Premortem

Quick facts

Mechanism: Generates a ranked risk register that stress-tests a planned decision by imagining it has already failed, surfacing the likely causes and pairing each with a mitigation, tripwire, and kill criterion.

Produces: risk-register

Evidence: S/M · Domain: Risk and Resilience

Use when about to commit to a launch, hire, investment, migration, or any risky, hard-to-reverse decision, or when you need risks surfaced before committing.

A premortem stress-tests a plan by assuming it has already failed and reasoning backward to explain why, then converting each cause into a mitigation, a tripwire, and a kill criterion. The shift from “what might go wrong?” to “it went wrong, why?” is what does the work: it licenses dissent, surfaces more and more specific causes than ordinary risk review, and turns vague worry into pre-committed action while you can still change course. The output is a risk register, not a discussion.

When to Use

• Before a launch, hire, investment, migration, vendor selection, or any consequential, hard-to-reverse commitment.
• When a plan has optimistic momentum and you suspect concerns are going unspoken.
• When you want risks expressed as observable signals and pre-decided responses, not a feeling of caution.
• Often after options have been compared and one has been chosen, as the last gate before committing.

When NOT to Use

• After the outcome is known. That is a postmortem, a different tool.
• For trivial or fully reversible (two-way-door) decisions. The ceremony is not worth it.
• To generate options or to choose among them. Use an ideation skill or a decision-option review; premortem is a risk tool.
• As a ritual to bless a decision already made. If the mitigations will not be acted on, skip it; a premortem nobody acts on is theater.

Instructions

When asked to run a premortem, follow these steps:

1. Frame the decision and the horizon. State the plan and intended outcome in one or two sentences, and pick a concrete time horizon (for example, “six months after launch”). If the decision is trivial or already irreversible, say so and stop.
2. Declare the failure vividly. Assert it in the definite past: “It is [horizon]. This plan has failed badly.” Make the failure concrete and specific, not “it underperformed.”
3. Generate causes broadly, before judging. List the plausible reasons the failure happened. Aim for breadth and specificity; include uncomfortable, political, and second-order causes, not just technical ones. Do not filter yet.
4. Cluster and rank. Group related causes and rank them by likelihood and impact (High/Medium/Low each). Keep the vital few; do not pad.
5. Convert each top cause into action. For each high-priority cause, define a leading signal / tripwire (the early sign it is happening), a mitigation (what reduces the risk now), an owner, and a kill criterion (the pre-decided condition under which you stop or change course). This conversion step is mandatory; a list of risks without it is not a premortem.
6. Emit the risk register and a short summary. Produce the artifact in references/TEMPLATE.md: a one-paragraph “top risks and what we will do” summary above a ranked register table.

Output Format

Use the template in references/TEMPLATE.md. The deliverable is the filled risk register plus its summary, not a prose essay.

Quality Checklist

Before finalizing, verify:

• The failure was declared in the definite past, with a concrete horizon.
• Causes go beyond the obvious technical ones (include people, process, second-order, and external causes).
• Every high-priority cause has a tripwire, a mitigation, an owner, and a kill criterion (the conversion step is done).
• Risks are ranked, not an undifferentiated list.
• The output is the risk register artifact, not prose.
• No overclaiming: the skill does not promise a better outcome, only better-surfaced and better-handled risk (see evidence/dossier.md).

Evidence

Tier S/M (contested). Prospective hindsight reliably increases the number and specificity of causes surfaced and reduces overconfidence in a plan (Mitchell, Russo & Pennington 1989; Veinott, Klein & Wiggins 2010). It does not have strong evidence of improving final outcomes, and the often-quoted “30%” figure measures the number of reasons generated, not decision quality. The evidence is transferred from human studies and has not been validated for AI-augmented use. Full grading, sources, and caveats: evidence/dossier.md.

Examples

See references/EXAMPLE.md for a completed premortem on a real decision.

Deep dive: worked example

A full worked run (the shared Northwind scenario)

Premortem Risk Register - Worked Example

A completed run of the premortem skill on a real, consequential decision. This is the quality bar a generated premortem should meet.

---

Decision under premortem

• Decision: Launch a self-serve free tier of our B2B SaaS in 6 weeks to accelerate top-of-funnel growth ahead of the Q3 board target.
• Intended outcome: 3x sign-up volume within one quarter and a measurable lift in paid conversions from self-serve, without degrading the existing sales-led motion.
• Horizon: 6 months after launch.
• Reversibility: One-way door in practice. Pulling a free tier after launch is possible but damages trust and is publicly visible, so treat it as hard to reverse.

