premortem

A Claude Code skill that runs a premortem on any plan, launch, product, hire, strategy, or decision. Imagines the plan failed 6 months from now, works backward to find every reason why, then produces a revised plan with the blind spots exposed.

The method

A premortem is the opposite of a postmortem. Instead of figuring out what went wrong after a failure, you imagine the failure has already happened and explain how, before you start.

• Method: Gary Klein, Harvard Business Review, 2007.
• Kahneman endorsed it in Thinking, Fast and Slow as a counter to overconfidence.

The mechanism: "what could go wrong?" produces hedged, polite answers. "This already failed, explain why" puts the brain into narrative mode and generates specific, honest causes. Wharton and Cornell call this prospective hindsight.

This matters for AI assistance. Claude defaults to agreeable. Asking "is this a good plan?" gets reasons it's good. The premortem reframe forces honest failure analysis instead of polite risk assessment.

How it works

Six steps. Linear flow. Parallel sub-agents do the heavy lifting in step 4.

flowchart TD
    A[Plan / launch / decision in] --> B[Step 1: gather minimum context]
    B --> B1{Three things known?What is it?Who is it for?What does success look like?}
    B1 -->|No| B2[Ask one question]
    B2 --> B1
    B1 -->|Yes| C[Step 2: set the premortem frameIt is 6 months from now. The plan failed.]
    C --> D[Step 3: generate failure reasonscomprehensive raw list]
    D --> E[Step 4: write context file, spawn parallel sonnet sub-agentsone per failure reason + one assumption excavator]
    E --> E1[Agent 1: failure A]
    E --> E2[Agent 2: failure B]
    E --> E3[Agent N: ...]
    E --> E4[Excavator: unstated assumptions]
    E1 --> F[Step 5: synthesise + gate]
    E2 --> F
    E3 --> F
    E4 --> F
    F --> F1[Most likely failure]
    F --> F2[Most dangerous failure]
    F --> F3[Hidden assumption]
    F --> F4[Revised plan]
    F --> F5[Pre-launch checklist]
    F1 --> G[Step 6: HTML report + 3-sentence chat summary]
    F2 --> G
    F3 --> G
    F4 --> G
    F5 --> G

Each sub-agent in step 4 takes one failure reason and writes the story of how it played out. Four things come back per agent: the failure narrative, the underlying assumption that made it possible, 1–2 early warning signs the user can watch for, and a probability/impact rating. A separate excavator agent returns the unstated assumptions the plan rests on, rated load-bearing and testable.

Cost discipline, learned from running this shape inside a larger research pipeline: the plan context is written to a file once rather than pasted into every prompt, every sub-agent runs on sonnet with the model set explicitly, and each returns under 300 words with no preamble. Synthesis stays with the parent model. Before the report is written, a gate checks that every failure maps to a revision and every checklist item names the failure it guards against.

What it catches

Failure classes that polite risk assessment routinely misses: