not asking for a verdict: instead it gives Claude a fact ('the plan failed') and asks for a causal narrative, which is a task Claude can perform without the sycophantic pull. Anthropic, OpenAI, and Meta have all documented this sycophancy behavior in their own research.

The article's 'weak input vs. strong input' rule is the key variable: a vague plan description produces a vague failure analysis that won't change any decision. The skill works best when the input is specific enough that Claude can reconstruct the plan's real assumptions, timelines, and constraints. The article includes a real worked example of a launch being torn apart and rebuilt, which illustrates concretely what a high-quality input and output look like.