Top risks and what we will do (summary)

The three most likely ways this fails: (1) the free tier cannibalizes paid rather than feeding it, because the entry plan is too generous; we will gate the highest-value features behind paid and instrument the free-to-paid path from day one. (2) Support and infrastructure load from unqualified free users swamps the team and blows the cost model; we will cap free usage, ship self-serve docs, and set a cost-per-free-user tripwire before launch. (3) The sales team undercuts or resents the motion because comp and qualification rules were not redesigned; we will align comp and lead-routing with sales leadership before any external announcement. Each has a tripwire and a kill criterion below.

Risk register

#	Cause of failure	Likelihood	Impact	Leading signal / tripwire	Mitigation	Owner	Kill criterion
1	Free tier cannibalizes paid: existing or prospective paying customers downgrade to free	H	H	Net new paid MRR growth slows in the first 4 weeks while free sign-ups rise; >5% of trials choosing free over paid	Gate the top 3 value features behind paid; instrument free-to-paid funnel before launch; A/B the free limits	PM (Growth)	Paid net-new MRR drops below the pre-launch trend for 2 consecutive weeks attributable to free downgrades
2	Support + infra cost from unqualified free users exceeds plan	H	M	Support tickets per 100 free users above threshold by week 2; cloud cost per free user above the modeled ceiling	Hard usage caps on the free tier; self-serve onboarding + docs; a cost-per-free-user budget set before launch	Eng lead + Support lead	Cost per free user exceeds 1.5x model for 3 weeks with no path to fix
3	Sales team undercuts or resents the motion (comp + qualification not redesigned)	M	H	Reps steering prospects away from free; complaints in pipeline reviews; lead-routing disputes in week 1	Redesign comp + lead-routing with sales leadership before announce; written rules of engagement; a shared dashboard	VP Sales + RevOps	Sales leadership withholds sign-off, or rep behavior measurably suppresses free sign-ups in month 1
4	The 6-week timeline forces shipping a broken or insecure self-serve flow	M	H	Billing/auth edge cases open in QA week 5; security review not complete by week 4	Cut scope to a thin, secure path; freeze the feature set at week 2; mandatory security review gate	Eng lead	Security review not green by the launch-minus-1-week gate
5	”Free” attracts the wrong segment (no ICP fit), so conversions never come	M	M	Free cohort firmographics diverge from ICP; week-4 activation among ICP-fit free users is low	Light qualification at sign-up; track activation by ICP fit, not raw sign-ups	PM (Growth)	After 8 weeks, ICP-fit free-to-paid conversion is below the breakeven the model requires

Watch list (lower-priority causes)

• Brand perception shift (“they went freemium, they must be struggling”) - monitor inbound sentiment, low likelihood.
• Free tier abused for fraud/spam - rate-limit and monitor, standard controls likely sufficient.
• Internal analytics not ready to attribute free-to-paid - ensure tracking is in the launch scope, not after.

---

Note how the value is in the conversion: every top cause carries a tripwire, a mitigation, an owner, and a kill criterion decided in advance. A list of five risks without those columns would not be a premortem.

Grounding: the full evidence dossier

What the research does and does not show, with graded sources

Evidence Dossier: Premortem

The single source of truth for the premortem skill. The SKILL.md, the sidecar (skill.meta.yml), and the eval cases all derive from this file. If a claim is not here, it does not belong in the skill.

Skill	thinking-framework-skills.premortem (installable name think-premortem)
Family	risk-and-resilience
Evidence tier	S/M (contested - see “What the evidence shows” below)
Confidence	Moderate-high that the mechanism helps; low that the published effect sizes mean what they are usually quoted to mean
Status	draft (first authored 2026-05-31, against discovery corpus)

---

1. The mechanism (what actually does the work)

A premortem is a deliberate act of prospective hindsight: instead of asking “what could go wrong?”, you assert that the plan has already failed and reason backward to explain why. The shift from a conditional (“might fail”) to a definite past (“has failed”) is the load-bearing move. It does three things:

1. Licenses dissent. Once failure is assumed, naming a reason is no longer disloyalty or pessimism; it is the assigned task. This is why a premortem surfaces concerns that normal risk review and optimistic planning suppress.
2. Recruits memory and imagination differently. Explaining a concrete past event is a richer retrieval cue than forecasting an abstract future one, so people generate more, and more specific, causes.
3. Converts vague worry into pre-committed action. The output is not a feeling of caution but named causes, each paired with a mitigation, a leading signal (tripwire), and a kill criterion decided before sunk cost and momentum distort judgment.

The mechanism is what we implement. The branded “premortem” ritual is the packaging; the durable move is prospective hindsight plus structured conversion to mitigations.

