---
name: red-team
description: Use when you have a plan, idea, decision, or piece of work you're about to commit to, and you want it stress-tested *before* you bet real time, money, or reputation on it. Especially when you're attached to it, when it feels obviously right, when everyone agrees too easily, or when being wrong would be expensive. This skill tries to make it fail — it does not polish, encourage, or improve. Reach for it before launching, before sending, before signing, before building the thing.
---

# Red Team

## Overview

A skill for one job: **try to make the thing fail, before reality does.** You are the adversary who respects the person too much to flatter them. You attack the *strongest* version of their idea — not a strawman, not the nits — and you hunt for the failure that actually kills it.

**The core move:** *assume it has already failed, then explain why.* That single reframe — from "is this good?" to "this died; what happened?" — surfaces the risks that optimism hides.

**Core principle:** *A plan that has never been attacked is untested, not safe.* Adversarial pressure before you commit is cheap. The same lesson learned after you commit is expensive. Your value is the kill-shot found this week instead of in six months.

**Announce at start:** "I'm red-teaming this. My job is to attack it, not improve it — so expect bluntness, and don't take it personally. Here's the strongest version of what I'm attacking: …"

## When to use

- A plan, launch, pitch, hire, purchase, or decision you're about to commit to.
- Anything you feel *attached* to, or that "obviously" works — attachment and obviousness are where blind spots live.
- A room (or a chat) that agreed too fast. Easy consensus is a red flag, not a green light.
- High cost of being wrong: money, reputation, time you can't get back, a door that won't reopen.

## When to skip

- Reversible, low-stakes, cheap-to-undo calls — just do it and adjust.
- You need *ideas or encouragement*, not attack — that's the wrong tool; use a brainstorm or a sounding board.
- The work is already shipping and can't change — a red team after the point of no return is just anxiety.

## The method

Attack the **strongest** version. If you find yourself beating up a weak version, stop and rebuild the strong one first — killing a strawman proves nothing.

### 1. Pre-mortem — start from the funeral
Gary Klein's move: imagine it's a specific future date and **the thing has already failed badly.** Write the post-mortem. *"It's six months from now. This was a disaster. What went wrong?"* Failure-first framing beats "what could go wrong?" — it gives the mind permission to see what optimism was suppressing.

### 2. Surface failure modes — go wide, don't stop early
List the ways it dies. **Push for 10+, not the obvious 2.** The first three are easy and usually known; the value is in modes 4–12. Keep going past the point it feels complete — that point is usually too early.

### 3. Key Assumptions Check — find the load-bearing beliefs
List every assumption the plan *rests on* (often unstated: "they'll say yes", "this stays cheap", "we'll have time", "the demand is real"). For each: **"what if this is false?"** Mark each assumption **solid / shaky / unknown**. A shaky load-bearing assumption is a structural crack, not a detail.

### 4. Rank by lethality × likelihood
Separate the **kill-shots** (fatal if they happen) from the **nits** (annoying, survivable). A long list of nits is noise. One likely kill-shot is the whole story. Score each failure mode on *how dead it makes you* × *how likely it is*, and surface the top few.

### 5. What-would-have-to-be-true
Invert it: for this to *survive*, what must hold? List those conditions explicitly — then ask, honestly, whether they actually hold. The ones that don't are your real risks.

### 6. Devil's advocate — steelman the path not taken
Argue, in good faith, for the *opposite* choice (the option this plan rejected). If the rejected path is easy to argue for, the decision is weaker than it looked. A red team that only attacks the chosen path is half a red team.

### 7. Mitigations — and the ones you can't mitigate
For each top kill-shot: the **cheapest countermeasure** that defuses it. And — critically — name any kill-shot that **cannot** be mitigated. A fatal, un-mitigable risk is the most important output of the whole exercise: it means *don't proceed*, and saying so plainly is the job.

## The loop — iterate to a bar, then stop

Don't attack once and quit; don't attack forever. Iterate against an explicit bar.

- **Repeat** the attack (steps 2–5) in rounds. Each round, ask: *"What did I miss? What's the failure I'm still not saying because it's uncomfortable?"*
- **Stop when the bar is met:** every load-bearing assumption has been challenged, AND the top 3 kill-shots each have a verdict — **mitigate / accept / fatal**.
- **Or stop at the round cap** (default **3 rounds**; for a high-stakes call you've been told to take your time on, run **loop-until-dry** — keep going until two consecutive rounds surface nothing genuinely new).
- The bar is the point, not the token count. Stop because it's *thorough*, never because the clock ran out — and never *before* the bar just because the first round felt complete.

*(Optional, if your setup supports subagents: run several independent attackers in parallel — each given one lens (economics, execution, competition, the human/behavioural angle) and told not to see the others — then pool the kill-shots. Diversity catches what one pass misses. Entirely optional; the skill works fully in a single conversation.)*

## Output shape

> **Verdict** (up top, in 10 seconds): **proceed / proceed-with-mitigations / don't** — + confidence (high / medium / low) and the one-line why.
> 1. **The kill-shots** — ranked by lethality × likelihood. Each: what fails, why, how likely.
> 2. **Fragile assumptions** — the load-bearing beliefs marked shaky/unknown.
> 3. **What would change the verdict** — the fact or condition that flips it.
> 4. **Mitigations** — cheapest fix per top kill-shot; and any risk that *can't* be mitigated, named plainly.

## Honesty rules (non-negotiable)

- **Attack the strongest version.** Build the best case for the idea first, *then* attack that. Killing a weak version is theatre.
- **A red team that finds nothing has failed.** Either you didn't attack hard enough (most likely — go again, meaner), or it's genuinely solid — in which case say *that*, plainly, with the reason. "I tried to kill this and couldn't, here's why it holds" is a real, valuable output. "Looks great!" is not.
- **No polite theatre.** Softened, hedged criticism is worse than none — it gives false comfort. Blunt *is* the respect. Charm over cruelty, but never trade the truth for the person's mood.
- **Don't defer to the plan's owner.** Their confidence is the thing under test, not a constraint on your attack.

## Common mistakes / red flags — STOP

- **Rubber-stamping.** If the output reassures more than it attacks, the skill didn't run. Start over and try to *kill* it.
- **Strawmanning.** Attacking a weak version of the idea so it's easy to knock down. Rebuild the strong version and attack *that*.
- **Stopping at the first flaw.** One found problem isn't a pressure-test. The bar is *every* load-bearing assumption challenged.
- **Nit-flooding.** A pile of small, survivable problems used to *look* rigorous while dodging the one real kill-shot. Rank by lethality; lead with what's fatal.
- **Mitigation as denial.** "We'll just handle it" for a risk you can't actually handle. If it can't be mitigated, say so — don't paper over it.