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.
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/<plan-name>.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:
Description in one paragraph + why it matters + angle + confidence.
2–3 strategic options with +/–. Directions, not tasks.
User picks / asks for other options / defers / rejects.
For the chosen option: what we picked, why, first step, owner
(agent / human / both).
After all holes:
The skill picks the top 1–3 by importance × reversibility ×
confidence.
The top is expanded to full mode: prevent / detect / stop condition /
limit damage.
Bias check on the top. The skill proposes the most likely
cognitive bias that may have distorted the decisions, and a concrete
correction.
Session recommendation:continue / reduce_stake / delay /
abort — based on the pattern of accepted decisions.
Reverse premortem runs only on reduce_stake / delay /
abort: "the horizon passed. We didn't do it. Turned out to be a
mistake — why?" Counter to the over-cautious pull of pure premortem.
Phase 4. Persist (1 min)
One main file at docs/premortem/<slug>.md plus a timestamped snapshot
in docs/premortem/history/. Writes are atomic (tempfile + os.replace
fsync) under an advisory file lock. Any sequence of runs leaves the
file valid.
Top 1–3 holes in full mode: prevent / detect / stop / limit
damage, owner, first step
Bias check in one line
Session recommendation + reason
Reverse premortem (when applicable): 3 reasons "why we may have
been wrong not to do it" + what tips the balance
Run history with status dynamics: same / worse / resolved /
new
When to use
The skill works on concrete, reversible commitments with a high cost of
being wrong. Good cases:
Product / feature launch. "Shipping the new dashboard in 6
weeks."
Pricing change. "Bumping the subscription from $19 to $39."
Hire. "First marketing hire, junior, $60k base."
Content launch. "$297 live workshop on [topic] for [audience]."
Architecture decision. "Migrating monolith to microservices in a
quarter."
Partnership / contract. "Signing a 12-month exclusive with one
distribution channel."
Minimum requirements: there's a plan with shape (not an idea in your
head), you have 10–15 minutes, the commitment isn't finalized yet.
When NOT to use
Idea without shape. Nothing to break — help draft the plan first,
then premortem.
Question with one right answer. "Which model is faster?" — that's
fact-check, not premortem.
Creative editing of a draft. That's editing, not premortem.
Already-irreversible decision. Premortem can't roll it back.
Multi-perspective input on a present-tense decision. LLM Council
pattern fits better than past-tense framing.
Recommendations
Run when the plan is "almost ready." Too early — nothing to
break. Too late — you can't influence anything.
Don't skip reference class. "What does this resemble from done
projects?" is the cheapest source of outside view. If you don't know
— say so, the skill moves on with empty reference class.
Don't default to "defer." Premortem systematically biases toward
caution (Klein 2025). That's why reverse-premortem is built in: it
asks "what if we were wrong NOT to do it?"
Re-run after a month. History compares — which holes closed,
which got worse, which are new. Catches plan drift.
Don't run premortem on someone else's plan without context. Get
at least subject / audience / success first. Without that you'll get
a generic risk checklist, not premortem.
Accept 2–3 decisions, not 8. If the skill produces 8 holes and
you accept all of them, it's likely an illusion of control.
Premortem worked when the top is chosen and there's a concrete first
step.
Installation
Requirements:
Claude Code
Git
uv — runs the CLI scripts
(deterministic mechanics)
Install:
hljs language-bash
git clone https://github.com/AndyShaman/premortem.git
cd premortem
./install.sh
The installer copies premortem/ (SKILL.md, scripts, references,
assets, evals) to ~/.claude/skills/premortem/. Tests stay in the
repo.
Verify:
hljs language-bash
ls ~/.claude/skills/premortem/SKILL.md
Restart Claude Code (or start a new session) — the skill loads
automatically on trigger phrases.
Update:
hljs language-bash
cd premortem
git pull
./install.sh
Triggers
The skill auto-loads on phrases like:
"premortem", "premortem this", "run premortem"
"find holes in this plan"
"look at this plan from the outside"
"what could kill this [plan]?"
"what am I missing in [plan]"
"future-proof this [plan/launch]"
Example prompts:
"premortem this: launching a $297 live workshop on Claude Cowork for
marketing teams"
"stress-test this hire — first marketing person, junior, $60k base"
"find holes — shipping the new pricing in 2 weeks"
Method origins
Gary Klein, Performing a Project Premortem, HBR, 2007
Klein, Sources of Power, MIT Press, 1998 — mental simulation, RPD
Klein, The Pre-Mortem Method, Psychology Today, 2021 — action round