2. Lineage

• Prospective hindsight as a cognitive effect: Mitchell, D. J., Russo, J. E., & Pennington, N. (1989). “Back to the future: Temporal perspective in the explanation of events.” Journal of Behavioral Decision Making, 2(1), 25-38.
• The “premortem” technique as a management practice: Klein, G. (2007). “Performing a Project Premortem.” Harvard Business Review, 85(9). Popularized further in Kahneman, Thinking, Fast and Slow (2011).
• Direct evaluation: Veinott, B., Klein, G., & Wiggins, S. (2010). “Evaluating the Effectiveness of the PreMortem Technique on Plan Confidence.” Proceedings of ISCRAM 2010.

No trademark. “Premortem” is a generic descriptive term in common use; no attribution is required and none is claimed. We name the skill descriptively and cite the lineage here rather than branding it.

3. What the evidence shows, and what it does NOT show

This is the honest core of the dossier. The skill must not overclaim.

What is reasonably supported (the S part):

• Prospective hindsight (assuming an outcome and explaining it) increases the number and specificity of causes people generate relative to ordinary forecasting. Mitchell et al. (1989) reported roughly a 30% increase in the number of reasons correctly identified for a future outcome under the “has happened” framing.
• The technique reduces overconfidence in a plan. Veinott et al. (2010) found participants who ran a premortem were better calibrated about their plans than those who did not.

What is NOT shown (the caveat that keeps the skill honest):

• The widely-quoted “premortems make decisions ~30% better” claim is a misreading. The 30% figure measures the number of reasons identified, not any improvement in decision quality, outcome, or accuracy. Generating more reasons is not the same as deciding better.
• There is no strong evidence that premortems improve final outcomes (project success rates, ROI, fewer failures). The mechanism is plausible and the calibration effect is real, but the chain from “more reasons surfaced” to “better real-world result” is not established by controlled study.
• General “thinking tools improve thinking” claims are weak: a 2024 meta-analysis in the problem-solving-pedagogy literature found no significant difference in some downstream measures between instruction that uses thinking tools and instruction that does not. (To be primary-source verified before any public claim; cited here as a humility prompt, not a settled fact.)

Net grade: S/M. The reason-generation and overconfidence-reduction effects are well-supported (S-leaning); the decision-quality improvement that the technique is usually sold on is not (M/contested). The skill should claim the former and explicitly disclaim the latter.

4. Transferred-evidence flag (required honesty for this library)

All of the evidence above comes from human subjects in workshop, lab, and team settings. There is no direct study of premortems run by, or with, an AI agent, and none of whether an AGENT-produced premortem improves a human’s decision. The evidence supporting this skill is therefore transferred from human contexts, not validated for AI-augmented use. This skill must say so. Treat the AI value as: the agent makes the mechanism cheap to run, enforces the structure, and produces a durable artifact - benefits that do not depend on the contested decision-quality claim.

5. When it works / when it fails (drives the eval negative cases)

Works best when:

• The decision is real, consequential, and not yet committed (you can still change course).
• There is genuine uncertainty and the plan has optimistic momentum behind it.
• Causes can be turned into observable signals and pre-decided responses.

Fails or misleads when (poor-fit / anti-patterns):

• Run after the fact - that is a postmortem, a different tool. (Anti-trigger.)
• Run as ritual - rote “imagine it failed, list five risks, done” with no conversion to tripwires/kill criteria produces cargo-cult comfort, not better risk handling. The skill must force the conversion step.
• Trivial or fully reversible decisions - the ceremony is not worth it; a two-way door does not need a premortem.
• Used to launder a decision already made - if mitigations are never acted on, the premortem becomes theater.
• Substituted for ideation or for option comparison - it is a risk tool, not a way to generate options (use SCAMPER/Question Burst) or to choose among them (use Decision Option Review).

6. Output artifact

The skill must emit a risk register, not prose: a ranked table of causes with likelihood, impact, a leading signal/tripwire, a mitigation, an owner, and a kill criterion, preceded by a short “top risks and what we will do” summary. The artifact is the deliverable; the conversation is not.

7. Sources

1. Mitchell, Russo & Pennington (1989), J. Behavioral Decision Making 2(1):25-38 - prospective hindsight; the ~30%-more-reasons finding.
2. Klein (2007), Harvard Business Review 85(9) - “Performing a Project Premortem.”
3. Veinott, Klein & Wiggins (2010), ISCRAM 2010 - premortem reduces overconfidence / improves plan calibration.
4. Kahneman (2011), Thinking, Fast and Slow - popularization; ties premortem to overcoming optimism bias and groupthink.

Verification status: citations 1-4 are standard and well-attested in the discovery corpus, but the exact effect-size phrasings and the 2024 meta-analysis claim in section 3 were drawn from a secondary research synthesis and should be confirmed against the primary papers before they appear in any public-facing README. They are safe to use inside this dossier because the dossier’s job is to be honest about exactly this uncertainty.