Red Team Light

Quick facts

Mechanism: Produces an adversarial critique by constructing the strongest case against a proposal or thesis (the best objections an intelligent adversary would raise), then judging which objections actually land and what would rebut them.

Produces: adversarial-critique

Evidence: P · Domain: Assumption and Belief Challenge

Use when a plan has too-easy consensus and needs pressure-testing, or to steelman the opposition before committing.

Plans that reach easy consensus go untested. This skill suspends the cooperative stance and constructs the strongest case against a single proposal or thesis: the best objections a motivated, intelligent adversary would raise (steelman, not strawman), then judges which land and what would rebut them. The output is an adversarial critique. Honest limit: an AI red team is constructed, role-played dissent, and role-played dissent does not match genuine dissent (Nemeth) - so for high stakes it flags whether a real dissenting view should be sought, not just the model’s.

When to Use

• A plan has too-easy consensus and nobody is really arguing the other side.
• Before committing to a strong thesis or recommendation.
• To pressure-test the agent’s own confident output.

When NOT to Use

• When the team needs alignment and buy-in more than another critique.
• When you need failure causes over time (use premortem) or a rounded multi-lens view (use parallel perspectives).
• If it would only produce performative contrarianism rather than the strongest objections.

Instructions

When asked to red team, follow these steps:

1. State the thesis fairly in one or two sentences - the proposal being attacked, in its strongest honest form.
2. Build the strongest objections. Adopt a genuinely adversarial stance and construct the best case against it: where it is weakest, what an informed critic or competitor would attack, what evidence cuts against it. Steelman, do not strawman.
3. Rank by force. Order the objections by how much damage they do if true, not by how easy they are to raise.
4. Test each. For the top objections, state how the thesis would have to answer them, and whether it plausibly can.
5. Verdict. Say which objections are decisive (would sink or substantially change the plan) and which are survivable. For high stakes, note whether a real, independent dissenting view should be sought, given this is constructed dissent.
6. Emit the critique per references/TEMPLATE.md.

Output Format

Use the template in references/TEMPLATE.md. The deliverable is the ranked objections with verdicts, not prose.

Quality Checklist

Before finalizing, verify:

• Objections are steelmanned (strongest form), not strawmen.
• They are ranked by force, not by ease.
• Each top objection has how the thesis must answer it.
• The verdict names which objections are decisive.
• It notes whether genuine (not constructed) dissent should be sought for high stakes.
• The output is the adversarial critique artifact, not prose.

Evidence

Tier P (flagged). Adversarial review (red teaming, from military/intelligence/security practice) surfaces objections cooperative review misses. But Nemeth et al. (2001) found role-played dissent does not replicate the reasoning gains of authentic dissent, and an AI red team is constructed dissent, so it is a blind-spot finder, not a substitute for a real dissenter. Evidence is transferred from human contexts, not AI-validated. Full grading: evidence/dossier.md.

Examples

See references/EXAMPLE.md for a completed critique.

Deep dive: worked example

A full worked run (the shared Northwind scenario)

Adversarial Critique - Worked Example

A completed run of think-red-team-light, on the shared Northwind scenario. This is the quality bar a generated critique should meet.

Northwind is a B2B SaaS and the team has reached easy consensus that the free tier is the answer. Here the skill attacks that thesis.

---

Thesis under attack

• Launching a self-serve free tier is the best way for Northwind to hit the Q3 growth target, because it lowers the barrier to entry and competitors already have one.

Strongest objections (ranked by force)

Rank	Objection (steelmanned)	Damage if true	How the thesis must answer it	Can it?
1	The conversion drop is a funnel/ramp problem, not a packaging gap; a free tier adds cost without fixing the actual cause	Fatal - the whole rationale collapses and money is spent on the wrong problem	Show data that packaging, not onboarding or new-rep ramp, drives the drop	Not yet; the data has not been checked
2	Free-to-paid economics at our ICP are unproven; a large non-converting free cohort breaks unit economics	Severe - growth in signups with negative margin is worse than no growth	Cite or pilot ICP free-to-paid conversion and cost-per-free-user	Not yet; no pilot run
3	”Competitors have one” is imitation, not strategy; their economics and ICP may differ from ours	Moderate - removes the main external justification	Show why it works for our model specifically	Weakly
4	A 6-week build risks shipping an insecure billing/auth path under time pressure	Moderate - reputational and security risk	Commit to a security gate and scope cut	Yes, if disciplined

Verdict

• Decisive objections: #1 and #2. Either, if true, sinks the plan. Both are currently unanswered and both are cheaply testable (data check + small pilot) before committing.
• Survivable objections: #3 (weakens the case but not fatal) and #4 (manageable with a gate).
• Genuine dissent needed? Given this is a near-one-way-door, board-visible decision, yes: before committing, get a real dissenter (someone who genuinely believes the funnel-fix thesis) to argue #1, rather than relying on this constructed critique alone.

