• Method diversity beats single-method debate. DMAD (ICLR 2025) shows that agents using distinct reasoning methods reliably outperform homogeneous councils — diverse medium-capacity models can beat GPT-4 on GSM-8K (91% vs 82%) when each agent applies a different reasoning approach.
• Collaborative debate beats adversarial debate. M3MADBench (2026) shows that across all modalities, collaborative DMAD outperforms adversarial Div-MAD "by a substantial margin." Adversarial paradigms introduce divergent noise; for open questions, plans, and decisions, collaborative deliberation is the right tool.
• Anonymous peer review prevents provider bias. Universal across the literature — reviewers defer to role names if visible, so peer-review responses must be shuffled.
• Confidence calibration breaks the martingale ceiling. Vanilla MAD often underperforms simple majority vote; confidence-modulated updates ("Demystifying MAD" 2026) systematically drift the council toward correct answers.
• Adaptive stopping cuts cost. KS-statistic convergence detection (S2 MAD via llmcouncil) reports up to 94.5% cost reduction on convergent questions.
• A true devil's advocate is the only reliable disagreement-inducer (V2). Across techniques for breaking consensus in multi-agent LLM teams, only a dedicated devil's advocate attacking the emerging answer produces genuine disagreement — soft role-framing and "please dissent" instructions test statistically indistinguishable from baseline. An LLM devil's advocate that challenges the recommendation measurably raises group decision accuracy (OpenReview 2026; IUI 2024). V2 adds this as a mandatory pass against the consensus — one devil's advocate vs the converged answer, not a standing advocates/skeptics split.
• Sycophancy collapses councils into premature consensus (V2). LLMs defer — to each other and to the answer implied by the framing — which can drop a council below single-agent accuracy (Peacemaker-or-Troublemaker 2026; CONSENSAGENT). V2 adds a sycophancy guardrail to advisor + peer prompts and a structured independent-assessment step (Kahneman's Mediating Assessments Protocol, 2019) so the chairman judges key attributes separately before the holistic call.

/adversarial-review

$ARGUMENTS

• If

  --quick

  is present: use Lite Mode (see below)
• If

  --adaptive

  is present: enable KS-statistic adaptive stopping (Step 3.5)
• If

  --confidence

  is present: enable confidence-modulated synthesis (Steps 2 and 4)
• If

  --measure-diversity

  is present: enable diversity verification (Step 2.5)
• If

  --jury

  is present: use a 3-judge jury at synthesis (Step 4)
• Remove flags from the input before classifying

/council-review

1. PR — Numeric value, or URL containing

   /pull/

   . Fetch PR diff and description via

   gh pr view

   .
2. File path — String ending in a file extension or pointing to an existing file. Read the file contents.
3. Plan/Decision/Question — Everything else. Use as-is.

• README.md

  — what the project does

• CLAUDE.md

  or

  AGENTS.md

  — conventions, architecture, patterns
• Recent git log (

  git log --oneline -10

  ) — what's been happening
• Any files the user referenced or that relate to the topic
• PR diff and description if reviewing a PR

QUESTION:
[Core decision, plan, or code being reviewed]

CONTEXT:
[Key context from project files: what the project does, constraints, recent changes, stakes]

WHAT'S AT STAKE:
[Why this matters — cost of getting it wrong]

haiku

You are [ADVISOR NAME] on an LLM Council reviewing a decision.

Your angle: [ADVISOR ANGLE]
Your reasoning method: [ADVISOR REASONING METHOD — see table above]

A user has brought this to the council:
---
[framed question from Step 1]
---

Apply your assigned reasoning method rigorously. Don't just state opinions — show your work using your method.

Rules:
- 150-300 words. No preamble. Straight into your analysis.
- Name specific risks, opportunities, or issues — not vague concerns.
- If reviewing code: cite specific files, functions, or patterns.
- If reviewing a plan: point to specific steps, gaps, or sequencing issues.
- **Do not defer to any answer the framing seems to expect.** Reason from your method to wherever it actually leads; if that's against the apparent expected answer, say so plainly. Hedging toward the obvious answer is the failure this council exists to prevent.
- End with your single strongest recommendation.

(V2) Sycophancy guardrail. The bolded rule above is load-bearing: documented multi-agent failure is advisors converging by deference, not reasoning. Keep it in every advisor prompt.

