pre-mortem

name	pre-mortem
description	Imagine the project has already shipped and failed catastrophically — work backwards from the failure to identify the most likely causes BEFORE building. Forces concrete risk identification, not vague "what could go wrong" lists.
when_to_use	Apply BEFORE implementation begins: - architect, after writing ARCH but before gate:plan - pm, while breaking work into tasks (Pre-mortem section in PLAN-*.md) - security-officer, when threat-modeling - any time the feature is irreversible or high-blast-radius
effort	medium
allowed-tools	Read, Write
paths	["docs/plans/**","docs/architecture/**","docs/threat-models/**"]

Pre-mortem — fail-it-before-you-build-it

A retrospective for a project that hasn't happened yet. Surfaces real risks that "list every risk" prompts miss.

Originated in Gary Klein's research at MIT Sloan, now standard at AWS and other ops-mature orgs.

The 5-step pre-mortem

Step 1. Imagine you're 6 months in the future

The project shipped. It is a clear, public failure. There's a Reddit thread about it. The CEO is asking what went wrong.

Step 2. Write the post-mortem newspaper headline

One sentence. Concrete. Specific. Examples:

• ❌ Bad: "We had some quality issues."
• ✅ Good: "On 2026-09-12, the Stripe webhook handler deduplicated by raw body hash, so 30K customers were double-charged after Stripe retried delivery during a network blip."

The headline forces you to name the failure mode SPECIFICALLY.

Step 3. List every individual reason this exact failure happened

Brainstorm 10-15 reasons. Be specific. Each item should reference:

• A real component / file
• A real failure mode (race condition, schema mismatch, expired credential)
• A real human factor (oncall didn't see alert, runbook was outdated)

Reject hand-waves like "testing was insufficient." Replace with "we didn't write a property-based test for the dedup-key collision case."

Step 4. Rank by likelihood × severity

For each cause, score:

• Likelihood: 1-5 (1=once-in-a-decade, 5=monthly)
• Severity: 1-5 (1=cosmetic, 5=data loss / regulatory breach)
• Risk score: likelihood × severity

Top 3 by risk score → these are your highest-priority mitigations.

Step 4b. Classify risks — Tigers / Paper Tigers / Elephants

After scoring, classify each risk into one of three types:

🐯 Tigers — Real problems you personally believe could derail the project

• Based on evidence, past experience, or clear logic
• Should keep you awake at night
• Require concrete action
• Classify each Tiger by urgency:
  • Launch-Blocking: Must be resolved before shipping (broken core feature, regulatory blocker, data integrity risk)
  • Fast-Follow: Must be resolved within 30 days post-launch (performance issues, secondary features)
  • Track: Monitor post-launch, fix if it becomes an issue (edge cases, nice-to-haves)

📄 Paper Tigers — Concerns others might raise that you don't believe are real risks

• Valid-sounding on the surface but unlikely or overblown
• Not worth significant resource investment
• Worth documenting to align stakeholders and avoid repeated debates
• For each: explain WHY you don't believe it's a real risk

🐘 Elephants — Things the team knows about but isn't discussing openly

• Uncomfortable concerns: technical debt, team tension, unrealistic timeline, design that nobody likes
• Uncertain — you're not sure if it's a problem, but nobody is investigating
• Deserve explicit surfacing before launch — silent elephants become Tigers post-launch

Step 5. For each top-3 cause, write a guardrail in the plan

Each guardrail is a concrete change to the plan:

• A test that would have caught it
• A circuit breaker / feature flag
• A runbook entry
• A monitoring alert with specific SLO

If a top-3 cause CANNOT be mitigated within the time/budget, escalate to the user: "This plan accepts the risk of X with no mitigation."

Template — add to PLAN-*.md

## Pre-mortem

Six months from now, this project failed. Headline:

> <one-sentence failure headline>

### Top reasons (likelihood × severity)

| Cause | L | S | Risk | Mitigation in plan |
|---|---|---|---|---|
| <specific cause> | 4 | 5 | 20 | <Task #N: write idempotency test> |
| ... | | | | |

### 🐯 Tigers (real risks — require action)

| Tiger | Classification | Mitigation | Owner | Due |
|-------|---------------|-----------|-------|-----|
| <risk> | Launch-Blocking | <concrete action> | <team/person> | <date> |
| <risk> | Fast-Follow | <concrete action> | <team/person> | <date> |
| <risk> | Track | <monitoring approach> | <owner> | post-launch |

### 📄 Paper Tigers (overblown — document to align stakeholders)

- **<concern>**: Not a real risk because . If  changes, revisit.



 : Nobody is talking about this. Suggested conversation: "".



  — accepted because . Owner: .

Common failure modes by archetype

Quick start — most-common pre-mortem causes per archetype:

Archetype	Common failure
fintech / commerce	Idempotency-key collision; double-charge during retry storm
healthcare	PHI leak via debug log; BAA not signed with vendor
web3	Oracle staleness; flash-loan exploit on bonding curve
mlops	Training/serving skew; model drift undetected
iot-embedded	OTA bricks devices in a region with no recovery path
data-platform	Late-arriving data overwrites correct values
ai-system / agent-product	Prompt injection exfiltrates other users' data
enterprise-saas	Cross-tenant data leak via RLS gap
cli-tool	Destructive flag with no confirmation (rm -rf equivalent)
library	Breaking change in minor version bump

Anti-patterns in pre-mortems

❌ Vague risks. "Performance might be a problem." Be specific: which operation, at what load, what's the SLO.

❌ Cosmic risks. "AWS could go down." Yes, but that's not actionable. Focus on what you can mitigate.

❌ Defensive list. Listing risks you've already mitigated to look thorough. Only list risks the current plan does NOT yet address.

❌ Skip the headline. Without the headline, the team won't believe the failure scenario is real.

When to skip

• nano project_size — pre-mortem is overhead.
• Pure refactor with full test coverage — guardrails already exist.
• Bug-fix with one-line repro — risk is well-bounded.