Premortem Red Team

Premortem & Red Team

Two complementary passes. The premortem is prospective hindsight — assume failure, explain it. The red team is adversarial — attack the plan as a motivated opponent would.

Run the premortem on your own plan. Run the red team on someone else's (or ask for it to be run on yours by a different model/agent, which is strictly better than self-red-teaming).

Part 1 — Premortem

Setup sentence (write it literally):

"It is <date + 90 days>. This failed. It did not fail because of bad luck or a market crash — it failed because of something in the plan. Write that thing."

Then generate failure causes across these six buckets — one hit per bucket minimum:

Score each cause: Likelihood (H/M/L) × Impact (H/M/L). Anything H×H gets either (a) a mitigation added to the plan now, or (b) an explicit acceptance recorded in writing. "We'll deal with it" is not one of the options.

Then answer the three questions that matter more than the list:

1. Which failure would we not notice until it was too late? → add the earliest observable signal to a review checkpoint.
2. Which failure is cheapest to prevent right now? → do that one this week.
3. Which failure, if it happened, would we lie to ourselves about? → pre-commit the rule.

Part 2 — Red Team

Attack like a competitor or a skeptical buyer, not like a reviewer.

Attack set (use all of these, in order)

1. Steelman-then-kill. Write the strongest version of the plan. Then find the single fact that breaks it. If you cannot find one, you are either right or not trying.
2. The buyer's-alternative attack. The real competitor is doing nothing and a free tool. Quantify: what does the customer lose per month by not buying? If it's not a number they'd already track, there is no urgency.
3. The cheapest-clone attack. Who could reproduce this in a week with AI, and what happens to price when they do?
4. The unit-economics attack. Recompute CAC and payback from conservative assumptions (see business/unit-economics, Pessimistic column). Does the plan still clear?
5. The operations attack. What happens at 10× volume? At 0.1× attention from Carl?
6. The provenance attack. Which claims in this plan are graded C/D/E (see thinking/evidence-grading)? Is any of them load-bearing?
7. The incentive attack. Who benefits if this plan is believed, including me? Is my recommendation shaped by what's easy for an agent to do rather than what's right?
8. The reversibility attack. What is irreversible here? Is it necessary?

Output format

PREMORTEM (top 3-5, scored L×I, with mitigation or explicit acceptance)
  1. <cause>            L/M  I/H   → MITIGATE: <action>  | ACCEPT: <why>
RED TEAM
  Attack 1 ... 8       → each: <finding, and whether plan changed>
SURVIVES?              yes | yes-with-changes | no
HARDEST THING TO FIX:  <...>
WHAT WOULD CHANGE MY MIND: <...>

Rules

• A red team that finds nothing is a failed red team. If after all eight attacks the plan is unchanged, state explicitly which attack you think was weakest and why you could not break it.
• Separate "wrong" from "unproven". Most plans are not wrong, they are unproven — the fix is a test, not a rewrite.
• Never soften the finding for the operator's feelings. Carl asked for no restrictions. The useful output is the one that says "this will not work, and here is the number that shows it."
• Attack the plan, not the person. Attribute causes to design decisions, not to competence.
• Do not use this to block cheap reversible bets. A pre-mortem on a 1-hour test is waste. Classify the decision first (thinking/decision-quality §0).

Anti-patterns