premortem

A Claude Code skill that runs a structured premortem on a concrete plan, launch, hire, or decision — finds in 10–15 minutes the holes that would otherwise show up only after the failure.

Russian / Русский

---

Why this exists

You ship a product — feels well thought out. Three months later it's dead. And the whole way through, Claude was cheerfully nodding: "yeah, solid plan, good luck!"

That's not a bug in any specific model. It's the default behaviour of all LLMs. Ask "what could go wrong?" and you get hedged, polite, generic risk lists. Ask "is this a good plan?" and the model finds reasons to say yes.

Premortem breaks the pattern. Not "evaluate this plan," but "imagine the plan has already failed — explain how it died." The model switches into narrative mode and produces concrete, creative, honest causes instead of hedging.

Mitchell, Russo & Pennington (1989) measured ~30% more specific causes of future outcomes under past-tense framing than under conditional framing. Kahneman called premortem his single most valuable decision-making tool.

This skill ports the method into Claude Code as a reproducible session with deterministic mechanics and persistent history.

---

What it does

The session has four phases:

• Pre-flight — syncs context with what the project already knows (CLAUDE.md, attached briefs), only asks for what's missing.
• Silent scan — three parallel helpers look at the plan from six angles (Customer, Operator, Adversary, Assumptions, situational angle, and always "Future Maintainer" for the WYSIATI micro-step). Semantic dedup leaves 5–8 unique holes.
• Live dialog — for each hole the skill proposes 2–3 strategic options, the user picks one. The top 1–3 are expanded to full mode (prevent / detect / stop condition / limit damage).
• Persist — atomic write to docs/premortem/.md plus a snapshot in history/. A month later you can re-run premortem on the same plan and compare: which holes closed, which got worse, which are new.

Built into the synthesis: outside view (Kahneman & Lovallo 2003), bias scan with 6 items from the Kahneman/Lovallo/Sibony HBR 2011 checklist, and a conditional reverse-premortem (Klein 2025) when the recommendation is delay/abort — countering the over-cautious bias of pure premortem.

---

How it works

Phase 1. Pre-flight (1–2 min)

The skill first scans available context — what's already in the chat or in project files. It only asks for what's missing. Minimum to start:

• subject (what's launching / being decided)
• audience
• success criterion
• evaluation horizon (30 days / 3 months / 12 months / after pilot / after launch)

Reference class is asked separately: "What does this resemble from projects already done?" — the foundation for the outside view.

Phase 2. Silent scan (3–5 min)

The skill classifies the plan type and picks 6 angles for it. Three parallel LLM helpers run in one message (2 angles each). Each returns up to 2 holes with a confidence marker: "backed by context" / "needs verification" / "assumption."

Hash dedup removes obvious duplicates; semantic dedup merges close holes while preserving sources. Result: 5–8 unique holes, sorted by importance × confidence.

The user only sees the chosen angles in one line, then the prepared list of holes.

Phase 3. Live dialog (5–8 min)

For each hole: