Devils Advocate

Devil's Advocate (s35)

Purpose

Systematically dismantle ideas, strategies, designs, and hypotheses to expose every flaw, assumption, contradiction, and logical fallacy — so only the strongest arguments survive. This skill operates as a quality gate across the entire workflow, ensuring that no critical decision goes unchallenged.

Adapted from the Devil's Advocate Agent (github.com/dungnotnull/devils-advocate-agent), which provides:

• Multi-intensity adversarial critique (Skeptic → Critic → Prosecutor → Demolisher)
• Real-time logical fallacy detection (14+ fallacy types)
• Multi-dimensional argument strength scoring (0-100 across clarity, evidence, logic, novelty, defense)
• RAG-grounded counter-arguments using external evidence
• Assumption hunting and implicit premise extraction

This skill integrates those capabilities into the workflow as a callable quality gate.

---

Input Contract

| Input | Source | Required | |---|---|---| | Subject to critique | Any skill output, PRD, design, hypothesis, strategy | Yes | | Critique intensity level (1-4) | User specification or auto-determined | No (default: 2) | | Domain context | s01 PRD, s02 taste | Yes | | Relevant evidence/research | s32 deep-research (if available) | No | | Current phase context | s03 progress tracker | Yes |

---

Output Contract

| Output | Destination | Format | |---|---|---| | Counter-argument report | .commandcode/artifacts/critique/ | Markdown | | Argument strength scores | .commandcode/artifacts/critique/ | JSON | | Identified fallacies | .commandcode/artifacts/critique/ | JSON | | Assumption inventory | .commandcode/artifacts/critique/ | Markdown | | Recommended revisions | Target skill's output (feedback loop) | Markdown | | Quality gate verdict | s03 progress tracker | JSON |

---

Prerequisites

• [ ] Subject matter (idea, design, hypothesis, strategy) clearly defined
• [ ] Domain context available (PRD, taste, or relevant background)
• [ ] User or orchestrator has explicitly invoked critique

---

Intensity Levels (Adapted from Devil's Advocate Agent)

| Level | Name | Behavior | When to Use | |---|---|---|---| | 1 | Skeptic | Gentle probing questions; assumes good faith | Early ideation, brainstorming sessions | | 2 | Critic | Direct challenges with evidence; names gaps explicitly | Design reviews, hypothesis validation, PRD reviews | | 3 | Prosecutor | Systematic deconstruction of all premises; counts contradictions | Architecture decisions, security reviews, production readiness | | 4 | Demolisher | Maximum adversarial intensity; tears apart piece by piece | Pre-deployment gates, game day pre-mortems, compliance audits |

Default: Level 2 (Critic) Auto-escalation: Increase by one level when:

• The same weak argument is repeated without new evidence
• Score drops ≥5 points between consecutive critique rounds
• Critical assumption is left unaddressed after one challenge

---

Critique Framework

Phase 1: Input Analysis

`yaml inputanalysis: claimextraction:

• Identify every factual claim in the subject
• Rate each claim: verifiable / unverifiable / opinion
• Flag statistical claims without cited sources

assumption_hunting:

• Identify explicit assumptions (stated "we assume...")
• Identify implicit assumptions (unstated but necessary for the argument)
• Categorize: causal / universal / value / existence / comparison

fallacy_detection: types:

• ad_hominem: Attacking the person, not the argument
• straw_man: Misrepresenting the opposing position
• hasty_generalization: Drawing broad conclusions from limited data
• false_dichotomy: Presenting only two options when more exist
• appealtoauthority: Citing authority without domain expertise
• slippery_slope: Assuming one step inevitably leads to extreme outcomes
• circular_reasoning: The conclusion is assumed in the premise
• bandwagon: "Everyone is doing it" as sole justification
• anecdotal_evidence: Generalizing from personal stories
• confirmation_bias: Selectively citing supporting evidence only
• sunk_cost: Continuing because of prior investment
• survivorship_bias: Only considering successful cases
• post_hoc: Assuming correlation implies causation
• appealtotradition: "This is how we've always done it"

`

Phase 2: Multi-Perspective Challenge

Every subject is challenged through four lenses (adapted from s01-1 user-flow review):

`yaml perspectives: business_lens: questions:

• "Does this actually solve the stated business problem?"
• "What is the ROI? Can you quantify it?"
• "What happens if adoption is 10% of projections?"
• "What are the regulatory/compliance risks?"
• "How does this impact revenue, costs, or customer churn?"

engineering_lens: questions:

• "Is this implementable with the current architecture?"
• "What are the hidden complexity costs?"
• "How does this scale to 10x current load?"
• "What are the failure modes?"
• "What are the blast radius implications?"

reliability_lens: questions:

• "What happens when this fails?"
• "How do we detect failure? (Observability)"
• "What is the recovery time objective?"
• "Has this been tested under chaos conditions?"
• "What are the cascading failure risks?"

security_lens: questions:

• "What is the attack surface of this change?"
• "Are there privilege escalation paths?"
• "Does this introduce new secrets or credentials?"
• "Does this comply with security policies (s24)?"
• "What is the worst-case security scenario?"