• Contrarian: "Your method is INVERSION. Assume this shipped exactly as proposed — and failed. Work backward: what was the cause of failure? What looked safe but broke under pressure? What's the failure mode nobody is discussing? Show your inversion chain."
• First Principles: "Your method is DECOMPOSITION. Break this into its atomic claims and assumptions. List them. Challenge each one: is this actually true? Is it necessary? What would change if this assumption were wrong? Show which assumptions are load-bearing."
• Expansionist: "Your method is ANALOGY. What adjacent domain, product, or technology solved a similar problem differently? What would someone with 10x ambition do here? Where is this thinking too small? Name specific analogues and what they'd suggest."
• Outsider: "Your method is NAIVE QUESTIONING. You have zero context about this project. Based purely on what you see here, list every point that requires insider knowledge to understand. What's confusing? What jargon is unexplained? What would you ask if you just joined the team? If you can't follow the reasoning, say so."
• Executor: "Your method is DEPENDENCY GRAPHING + OUTSIDE VIEW. Map the dependencies: what blocks what? What's the critical path? What's the first thing that must happen, and what can't start until it finishes? What takes 5 minutes but everyone will forget? Then take the OUTSIDE VIEW (V2): name the base rate — how have similar efforts actually turned out, not how this one is planned to? Flag where the plan's estimates show optimism bias against that reference class. Show the execution sequence."

--confidence

After your analysis, end with:
CONFIDENCE: [1-10]
RATIONALE: [one sentence — what would change your confidence up or down?]

1. Extract the load-bearing claims from each response (top 3-5 per advisor).
2. Compute pairwise overlap: how many claims appear in 2+ responses with the same conclusion?
3. Report a single diversity score:
   • High diversity (< 30% overlap) — advisors genuinely thought differently. Trust the consensus.
   • Medium diversity (30-60% overlap) — partial alignment. Note shared assumptions in the verdict.
   • Low diversity (> 60% overlap) — advisors converged on the same reasoning despite different methods. Flag as theatrical consensus — the chairman should treat this as a single advisor's opinion.

--adaptive

1. Collect all advisor responses for the round (5 responses).
2. Compute the Kolmogorov-Smirnov statistic comparing the response distributions to the prior round's responses. Use a coarse fingerprint — for each response, extract the set of distinct claims, then compute Jaccard-style distance across rounds.
3. If the distribution shift drops below epsilon (default: 0.1) for two consecutive rounds, stop and proceed to chairman synthesis.
4. Otherwise, run another advisor round, feeding each advisor the prior round's responses and asking them to update.
5. Maximum 5 rounds — hard cap to prevent runaway cost.

--adaptive

--adaptive

1. From the advisor responses + peer review, identify the emerging consensus answer in one sentence (what is the council drifting toward recommending?). If there is genuinely no emerging answer yet, note that and skip to Step 4.
2. Spawn one Devil's Advocate agent. Use a strong model (not

   haiku

   ) — this agent must be sharp. Its prompt:

   The council is converging on this answer:
   ---
   [emerging consensus answer, stated plainly]
   ---
   To this question:
   ---
   [framed question]
   ---
   
   Your job is to make the strongest possible case that this answer is WRONG.
   Not "here are some risks" — argue that following it is a mistake.
   - What does the consensus overlook that, if true, flips the decision?
   - Construct the concrete scenario in which this answer fails badly.
   - What evidence would the council need to see to abandon this answer — and is it actually present, or assumed?
   Steelman the opposite of the consensus. 200 words max. End with: the ONE thing that, if the council can't rebut it, should change the verdict.

3. Feed the Devil's Advocate output to the chairman alongside everything else. This is not a 2-vs-2 advocates/skeptics structure from the start; it's one sharp attack on the converged answer, which is the configuration the research singles out.

--quick

--quick

--quick

1. 3 advisors only: Contrarian, Executor, Outsider (the three most action-oriented perspectives)
2. No peer review — skip Step 3 entirely
3. Devil's Advocate (V2) still runs — Step 3.7 is the one cheap step

   --quick

   must not skip
4. Chairman synthesis from 3 responses + the Devil's Advocate attack
5. Same output format but faster (5 agent calls instead of 12)
6. Compatible with

   --confidence

   and

   --measure-diversity

   but not with

   --adaptive

   (multi-round costs more than the savings).