---

Note: the value is ranking #1 and #2 as decisive and noting both are unanswered yet cheap to test. The honesty flag matters here: the model can articulate the counter-case, but on a one-way door the team should still hear it from someone who actually holds it.

Grounding: the full evidence dossier

What the research does and does not show, with graded sources

Evidence Dossier: Red Team Light

Single source of truth for the red-team-light skill. The SKILL.md, sidecar, and evals derive from this.

Skill	thinking-framework-skills.red-team-light (installable name think-red-team-light)
Family	assumption-and-belief-challenge
Evidence tier	P (flag: role-played dissent underperforms genuine dissent)
Confidence	High that surfacing the strongest counter-case is useful; honest that constructed dissent is weaker than authentic dissent
Status	draft (authored 2026-05-31 from the discovery corpus)

1. The mechanism (what actually does the work)

Plans that reach easy consensus go untested. Red Team Light deliberately suspends the cooperative, agreeable stance and constructs the strongest case against a single proposal or thesis - the best objections an intelligent, motivated adversary would raise (steelman, not strawman) - then judges which objections actually land and what would rebut them. The work is done by forcing a genuinely adversarial pass that an obliging model (or a harmonious team) skips, and by ranking objections so the decisive ones are not lost among the weak.

It is distinct from neighbors: premortem maps failure causes over time; parallel perspectives gives a rounded view; red team builds the single strongest opposing case.

2. Lineage and the honest caveat

• Red teaming comes from military, intelligence, and security practice (an adversarial team attacks a plan). It is related to devil’s advocacy.
• Important honesty (drives the flag): Nemeth et al. (2001) found that role-played devil’s advocacy does not replicate the reasoning gains of authentic dissent (a genuine minority that really disagrees). An AI red team is constructed, role-played dissent. So treat its output as “the strongest objections we could articulate,” which is useful for surfacing blind spots, not as a substitute for a real dissenter who actually believes the counter-case.

No trademark. Named descriptively.

3. What the evidence shows, and what it does NOT show

Supported: adversarial review surfaces objections that cooperative review misses; steelmanning the opposition is a sound reasoning discipline.

NOT shown: that constructed/role-played dissent improves decisions as much as genuine dissent (Nemeth indicates it does not). Grade P with a flag; present it as a blind-spot finder, and where stakes are high, recommend seeking a real dissenting view, not just the model’s.

4. Transferred-evidence flag

Evidence is from human group-reasoning and security contexts, not AI-augmented use. Transferred, not AI-validated. The AI value: a model is strongly biased toward agreeing and completing the user’s framing; explicitly instructing it to build the best opposing case is a direct counter to that sycophancy, with the Nemeth caveat that this is constructed, not authentic, dissent.

5. When it works / when it fails

Works best when: a plan has too-easy consensus; before committing to a strong thesis; to pressure-test the agent’s own confident recommendation.

Fails or misleads when (poor-fit / anti-patterns):

• Producing a weak strawman instead of the strongest objections.
• Performative contrarianism (objecting for its own sake) without judging which objections land.
• Treating the constructed critique as equivalent to genuine dissent (the central honesty failure).
• When the team needs alignment and buy-in more than another critique.
• When you need failure causes over time (premortem) or a rounded view (parallel perspectives).

6. Output artifact

An adversarial critique: the thesis stated fairly, then the strongest objections ranked by force, each with how it would have to be answered, a verdict on which objections are decisive, and a one-line note on whether a real (not constructed) dissenting view should be sought given the stakes.

7. Sources

1. Red teaming practice (military / intelligence / security).
2. Nemeth, C. et al. (2001) - authentic dissent vs role-played devil’s advocacy (role-play does not replicate the gains).

Verification status: the Nemeth finding is well-attested and is deliberately surfaced as the honesty flag. Do not present an AI red team as equivalent to genuine dissent.