Pre Mortem And Red Team

Pre-Mortem and Red Team

When to use

Stress-test a recommendation, deal, or program before commitment — structured pre-mortem (assume failure, work backward), dis-confirming evidence hunt, devil's-advocate pass, and decision-quality score.

Use after analysis is drafted (business-case, evaluate-strategic-option, map-competitive-landscape) and before board, IC, or sponsor sign-off.

What this skill does not do

• Does not veto decisions — surfaces failure modes and evidence gaps; strategist decides proceed / pivot / stop.
• Does not replace legal, compliance, or security review — flags when those lanes are needed.
• Does not invent dis-confirming data — names searches and witnesses needed; tags model knowledge [verify].

Preconditions

| Input | If missing | |---|---| | The recommendation or option under test | Ask — cannot pre-mortem without a stated bet | | Decision audience and stakes | Ask once; calibrates depth | | Supporting analysis artifact | Proceed from user summary; flag thin evidence |

Provisional mode

Thin input: run abbreviated pre-mortem on stated recommendation only; confidence structured first pass; list evidence needed for full red-team.

Trust spine

• Confidence bands (option-evaluation):
• High: Dis-confirming search run; devil's-advocate challenges tied to evidence; decision-quality score with explicit kill triggers.
• Medium: Pre-mortem complete; some dis-confirming paths unexplored — listed.
• Low: Recommendation paraphrase only — scaffold, not clearance.
• Strategic advice vs. support (mandatory): Surfaces challenges; does not silently pick stop/go — [review] on proceed recommendation.
• Accountability gap: Decision-quality score forces engagement; no "looks fine" conclusion without scored dimensions.
• Per ../../references/trust-conventions.md for GATE before external distribution.

Assumption audit

| Assumption | Status | If wrong, what breaks | |---|---|---| | Success conditions are explicit | [stated / implied / missing] | Pre-mortem targets wrong failure | | Key dependencies named | [complete / partial / unknown] | Blind spots in failure tree | | Dissenting stakeholders identified | [named / unknown] | Devil's advocate too friendly | | Evidence for headline claim | [sourced / partial / narrative only] | Red team argues past the analysis |

Red flags

Pre-commitment review non-negotiables:

• MUST NOT declare proceed without a completed devil's-advocate pass — consensus without challenge tends to cause groupthink approval.
• Do not proceed to High confidence without listing at least three dis-confirming evidence items sought (found or NOT FOUND — gap).
• Hard stop: recommendation unchanged after red-team with zero mitigation actions — that signals checklist theatre, not decision quality.
• MUST NOT treat absence of dis-confirming evidence as confirmation — absence may mean search was too shallow.

Outside-view step

Before the pre-mortem narrative, anchor with a base-rate failure check:

1. Name the reference class of similar decisions (deals, programs, strategies).
2. State the class base rate for failure, delay, or value miss (tag sources).
3. Ask: does this proposal assume it will beat the class base rate? If yes, name the mechanism — otherwise flag optimism.

Always compare the proposal to class base-rate failure because teams systematically believe their plan is the exception — untreated base rates tend to cause repeat of known failure modes.

Workflow

Before Step 1: Read and apply ../../references/trust-conventions.md — source-tagging, [verify] on model-only numbers, load-bearing assumptions at top, numbers provenance, confidence labeling, and board-ready gate.

Step 1: Freeze the bet

Restate recommendation, success criteria, timeline, and stakeholders in one paragraph. Confirm with user.

Step 2: Pre-mortem (prospective hindsight)

Assume the initiative failed badly 12–18 months out. Work backward: list 5–8 plausible failure stories — each with early warning signal and leading indicator that could be monitored now.

Step 3: Dis-confirming evidence pass

For each load-bearing claim in the recommendation, ask: what evidence would prove this wrong? Run (or specify) searches for:

• Contradicting market or competitor data
• Internal track record contradicting the plan
• Expert or stakeholder dissent

Record each item: DIS-CONFIRMING: [claim] — sought [evidence] — result [found / not found / gap].

Must actively hunt dis-confirming evidence because confirmation bias causes teams to assemble only supporting facts before major commits.

Step 4: Devil's-advocate pass

Assign a explicit challenger stance (even if solo user — "argue as skeptical IC member"):

• Steelman the stop or defer case in three bullets.
• Name the weakest link in the recommendation's logic chain.
• State what would change the decision — kill triggers.

Do not soften the devil's-advocate voice to preserve rapport — softened challenge tends to cause false confidence.

Step 5: Decision-quality score

Score 1–5 (1 = weak, 5 = strong) on:

| Dimension | Score | Rationale | |---|---|---| | Clarity of success criteria | | | | Quality of dis-confirming search | | | | Outside-view / base-rate alignment | | | | Mitigations for top failure modes | | | | Kill triggers defined and dated | | |

Overall decision quality: [average or min dimension] — proceed only with [review] if any dimension ≤2.

Step 6: Red-team verdict

Synthesize: Proceed / Proceed with conditions / Defer / Stop — each condition or kill trigger explicit. GATE before board/exec externalization.

Output format

CONFIDENCE: [defensible recommendation | structured first pass]
BET UNDER TEST: [one paragraph]

BASE-RATE CHECK: [reference class] — failure/delay rate [X%] — proposal assumes [beats class / aligned / worse than class]

PRE-MORTEM FAILURE STORIES:
1. [story] — early warning: [...] — leading indicator: [...]
[repeat]

DIS-CONFIRMING EVIDENCE:
| Claim | Evidence sought | Result |
|---|---|---|

DEVIL'S ADVOCATE (stop/defer case):
- [...]

DECISION-QUALITY SCORE:
| Dimension | 1-5 | Note |
|---|---|---|
OVERALL: [score] — [review] if ≤2 on any dimension

VERDICT: [Proceed | Proceed with conditions | Defer | Stop]
CONDITIONS / KILL TRIGGERS: [...]

Worked example

Input: Approve $8M CRM replacement; 9-month timeline; projected $5M benefits.

Expected output (excerpt):

BASE-RATE CHECK: CRM replacements in peer set — 45% miss timeline by >6 months [verify]
DIS-CONFIRMING: adoption at go-live — sought prior rollout NPS — NOT FOUND — gap
DEVIL'S ADVOCATE: Defer until integration dependency on Identity phase proven
DECISION-QUALITY SCORE: dis-confirming search = 2 — overall 2.4 [review]
VERDICT: Proceed with conditions — kill if Identity slip >8 weeks

Quality checks before delivering

• [ ] Pre-mortem has ≥5 failure stories with early warnings
• [ ] Dis-confirming evidence table complete (gaps explicit)
• [ ] Devil's-advocate pass present and not softened
• [ ] Decision-quality score filled; low dimensions flagged [review]
• [ ] Base-rate / outside-view check included

Propose profile update

When a stable convention surfaces during this run (thresholds, naming, tone, output format, or recurring corrections), propose a profile update: show the exact diff against ~/.claude/plugins/config/claude-for-strategy/consulting/CLAUDE.md (org-wide facts go to org-profile.md), ask for confirmation, and write only on yes. Only /consulting:practice-setup auto-applies a full profile write.

Outputs

Follows plugin CLAUDE.md § Outputs. Next: revise /transformation:business-case, run /consulting:exec-memo for IC, or /consulting:doc-reviewer on updated narrative.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: daddia
• Source: daddia/claude-for-strategy
• License: MIT

Install and usage instructions live in the source repository linked above.