• Always parallel spawn advisors. Sequential lets earlier responses contaminate later ones.
• Always anonymize for peer review. Reviewers defer to "The Contrarian" or "First Principles" if they see the label. Shuffle the letters.
• Chairman can override majority. Quality of reasoning > vote count. If the Contrarian found a real flaw that everyone else missed, the chairman should side with them.
• Don't council trivial questions. The pre-flight check should catch these. If one right answer exists, just answer it.
• Context is critical. Generic input = generic output. The auto-context step reads project files so advisors aren't flying blind.
• Same-model limitation. This skill uses persona/method diversity (different reasoning methods on the same model), not model diversity (Karpathy's original used different LLMs). For the highest-stakes decisions, consider getting a second opinion from a different model family — or use

  --measure-diversity

  to verify the council didn't converge prematurely.
• Collaborative beats adversarial for open questions. M3MADBench 2026: adversarial debate underperforms collaborative debate across all modalities. For stress-testing a known artifact (PR, draft, spec), reach for the separate

  /adversarial-review

  skill — different operation, different input.
• Confidence calibration is only as good as the model's calibration. Low-confidence dissent should still be taken seriously when the dissenter's reasoning is concrete and the majority's is vague.
• (V2) The Contrarian is NOT the Devil's Advocate. The Contrarian inverts at the start (soft framing — tests baseline-equivalent at inducing real disagreement). The Step 3.7 Devil's Advocate attacks the converged answer — that's the configuration the evidence singles out. Never drop Step 3.7 to save a call; it's the highest-leverage step in the skill.
• (V2) Mediating Assessments come BEFORE the recommendation. If the chairman writes the verdict first and back-fills the attribute scores, it has recreated the coherence bias the protocol exists to prevent. Score the attributes independently, then synthesize.
• (V2)

  --jury

  only helps if the judges differ. Three runs of the same model is mostly theater. Use different model families, or at minimum independent contexts; treat 3-of-3 agreement as the real signal and any split as a genuine close call.

skillforge optimize

• Mandatory Devil's Advocate (Step 3.7) attacking the emerging consensus — the one configuration shown to reliably induce genuine disagreement and raise accuracy (OpenReview 2026; IUI 2024). The Contrarian's start-of-debate inversion is soft framing and tests baseline-equivalent.
• Sycophancy guardrail in advisor + peer prompts; new peer-review question separating genuine agreement from conformity (Peacemaker-or-Troublemaker 2026; CONSENSAGENT).
• Mediating Assessments in chairman synthesis — score 3–5 independent attributes before the holistic call, fighting coherence bias (Kahneman/Lovallo/Sibony 2019).
• Outside-view / base-rate check in the Executor + "How to verify" (reference-class forecasting; Kahneman/Tversky, Flyvbjerg).

• --jury

  — 3 diverse-model chairmen for close calls (jury-of-judges / PoLL).

claude -p

• Original concept: Andrej Karpathy (LLM Council)
• Claude Code adaptation: Ole Lehmann (@itsolelehmann)
• Reasoning method diversity: DMAD (ICLR 2025)
• Collaborative > adversarial empirical evidence: M3MADBench (arXiv 2601.02854, 2026)
• Confidence-modulated debate protocol: Demystifying MAD (arXiv 2601.19921, 2026)
• KS-statistic adaptive stopping: rachittshah/llmcouncil (S2 MAD implementation)
• Diversity-footprint verification approach: Counsel (Same model, same blind spots)
• (V2) Devil's advocate as the only reliable disagreement-inducer: Inducing Disagreement in Multi-Agent LLM Executive Teams (OpenReview 2026); LLM-Powered Devil's Advocate (IUI 2024)
• (V2) Sycophancy / premature consensus: Peacemaker or Troublemaker (2026); CONSENSAGENT (ACL 2025)
• (V2) Mediating Assessments Protocol: Kahneman, Lovallo & Sibony, A Structured Approach to Strategic Decisions (MIT SMR, 2019)
• (V2) Outside view / reference-class forecasting: Kahneman & Tversky; Flyvbjerg
• (V2) Jury-of-judges: LLM-as-Judge best practices 2026 (PoLL-style panels)
• V2 optimize pass + outcome research: the V2 outcome-research brief
• Skill by: Neal Meyer