`

Phase 3: Argument Strength Scoring

Each argument is scored on a 0-100 scale across five dimensions:

json { "argument_strength": { "clarity": { "score": 0, "description": "How clearly is the claim stated?", "rubric": { "0-20": "Vague, ambiguous, or incomprehensible", "21-40": "Understandable but imprecise", "41-60": "Clear but could be more specific", "61-80": "Well-articulated with precise language", "81-100": "Crystal clear with measurable criteria" } }, "evidence_quality": { "score": 0, "description": "Quality and relevance of supporting evidence", "rubric": { "0-20": "No evidence or purely anecdotal", "21-40": "Weak evidence, cherry-picked data", "41-60": "Some evidence but gaps in sourcing", "61-80": "Well-sourced evidence from credible sources", "81-100": "Comprehensive, peer-reviewed, and quantified evidence" } }, "logical_consistency": { "score": 0, "description": "Freedom from contradictions and logical gaps", "rubric": { "0-20": "Major logical contradictions", "21-40": "Some contradictions or unsupported leaps", "41-60": "Mostly consistent with minor gaps", "61-80": "Consistent with well-reasoned connections", "81-100": "Ironclad logic with no gaps" } }, "novelty": { "score": 0, "description": "Originality and depth of reasoning", "rubric": { "0-20": "Repetitive of known approaches", "21-40": "Minor variations on existing ideas", "41-60": "Some original insight combined with known approaches", "61-80": "Novel approach with clear advantages", "81-100": "Breakthrough insight that shifts the paradigm" } }, "defense_improvement": { "score": 0, "description": "How well weaknesses were addressed in follow-up rounds", "rubric": { "0-20": "No defense offered or defense made it worse", "21-40": "Weak defense, ignored major counter-arguments", "41-60": "Partial defense, addressed some challenges", "61-80": "Strong defense, addressed most challenges effectively", "81-100": "Exceptional defense that strengthened the original argument" } } }, "overall_score": "weighted average: (clarity*0.2 + evidence*0.25 + logic*0.25 + novelty*0.15 + defense*0.15)" }

Phase 4: Verdict Generation

yaml verdict: overall_assessment: one of [STRONG, MODERATE, WEAK, CRITICAL] confidence_level: percentage (0-100%) key_weaknesses: [list of top 3 weaknesses] key_strengths: [list of top 3 strengths] recommended_revisions: [list of specific, actionable revisions] escalation_needed: boolean (true if any dimension scores below 40) gate_result: PASS | CONDITIONAL_PASS | FAIL PASS: overall >= 70, no dimension below 50 CONDITIONAL_PASS: overall >= 50, no dimension below 40 FAIL: overall < 50 OR any dimension below 40

---

Workflow Integration Points

The Devil's Advocate integrates with the workflow at critical decision gates:

Integration with s14 (Experiment Design)

yaml trigger: After chaos experiment is designed, before blast radius approval action: Challenge the hypothesis, question fault selection, test abort conditions output: Counter-arguments fed back to s14 for revision, or gate PASSED

Integration with s15 (Hypothesis Validation)

yaml trigger: After hypothesis is written, before acceptance criteria are locked action: Attack the hypothesis from the reliability lens — what if it's wrong? output: Strengthened hypothesis with better acceptance/rejection criteria

Integration with s28 (Release Management)

yaml trigger: Before Go/No-Go decision action: Full Prosecutor-level (3) critique of release readiness output: Release gate verdict — PASS, CONDITIONAL, or FAIL

Integration with s31 (Strategic Creator)

yaml trigger: After strategic proposals are generated action: Challenge every proposal with Demolisher-level (4) intensity output: Only proposals that survive are forwarded to user for acceptance

Integration with s01 (BA Requirements)

yaml trigger: After PRD is drafted, before ADRs are locked action: Critic-level (2) challenge of business assumptions, technical feasibility output: Revised PRD with addressed counter-arguments documented

