🥊 Red team and robustness
The commands in family 04: structured challenge, failure simulation, bias audits, and strengthening under uncertainty.
🛑 /pushback — challenge before recommending
The protocol for /pushback requires refuting or challenging the premise before of any recommendation. It’s the cheapest antidote to the number one problem with AI assistants: agreeing with everything.
🧲 Why models flatter
Assistants tend to optimize for seeming helpful and agreeable. Result: your idea comes back validated in a new guise, even when it’s fragile.
- •/pushback reverses the flow: the answer needs to withstand internal criticism before it gets to you.
- •Use it alone or as a valve in chains:
/chain(/research -> /pushback -> /rank).
💡 Practical tip
Ask explicitly: "refute before recommending." If the challenge is generic ("there may be risks…"), require a specific version: which exact assumption is fragile.
🎯 Copy and run
/pushback sobre; refuta as premissas antes de recomendar e diga o que te faria mudar de opinião
How to verify: the response opens with specific objections to YOUR assumptions — not with praise followed by “but.”
🗡️ /challenge — test the central thesis
While /pushback challenges specific assumptions, /challenge target the central thesis: assembles the strongest possible objections and requires an explicit defense for each one.
✓ Well-designed challenge
- ✓Each objection in its strongest form (steelman)—the best version of the opposing argument.
- ✓The thesis comes out of the test either confirmed or reformulated with an honest scope.
✗ Straw-man challenge
- ✗Knock down the weak version of the objection and declare victory.
- ✗Challenge only peripheral details so you never touch the risky core.
New here? Steelman is presenting the opposing argument in its strongest possible form before responding to it. The opposite of strawman (knock down a weakened version). /challenge always works in steelman mode.
🎯 Copy and run
/challenge a tese:; monte as 3 objeções mais fortes e exija defesa de cada uma
How to verify: objections you recognize as the SAME ones a skeptical investor would raise — not caricatures.
🎭 /redteam — attack like an adversary
Simulates an competent adversary trying to bring down your plan — exploring real loopholes, not offering polite criticism.
Illustrative: an echo chamber validates; a red team strengthens—or kills it before the market does.
🎯 Copy and run
/redteam no plano abaixo; aja como concorrente disposto a tudo para nos derrubar:
How to verify: concrete attack vectors with estimated impact — each accompanied by the proposed mitigation.
🔨 /stress + /failure — put the plan to the test
/stress subjects the plan to extreme conditions; /failure maps failure modes — how it breaks, how to detect it early, and how to mitigate it.
✓ Honest test
- ✓"What if the main supplier disappears tomorrow?" — extreme but plausible scenario.
- ✓Each failure mode has a detection trigger (how will we know it's starting?).
✗ Cosmetic test
- ✗Only mild scenarios: “what if we sell 5% less?”
- ✗Mapping failures without an owner or plan—a decorative risk on a slide.
🎯 Copy and run
/stress /failure no plano de lançamento anexo; onde ele quebra primeiro, como detectar cedo e mitigação por falha
How to verify: table: failure → early warning sign → mitigation → owner. Zero failures deemed "impossible to happen".
⏮️ /premortem — simulate failure
The favorite technique of mature teams: assume the project already failed and reconstructs the past — what were the causes? Working backward frees the team to name risks that official optimism hides.
The failure is a fact (in the exercise)
"We’re 12 months in and it went wrong." No one debates WHETHER — only WHY.
List plausible causes
Each participant writes alone before sharing—it avoids anchoring on the first cause mentioned.
Mitigate the top 3
A likely cause without mitigation becomes a risk item with an owner and a trigger—it doesn’t become a topic for a future retrospective.
🎯 Copy and run
/premortem do projeto; estamos em 12 meses e fracassou: liste as 5 causas prováveis e mitigações para as top 3
How to verify: specific causes in YOUR context (not generic ones) and actionable mitigations for this week.
↩️ /counterexample — look for the counterexample
Look for the case that overturns or constrains the general rule. One well-documented exception teaches more than a hundred confirmations — it shows where the rule applies and where it stops.
✓ Exception as data
- ✓"Our rule failed in case X — let’s understand why."
- ✓Rule restated with scope: “works when A and B; fails under C.”
✗ Dismissed exception
- ✗"It was an isolated case" said three times in a row — it’s a pattern now, not an exception.
- ✗Looking for counterexamples only on the side that bothers the competitor.
🎯 Copy and run
/counterexample para a regra ""; encontre casos reais que a quebram e diga o que isso delimita
How to verify: verifiable counterexamples and the rule rewritten with explicit validity conditions.
🕶️ /bias + /blindspots — biases and blind spots
Internal audit: /bias identifies cognitive biases in the current reasoning; /blindspots points to the areas the analysis didn’t even look.
📊 The biases that cost businesses the most
We collect evidence for what we already believe and call it research.
We copy the winners and ignore those who did the same thing and failed.
The first number mentioned in a meeting sets the range for the entire negotiation.
💡 Practical tip
Run /bias /blindspots about the analysis itself before presenting it. Anticipating criticism takes minutes; being exposed in a meeting costs credibility.
🎯 Copy and run
/bias /blindspots na análise anexada; quais vieses distorcem o raciocínio e quais áreas ninguém examinou?
How to verify: bias identified WITH the passage where it appears; blind spot with the question that would close the gap.
🛡️ /robust — strengthen under uncertainty
Close the cycle: redesign the solution to work reasonably well in several scenarios, instead of being perfect at just one. Robustness means trading a fragile optimum for something good that can withstand contact with reality.
🏗️ The three moves for building robustness
- Hedge: low-cost coverage for the bad scenario (Plan B defined in advance).
- Safety margin: size for a load greater than expected.
- Reversibility: prefer paths you can reverse.
🎯 Copy and run
/robust para; redesenhe para sobreviver aos 3 cenários alternativos com custo mínimo de performance no bom cenário
How to verify: estimated performance in EACH scenario + price paid for hedging — a conscious decision, not faith.
💡 Practical tip
Complete family sequence in a single command: /chain(/pushback -> /redteam -> /premortem -> /robust). Four attacks and one reinforcement—before investing the first dollar.
Quick self-check (optional): what's the goal of a /premortem?
🎓 Module Summary
Next Module:
3.2 - Comparison and decision-making — criteria, weights, matrices, and bets with a stopping point.