Home / Techniques that change the result

The premortem: assume the plan has already failed

The premortem reverses the question. Instead of asking whether the plan is good, which invites looking for reasons to say yes, you state that it has already failed and ask for the autopsy. It brings out the causes of failure, which remain hypotheses to check.

Asking Claude "is this plan good?" invites it to look for reasons to say yes: it rereads the plan, notes what holds and adds two minor reservations. The premortem reverses the question. You state that the plan has already failed, three months from now, and you ask for the autopsy. The psychologist Gary Klein described this method for human teams; it applies to a plan Claude has just handed back.

Five requests, in this order

The autopsy ranks the causes of failure by likelihood, names the assumption in the plan that proved false and the first sign that would have shown it. The verdict separates the most probable cause from the most dangerous one, since the second may be rare and cost the most. The reconstruction turns each retained cause into a check, with a threshold beyond which the plan is dropped. The adversary asks who wins if the plan fails. The triggers set, for each cause, a measurable signal that requires stopping.

The following demonstration writes the five requests into a new file, then prints their titles:

mkdir -p premortem-demo && cd premortem-demo
cat > premortem.md << 'EOF'
1. Autopsy: the plan has failed. List the causes ranked by likelihood, name the faulty assumption and the first sign that would have shown it.
2. Verdict: separate the most probable cause from the most dangerous one, and say which to handle first.
3. Reconstruction: for each retained cause, give one check to run before acting and the threshold beyond which the plan is dropped.
4. Adversary: who wins if this plan fails, and what would that person do to make it happen?
5. Triggers: for each cause, give one measurable signal that requires stopping as soon as it is observed.
EOF
cut -d: -f1 premortem.md

1. Autopsy
2. Verdict
3. Reconstruction
4. Adversary
5. Triggers

Paste the contents of the file after the plan. Each request fits on one line, because a vague autopsy produces vague causes: "the plan lacks rigor" cannot be checked, "the customer export is in a format the target tool rejects" can.

When to run it, and who gets the questions

Run it when the plan comes back, before the first action. The method costs one more exchange: it pays off when the work touches several repositories or several systems, or when its effect is hard to undo. On a short, reversible plan it costs more than it returns. To avoid relying on memory, a hook can add the five requests when the plan returns.

Ask Claude to check for itself what can be read in the files, and to put to you one question at a time on whatever depends on a decision, an outside agreement or data only you hold, instead of assuming it settled.

What the autopsy establishes

An autopsy written by a model is a list of plausible stories, not a finding. Each cause stays a hypothesis, not yet proven in the sense of the lesson on proven versus not yet refuted, until the check from the reconstruction has run and its output has been read. The premortem also works on the plan, before any result exists: it does not replace the verifier subagent, which later judges a finished artifact.