red-team-review
GitHub提供红队对抗审查技能，支持战略矩阵与交付物审计两种模式。通过结构化流程识别逻辑漏洞、偏见及盲点，适用于论证、计划或代码等产物的批判性评估与安全加固。
Trigger Scenarios
Install
npx skills add winstonkoh87/Athena-Public --skill red-team-review -g -y
SKILL.md
Frontmatter
{
"name": "red-team-review",
"model": "default",
"description": "Unified adversarial review, two modes. Strategic Matrix v4.3 (MTA-004) for arguments\/plans\/decisions: Priors → Rubric → Adversarial Lenses → SWOT\/TOWS → MCDA → Blind Spot\/Kill Switch → Summary. Deliverable Audit for built artefacts (assignments, reports, code outputs): isolated fresh-context hunt against the brief, executed lenses, evidence-only findings, red-run fixes. Absorbs: bias-detector.",
"argument-hint": "review this | red team | what did I miss | QA | bias check | is this anchored | base rate",
"context_trigger": "review, red team, what did I miss, QA, bias check, critique, pre-mortem, is this ready, stress test, adversarial, deliverable audit, audit the deliverables, audit the assignment, pre-submission check",
"epistemic_status": "agent-discretion"
}
Red-Team Review (v4.3 — Strategic Matrix)
Canonical Protocol: MTA-004-red-team-v4-3.md Absorbs:
bias-detector(v4.0 Phase 2 anchoring/base-rate suite retained below) Upgraded: 2026-09-04 — unified skill v4.0 + protocol v4.3 into single version Version History: v4.0 (5-phase) → v4.1 → v4.2 → v4.3 (Strategic Matrix, 7-phase)
This skill wraps MTA-004 (the canonical red-team protocol) and adds trigger routing + the bias detection suite from v4.0. For the full framework, load MTA-004 directly.
Mode Routing (pick before Phase 0)
| Target | Mode |
|---|---|
| An argument, plan, strategy, price or claim to believe | Strategic Matrix: the 7 phases below |
| A built artefact someone will receive or mark (docx/PDF, slides, notebook, code output, handoff folder) | Deliverable Audit: load deliverable-audit.md. Skip SWOT/MCDA. Severity = marks or users at risk. |
Mixed targets (an essay's argument and its compiled file) get both: Strategic Matrix on the draft, Deliverable Audit on the compiled file. The Deliverable Audit core rule: cut the auditor off from the work's history (answer key, changelog, earlier audits, "verified" claims), never from the spec, and make it run things, not just read.
Triggers
"review this", "red team", "what am I missing", "is this ready to ship", "QA", "critique", "is this a fair price", "too good to be true", "scam", "is this legit", "grill me"
The 7-Phase Framework (MTA-004 v4.3)
Phase 0: Priors & Premises (Mandatory)
Assume the "Alien Observer" stance. Declare:
- Thesis: core argument/frame of the artifact
- Falsification: what specific metric/event proves this wrong
- Excluded Reality: what valid perspective is strictly ignored
- Reality Check: the unspoken cultural narrative or meta-game
Do not proceed until these 4 are answered.
Phase 1: The Objective Rubric
Score the artifact against this matrix. Start at 50%.
| Component | Weight | Criteria |
|---|---|---|
| Logic & Rigor | 30% | Deductively sound. Checks survivorship bias & sample size. No orphan stats. |
| Bias Mitigation | 20% | Steelmans opposition. Checks for sycophancy. |
| Completeness | 20% | Addresses "The Victim." Covers edge cases and cultural nuance. |
| Strategic Alignment | 15% | Aligns with Law #1. Respects physics. High utility. |
| Actionability | 15% | Recommendations are distinct, executable, and ranked. |
Confidence Cap: Low confidence → max score capped at 60/100.
Phase 1.5: The "Grill Me" Test (Echo Chamber Breaker)
If invoked via /grill or explicitly asked to "grill me":
- Reverse Roles: Act as the senior, uncompromising reviewer.
- Challenge Premises: Explicitly challenge at least one core premise.
