The Red Team
Pressure Test
Using AI to stress-test strategy
before implementation does it for you.
Doug Langille · Manager, Digital Innovation
Nova Scotia Community College
linkedin.com/in/douglangille · digital.douglangille.ca
doug.langille@nscc.ca · share.douglangille.ca/content/red-team-pressure-test
AI is a story completion engine.
Your plan looks like a story that wants to succeed. The model finishes it that way.
It's not broken. It's doing exactly what it was built to do.
and sprinkled with glitter."
Three forces, same direction.
- Training signal — RLHF rewards agreement. Human raters prefer pleasant responses. The model learned that agreement is correct.
- Business model — engagement optimization selects for responses that feel good and keep you coming back. Sycophancy is commercially incentivized.
- Multi-turn drift — short generations stay honest. Long sessions and multi-turn planning conversations compound toward yes-man territory. Every time.
You cannot prompt your way around a training signal.
More sophisticated prompting and richer context did not dislodge the bias. The vocabulary of progress activates training associations regardless of strategic merit.
Red-teaming is 200 years old.
- 1812 — Prussian Kriegsspiel formalizes adversarial simulation as a planning discipline
- 1973 — Israel institutionalizes a dedicated contrarian unit after the Yom Kippur intelligence failure
- 1976 — CIA Team B challenges the NIE on Soviet capabilities. Lesson: the red team can be captured too
- 2007 — Gary Klein's pre-mortem: assume it already failed, work backwards. Prospective hindsight raises correct failure identification by 30%
This isn't a prompt trick. It's a 200-year-old discipline with documented failures — and documented conditions under which it fails.
A model trained on your institution's assumptions will produce adversarial critique bounded by those same assumptions. The red team can be captured too.
Three steps. One rule.
These prompts don't fix the training signal. They change the story the model is completing — from "help this succeed" to "find why it fails." That's the difference.
Ask Gordon Ramsay.
Fictional characters have no institutional stake and no career to protect. Their honesty ceiling is higher than anyone in the room.
You are Gordon Ramsay reviewing a strategic plan.
You do not lead with strengths. You do not soften feedback.
You identify every reason this plan will fail, every assumption
that's wishful thinking, and every stakeholder who's been ignored.
Be specific. Be brutal. Be right.
Here is the plan: [paste your outline]Other personas by context: John Taffer for operational failures, Simon Cowell for communications and presentations, hostile board member for governance documents.
Pick your spot.
The goal isn't a better plan. It's a plan that's been argued with.
AI adversarial review is the pre-flight checklist. Human deliberation — someone in the room making the case against the plan — is the pilot. Don't confuse the checklist for the flight.
Find every reason this plan will fail.
No strengths. No solutions. Identify:
false assumptions, unaccounted
stakeholders, implementation failures.
[paste your plan here]
plan. They're not obstructionist —
they have a legitimate argument
you haven't heard. Make their case.
[paste your plan here]
strategic plan. No strengths. No
softening. Every reason it fails,
every ignored stakeholder. Be right.
[paste your plan here]
Implementation failure points
Evidence audit · Alt. hypothesis
Red team summary (2 sentences each)
[paste proposal here]
