Claim
A pre-mortem ("assume this failed; why?") becomes a feelings exercise unless you commit explicit kill criteria — quantified conditions under which you will stop or pivot — before launch. The criteria turn the pre-mortem into a real decision gate.
Mechanism
Without kill criteria, every "fail mode" surfaced in a pre-mortem can be rationalised away in flight ("the metric is bad but the launch is great in other ways"). Kill criteria, written down before commitment, remove the rationalisation surface. The pre-launch decision binds the post-launch self. The same dynamic that makes Ulysses tie himself to the mast applies to teams under sunk-cost pressure.
Conditions
Holds when:
- The team can articulate measurable signals tied to the failure modes.
- Leadership backs the kill criteria when triggered (not just performatively).
- The decision being pre-mortemed is reversible enough that "kill" is a real option.
Fails when:
- Failure modes are qualitative ("the team morale will drop") and resist quantification.
- Political pressure prevents anyone from invoking the criteria once triggered.
- The decision is one-way and "kill" is not actually available.
Evidence
"If we see >30% churn in month 2, we pivot messaging."
Annie's framing: a pre-mortem without kill criteria is a feelings exercise; with kill criteria it is a decision gate. Direct application: campaign launches, product bets, messaging changes — every pre-launch session ends with a written threshold for stop or pivot.
— Annie Duke on Lenny's Podcast, 2026-04-28
Signals
- Every pre-mortem document ends with a numbered list of kill criteria.
- Post-launch reviews open with "did we hit any kill criteria?" before any other discussion.
- Teams that adopt the practice report cleaner post-mortems and fewer zombie projects.
Counter-evidence
For exploratory or research bets where outcomes are unpredictable, premature kill criteria can shut down genuinely useful learning. The discipline is conditional on bets where success looks knowable in advance.
Cross-references
- Force intuitions into explicit predictions so you can find out where you are wrong — the broader practice
- There is no such thing as a long feedback loop — find a correlated short signal — the measurement complement