- Demand Proof: Make the user defend their decisions.
Phase 2: Adversarial Lenses (The Attack)
One concrete objection per lens:
| Lens | The Attack Question |
|---|---|
| The Skeptic | Survivorship bias or small sample size (n=1)? Show me the graveyard. |
| The Victim | This strategy harms me because [negative externalities]. |
| The Historian | Mean reversion? Does this work in a bear market? |
| The Anthropologist | What is the implied social contract or cultural narrative? |
| The Cynic | You are lying to yourself about [hidden incentive/ego]. |
| The System | This ignores the 2nd order effect on [broader ecosystem]. |
| The Entropy | This will rot via [vector] because [reason]. |
Entropy Vectors: Incentive Rot, Complexity Rot, Competition Rot, Regulatory Rot, Tech Obsolescence, Data Rot.
Phase 2B: Bias Detection Suite (from v4.0)
Anchoring Detection
- Source Check: Where did this number come from?
- Red Flags: Round numbers, comparison anchoring, pre-concession framing.
- Structural Test: Cost of production? Value to buyer? Alternative cost?
- Correction:
Structural Value = max(Cost * 1.5, Value to Buyer * 0.3)
Base Rate Audit
IF (Claimed Outcome) >> (Expected Outcome for Demographics)
THEN (Hidden Variable) EXISTS
Hidden Variable Spectrum (descending probability):
- Hidden Capital — family money, inheritance
- Hidden Cost — extreme leverage, illegal activity
- Hidden Variance — survivorship bias
- Fabrication — the claim is a lie
- Genuine Outlier — the <0.1%
Standard Bias Checklist
Sycophancy, Cherry-Picking, False Precision, Assumed Context, Complexity Bias.
Phase 3: Strategic Expansion (SWOT + TOWS)
Use Phase 2 objections to populate Weaknesses/Threats. Build the TOWS matrix:
- SO (Attack): Use Strength to capture Opportunity
- WO (Improve): Fix Weakness to capture Opportunity
- ST (Defend): Use Strength to block Threat
- WT (Mitigate): Survival/exit strategy
Phase 4: Decision Engine (MCDA)
Rank top 3 recommendations using:
Final Score = ((Impact + Ease) / Risk Factor) * Entropy Multiplier
- Risk Factors: Low = 1.0 | Med = 1.5 | High = 2.0
- Entropy Multipliers: Stable = 1.0 | Decaying = 0.8 | Rotting = 0.5
Invoke Protocol 121 (MCDA/EEV/Pairwise) for complex multi-criteria rankings.
Phase 5: Blind Spot & Kill Switch
- "I am least confident about ______ because ______." (Check for survivorship bias.)
- The Kill Switch: We must ABANDON this strategy immediately if [specific condition].
Executive Summary
- Final Score: XX/100 (Confidence: X)
- The Kill Shot: The single most dangerous flaw found.
- The #1 Recommendation: The highest scoring Phase 4 option.
Rules
- Quote directly. No vague complaints.
- Steelman opposing views BEFORE critiquing them.
- Empty sections are allowed — don't invent issues.
- Every HIGH must have a fix achievable in ≤10 minutes.
Reference Protocols
- MTA-004: Red-Team Review v4.3 — canonical protocol (this skill wraps it)
- Deliverable Audit mode — isolation packet, lenses L1–L6, evidence format, red-run fixes, stop/oscillation rules
- WFL-247: Red-Team Handoff — cross-model handoff prompt
- DEC-121: MCDA / EEV / Pairwise — Phase 4 ranking engine
- DEC-500: GTO Problem Solver — capstone decision protocol
Version History
-
2352f57
Current 2026-10-18 19:04
新增Deliverable Audit（交付物审计）模式，支持对文档、代码等成品进行隔离式独立审查；强化Daily Self-RSI启动器。
-
006a53a
2026-09-08 20:31
v4.3版本引入战略矩阵，扩展为7阶段框架；整合bias-detector模块；更新触发路由机制。
- e624e2d 2026-07-19 08:47
