The AI pre-mortem: run the failure analysis before you commit

The pre-mortem shifts the review from after the work to before it — when course changes are still cheap.

Why optimism is the default — and the problem

Every plan looks better at the moment you write it than it will six months later. The people who built it are invested. The assumptions feel solid. The risks feel manageable. And so the hard questions don't get asked until after the money is spent, the announcement is made, or the team is committed.

The pre-mortem is a decision-science technique for breaking that pattern. Invented by psychologist Gary Klein, it works by temporarily assuming failure: instead of asking "could this go wrong?" — which invites optimistic hedging — you ask "this has already failed; what went wrong?" The frame shift is small but the output is different. People volunteer the risks they were quietly worried about rather than defending the plan.

This pattern is drawn from Linas Beliunas's June 2026 write-up, which packages the pre-mortem as a reusable Claude skill and documents the specific prompt structure that produces the most actionable output. The approach pairs naturally with Claude's extended reasoning — the model is unusually good at holding a complex brief, generating diverse failure modes, and surfacing the assumption everyone treated as settled.

What the workflow produces

A pre-mortem session with Claude gives you four things in one output:

Pre-mortem output · four sections

Failure modesTen specific ways the plan could fail, each traced to a false assumption.

Risk scoresEach mode rated for likelihood and severity, so you see where effort is justified.

Hidden assumptionThe one thing everyone treated as settled but never validated.

MitigationsOne concrete action per top-three risk, with an owner and an early-warning signal.

The hidden assumption section usually produces the most useful output. It forces Claude to read across the whole brief and name the thing that, if wrong, invalidates the plan entirely. On consequential decisions — a new hire, a vendor contract, a product pivot — that alone is worth the ten minutes.

The prompt

Paste this into Claude with your plan or decision in place of the bracket. Use Claude Sonnet 5 or Opus 5 for the best output on complex briefs.

You are a senior strategic risk analyst reviewing a plan that has already failed. The plan: [Paste your plan, proposal, project brief, or key decision here. Include the goal, the main steps, the key assumptions, and any known constraints. Two to ten paragraphs is ideal.] It is now [insert a date three to six months from today]. The project or decision has failed. You are conducting a post-mortem. Step 1 — Failure modes List the ten most likely failure modes. For each, describe: what specifically went wrong, which assumption proved false, and who was most affected. Step 2 — Risk scores Score each failure mode on two axes: likelihood (1–5) and severity (1–5). Show one sentence of reasoning per score. Multiply the two scores for a combined priority number. Step 3 — Hidden assumption Identify the single assumption the team treated as settled but never validated. State it in one plain sentence starting with "The assumption was..." Step 4 — Mitigations For the three highest-scoring failure modes, write one concrete mitigation action each. Each mitigation must include: the action in one sentence, the named role responsible, and an early-warning signal that would give at least two weeks' notice before the failure becomes unrecoverable. Output format: a structured brief with four numbered sections, one per step above. Keep each mitigation to three lines maximum.

Why each part earns its place

The "already failed" frame. Optimism bias is automatic. The moment you ask "could this fail?" the team is already defending the plan. The past-tense failure frame sidesteps that reflex and gets you the honest list.

Asking for ten failure modes. Most people name three risks. Three is where the obvious problems live. Seven through ten is where the problems you were quietly worried about start appearing. Forcing the model past the comfortable number is the point.

Two-axis scoring. The combined priority number separates frequent-small failures from rare-catastrophic ones. A vendor who is slightly unreliable (likelihood 4, severity 2 = 8) ranks below a regulatory assumption that turns out to be wrong (likelihood 2, severity 5 = 10). The arithmetic makes that distinction visible.

The hidden assumption step. Plans fail because of things that were never questioned. This step asks Claude to read the whole brief and name the one load-bearing assumption — the thing that, if false, invalidates everything else. It is the part of the output most likely to make a room go quiet.

Early-warning signals. A mitigation without a trigger is a good intention. The early-warning signal gives the named owner something specific to watch for, with enough lead time to change course. Two weeks is a practical minimum for most operational decisions.

How to use the output

Do not treat the failure modes as a definitive list. Claude does not know your internal politics, your specific team's track record, or the vendor relationship you have been managing carefully for three years. It works from the brief you give it — so a vague brief produces vague failure modes.

The most effective use is to run the prompt, then bring the output to the planning meeting as a discussion document. Ask the room: which of these have we already mitigated? Which ones surprised you? Are there failure modes not on the list that we know about but haven't said out loud?

That conversation — seeded by the output, not replaced by it — is where the real value lands.

What this will not do

The pre-mortem does not replace domain expertise. If the plan involves a regulatory filing, a technical architecture, or a specific supplier relationship, Claude's failure modes will be as good as the context you provide and no better. Front-load the brief with the relevant detail.

It will not identify risks that are genuinely outside its training — novel market conditions, a specific competitor's undisclosed move, a team dynamic you have not described. The model is generating plausible failure modes, not predicting the future. Review the list with that in mind.

And it will not run itself. The prompt takes ten minutes of genuine attention: writing a clear brief, reading the output critically, and deciding which mitigations are worth pursuing. Paste-and-ignore produces the same confident plan you started with.