Adversarial Analysis Skill

Adversarial Analysis

RULE: Before implementing any significant proposal, systematically attack it to find weaknesses.

The Principle

Proposals that survive adversarial scrutiny are more robust. This skill provides a 6-phase framework for stress-testing ideas before commitment.

Loading This Skill

Agents can load via:

required_skills:
  - adversarial-analysis

Or dynamically invoke:

skill: adversarial-analysis

The 6-Phase Framework

Phase 1: Pre-mortem

"It's 6 months from now. This failed. Why?"

Assume complete failure and work backwards. What went wrong?

Process:

Vividly imagine the failure state
Identify 5-7 distinct failure scenarios
Trace each back to root causes
Note which causes were preventable

Output format:

### Pre-mortem
> "It's 6 months from now. This failed. Why?"
- [Failure scenario 1]: [Root cause]
- [Failure scenario 2]: [Root cause]
- [Failure scenario 3]: [Root cause]

Phase 2: Assumption Audit

Every proposal rests on assumptions. Surface and stress-test them.

Process:

List every assumption (explicit and implicit)
Rate confidence: H (high) / M (medium) / L (low)
For each, answer: "If wrong, what breaks?"

Output format:

### Assumptions (confidence: H/M/L)
| Assumption | Confidence | If Wrong |
|------------|------------|----------|
| [Assumption 1] | M | [Impact if false] |
| [Assumption 2] | L | [Impact if false] |
| [Assumption 3] | H | [Impact if false] |

Prioritize low-confidence, high-impact assumptions for deeper analysis.

Phase 3: Edge Case Storm

Generate 10+ scenarios that could break the proposal.

Process:

Think about boundary conditions
Consider rare but possible inputs
Imagine hostile actors
Consider scale extremes (0, 1, many, millions)
Consider timing/ordering issues
Consider resource exhaustion
Consider integration failures

Output format:

### Edge Cases
- [Breaking scenario 1]
- [Breaking scenario 2]
- [Breaking scenario 3]
... (minimum 10)

Phase 4: Counter-Arguments

"A senior engineer thinks this is wrong. Their case:"

Steel-man the opposition. Build the strongest possible case against.

Process:

Adopt an adversarial persona
Find the most compelling reasons to reject
Identify precedents where similar approaches failed
Note what a skeptic would demand as proof

Output format:

### Counter-Arguments
> "A senior engineer thinks this is wrong. Their case:"
- [Argument 1]: [Evidence/reasoning]
- [Argument 2]: [Evidence/reasoning]
- [Argument 3]: [Evidence/reasoning]

Phase 5: Failure Mode Catalog

Component-by-component failure analysis.

Process:

Identify each major component
Determine how each could fail
Assess blast radius (contained vs cascading)
Identify detection mechanisms

Output format:

### Failure Modes
| Component | How It Fails | Blast Radius | Detection |
|-----------|--------------|--------------|-----------|
| [Component 1] | [Failure mode] | [Impact scope] | [How to detect] |
| [Component 2] | [Failure mode] | [Impact scope] | [How to detect] |

Phase 6: Synthesis

Integrate findings into actionable recommendations.

Process:

Identify critical weaknesses (must address before proceeding)
Identify significant concerns (should address)
Render verdict: GO / CONDITIONAL GO / NO GO
State confidence level

Output format:

### Synthesis

**Critical Weaknesses (must address):**
1. [Weakness] - Mitigation: [Action]
2. [Weakness] - Mitigation: [Action]

**Significant Concerns (should address):**
1. [Concern] - Mitigation: [Action]

**Verdict:** [GO / CONDITIONAL GO / NO GO]
**Confidence:** [High / Medium / Low]
**Rationale:** [1-2 sentences explaining verdict]

Full Output Template

# Challenge: [Proposal Summary]

## The Attack
### Causal Failure Map: [root causes and cascade effects]
### Assumption Audit: [weakest assumptions]
### Counter-Arguments: [strongest case against]

## What the Analysis Caught

| Weakness | Source | Severity | Mitigation |
|----------|--------|----------|------------|
| [weakness] | [phase] | Critical/High/Medium | [action] |

## Verdict
**[GO / CONDITIONAL GO / NO GO]** (confidence: X.X)
[2-4 sentences: real risk picture, what would change the verdict]

**Required Mitigations (if CONDITIONAL GO):**
1. [mitigation]

## Where to Go Next
-> /challenge --deep "[proposal]"
   _[if verdict needs simulation/ethical analysis]_
-> /problem-solve "[how to mitigate top risk]"
   _[if proceeding and need to decide on approach]_
-> /deepthink "[specific uncertainty]"
   _[if a key assumption needs exploration]_

Using cognition-mcp

For rigorous analysis, use cognition-mcp to store structured analysis:

mcp__cognition-mcp__cognition
  operation: "thought"
  thought: "Beginning adversarial analysis of [proposal]. Phase 1: Pre-mortem..."
  thoughtNumber: 1
  totalThoughts: 12
  nextThoughtNeeded: true

Structure as 2 thoughts per phase (analysis + output), allowing revision and branching as needed.

Quick Mode (Phase 1 + Top 3 Risks)

When time is limited, run Phase 1 only and extract top 3 risks:

# Challenge (Quick): [Proposal Summary]

## Causal Map
[cause] -> [intermediate] -> [effect]
Root causes: [2-3 most critical]

## What the Analysis Found

| Risk | Likelihood | Impact | Mitigation |
|------|-----------|--------|------------|
| [risk] | H/M/L | [impact] | [mitigation] |

## Quick Verdict
[Proceed with caution / Needs more analysis / Red flags present]
[1-2 sentences: why]

## Where to Go Next
-> /challenge "[proposal]" (full analysis)
   _[why deeper analysis would help]_
-> /problem-solve "[decision point]"
   _[if ready to decide]_

Deep Mode (+ Cognition Integration)

For critical decisions, after completing all 6 phases, invoke:

/think --model inversion [proposal]

This adds mental model analysis (thinking backwards from failure) to complement the adversarial framework.

Anti-Patterns

DON'T:

Rush through phases
Accept vague "it might fail" without specifics
Skip the steel-man counter-arguments
Ignore low-probability, high-impact failures
Let optimism bias filter out uncomfortable findings

DO:

Be genuinely adversarial
Quantify where possible
Name specific failure modes
Provide actionable mitigations
Render an honest verdict

Integration with Other Skills

search-before-edit: Search for similar past failures before analysis
debugging-first: If analyzing a fix, debug the original issue first
linter-loop-limits: Don't iterate forever on edge case generation

When to Use

Use adversarial analysis when:

Making architecture decisions
Choosing between major alternatives
Proposing significant refactors
Introducing new dependencies
Planning migrations
Making irreversible choices

Skip (or use quick mode) when:

Changes are easily reversible
Scope is small and contained
Time pressure is extreme (but document the skip)

Agent Skills: Adversarial Analysis

Install this agent skill to your local

Skill Files