---

Critique Execution Protocol

Step 1: RECEIVE subject from invoking skill or user Step 2: LOAD domain context (PRD, taste, relevant research) Step 3: DETERMINE intensity level (default: 2, auto-escalate if needed) Step 4: EXTRACT claims, assumptions, and logical structure Step 5: DETECT fallacies (apply fallacy taxonomy) Step 6: GENERATE counter-arguments from each perspective lens Step 7: SCORE argument strength on 5 dimensions (0-100 each) Step 8: PRODUCE verdict (STRONG/MODERATE/WEAK/CRITICAL) Step 9: WRITE critique report to .commandcode/artifacts/critique/ Step 10: UPDATE progress.json with gate result Step 11: RETURN verdict + recommended revisions to invoking skill Step 12: IF gate FAIL → block progression, require revision IF gate CONDITIONAL → allow progression with documented risks IF gate PASS → proceed to next skill

---

Counter-Argument Templates

For Design Decisions

`

Counter-Argument: [DESIGN DECISION]

Claim: [What the design claims to achieve] Challenge: [Why it might not achieve this]

Evidence Against

1. [Evidence point 1]
2. [Evidence point 2]
3. [Evidence point 3]

Assumptions at Risk

• [Assumption 1]: [Why it might not hold]
• [Assumption 2]: [Why it might not hold]

Alternative Approaches

1. [Alternative 1]: [Trade-offs]
2. [Alternative 2]: [Trade-offs]

Score Impact

• Clarity: -[N] points ([reason])
• Evidence: -[N] points ([reason])
• Logic: -[N] points ([reason])

`

For Hypotheses (Chaos Engineering)

`

Counter-Argument: [HYPOTHESIS]

Hypothesis: [If X fails, then Y should happen, measured by Z] Challenge: [Why the hypothesis might be wrong]

What Could Go Wrong

1. [Failure mode 1]: [Description]
2. [Failure mode 2]: [Description]
3. [Failure mode 3]: [Description]

Missing Considerations

• [Observability gap]: [What metrics are missing?]
• [Blast radius risk]: [What else could be affected?]
• [Cascading failure]: [What if X failure triggers Y?]

Recommended Strengthening

1. [Add metric/condition]
2. [Add probe]
3. [Add abort condition]

`

For Release Decisions

`

Counter-Argument: [RELEASE GATE]

Decision: [Go/No-Go for release] Challenge: [Why this release might not be ready]

Open Risks

1. [Risk 1]: [Severity] - [Mitigation status]
2. [Risk 2]: [Severity] - [Mitigation status]
3. [Risk 3]: [Severity] - [Mitigation status]

Outstanding Items

• [ ] [Checklist item 1]
• [ ] [Checklist item 2]

Verdict: CONDITIONAL_PASS

Conditions: [List of conditions that must be met before Go] Escalation: [Who needs to approve if conditions not met] `

---

AI Agent Integration

Autonomy Level

| Aspect | Level | Description | |---|---|---| | Current | L1 | AI generates counter-arguments and scores; human decides on verdict | | Target | L2 | AI auto-critiques at designated gates; human reviews verdicts |

Harness AI Agent

Agent: None (adversarial by design — intentionally independent) Fallback: Uses the skill's built-in critique framework and fallacy taxonomy

Human Gates

• Final verdict acceptance (PASS/CONDITIONAL/FAIL)
• Intensity level selection for each critique session
• Escalation decisions when gate fails

MCP Integration

None required. This skill operates as a pure reasoning quality gate.

Integration with Devil's Advocate Agent

For enhanced critique, this skill can optionally integrate with the Devil's Advocate Agent (github.com/dungnotnull/devils-advocate-agent):

`yaml integration: mode: optional # Skill works standalone; agent enhances it benefits:

• ML-based fallacy detection (beyond rule-based taxonomy)
• RAG-grounded counter-arguments from research papers
• Argument strength scoring via cross-encoder NLI model
• Persistent debate sessions for iterative improvement

setup:

• Clone devil-advocate-agent repository
• Configure LLM provider (Anthropic/OpenAI/Ollama)
• Start agent: python main.py --port 8001
• Set environment: DEVILSADVOCATEURL=http://localhost:8001

usage:

• Skill sends subject to /api/v1/debate/start
• Receives counter-arguments with fallacy annotations
• Incorporates into critique report
• User can continue debate rounds for iterative strengthening

`

---

Success Criteria

• [ ] Every claim in the subject is extracted and categorized
• [ ] At least 3 counter-arguments generated per perspective lens
• [ ] All detected fallacies are named and explained
• [ ] Argument strength scored on all 5 dimensions
• [ ] Verdict includes specific, actionable recommended revisions
• [ ] Gate result (PASS/CONDITIONAL/FAIL) is unambiguous
• [ ] Critique report stored in .commandcode/artifacts/critique/
• [ ] Progress tracker updated with gate result
• [ ] If gate FAIL, progression is blocked until revision
• [ ] If gate CONDITIONAL, documented risks are recorded

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: dungnotnull
• Source: dungnotnull/hybrid-harness-chaos-process-prm
• License: MIT

Install and usage instructions live in the source repository linked above.