premortem
Run a premortem on any plan, launch, product, hire, strategy, or decision. Imagines it failed 6 months from now, works backward to find every reason why, then produces a revised plan. Triggers include "premortem this", "premortem my", "what could kill this", "stress test this plan", "find the blind spots", "poke holes in this", "where will this break", "am I missing anything", "what could go wrong", "future-proof this", "devil's advocate this".
b1rdmania · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add b1rdmania/claude-premortem-skill --skill premortem技能说明
name: premortem description: Run a premortem on any plan, launch, product, hire, strategy, or decision. Imagines it failed 6 months from now, works backward to find every reason why, then produces a revised plan. Triggers include "premortem this", "premortem my", "what could kill this", "stress test this plan", "find the blind spots", "poke holes in this", "where will this break", "am I missing anything", "what could go wrong", "future-proof this", "devil's advocate this".
Premortem
The opposite of a postmortem. Imagine the plan already failed and work backward to find why, before you start.
- Method: Gary Klein, Harvard Business Review, 2007.
- Kahneman praised it in Thinking, Fast and Slow as a way to counter overconfidence.
- Widely adopted in corporate strategy and project planning.
- 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/Cornell call this "prospective hindsight."
Why it 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.
When NOT to apply
- Vague ideas with no concrete plan yet (help plan first, then premortem)
- Questions with one right answer (just answer)
- Creative feedback on a draft (that's editing)
- Decisions already made and irreversible (premortem only helps when course correction is possible)
- Requests for multi-perspective decision support (use LLM Council instead — different mechanism, different output)
- Simple feedback or factual questions
Step 1 — Gather minimum context
A premortem is only as good as its input. You need three things:
- What is it? — describe the plan in one sentence
- Who is it for? — audience, customer, team, stakeholders
- What does success look like? — failure is the inverse
Scan first, ask second:
- Read the current conversation for context already provided
- Glob + Read the workspace for
CLAUDE.md, anymemory/folder, project briefs (~30 seconds max) - If all three are clear, proceed
- If not, ask for the most important missing piece. One question at a time. Conversational, not a form.
Step 2 — Set the premortem frame
Tell the user, naming the actual plan (not a placeholder):
"OK, premortem time. It's 6 months from now. The [actual plan: workshop / launch / hire / pricing change / etc.] has failed. It's done. Let's look back and figure out why."
The "this has already failed" framing is the active mechanism. Without it, the analysis collapses back into polite risk assessment.
Step 3 — Generate failure reasons
Run a single comprehensive pass. No prescribed categories, no lenses.
"This plan has failed 6 months from now. Generate every genuine reason it could have died. Be specific. Ground each reason in the actual details of the plan. Don't pad with weak reasons. Don't stop early if there are more."
Each reason should be specific to this plan, grounded in details the user provided, and a real threat (not minor inconvenience or extreme edge case). Use whatever count is real for this plan — could be 4, could be 9. Don't force a number.
Step 4 — Spawn deep-dive agents in parallel
Write the plan context once to a scratch file (premortem-context.md: what it is, who it's for, what success looks like, relevant workspace notes). Agents read that file; do not paste the context into each prompt — N agents each carrying the full context is the largest cost in the run.
For each failure reason, spawn one sub-agent with model: "sonnet" set explicitly. Never inherit the parent model: the deep-dives are narrative writing from a fixed brief, synthesis stays with you. All in parallel — sequential spawning lets earlier outputs influence later ones.
Spawn one extra sonnet agent alongside them, the assumption excavator: read the context file, return at least 5 unstated assumptions the plan takes for granted, each rated load-bearing (if wrong, does the plan collapse?) and testable (can it be checked inside 6 months?). It finds things the failure stories miss.
Pass each deep-dive agent the prompt body below as its task. Substitute the angle-bracket values with actual content before sending — do not pass the brackets through.
You are an investigator in a premortem analysis. You've been assigned one specific failure reason to analyse in depth.
THE PLAN: read <path to premortem-context.md>
PREMORTEM FRAME: It is 6 months from now. This plan has failed.
YOUR ASSIGNED FAILURE REASON: <the specific failure reason from step 3>
Your job: go deep on this one failure. Write the story of how it played out. Use details from the plan. Make it feel like a case study of something that actually happened.
Output three sections:
- The failure story — 2-3 paragraph narrative. Specific moments where things went wrong and why.
- The underlying assumption — the one thing the user took for granted that made this failure possible. One sentence.
- Early warning signs — 1-2 concrete, observable signals the user could watch for. Things you can see or measure, not vague feelings.
- Rating — probability (high/medium/low) and impact (high/medium/low), one line.
Keep total under 300 words. No preamble, no recap of the plan. Direct. No hedging. No sugarcoating.
Step 5 — Synthesise
Read every deep-dive and the excavator's list. Produce:
- Most likely failure — highest probability rating; break ties on the story's specificity. The one to focus on first.
- Most dangerous failure — highest impact rating, even if less likely. The one worth insuring against.
- Hidden assumption — the single most load-bearing assumption across the excavator's list and the deep-dives' underlying assumptions. Often where the real value of the premortem lives.
- **Revised