Claude Code Skill · v3.5

Adversarial Verify

Chain-of-Verification meets stress testing and tri-modal reasoning.
Catch what normal reviews miss. Know what's genuinely robust.

Install Now View on GitHub

How It Works

A structured pipeline with tri-modal reasoning and stress testing

Pre-verification (Steps 0–0b)

0

Identify & Gather

Detect the verification domain, collect artifacts, and establish ground truth from tests, specs, and production state.

Chain-of-Verification (Steps 1–2b)

1

Decompose & Classify

Break artifacts into verifiable claims. Classify each by reasoning mode: deductive, inductive, or abductive.

→
2

Adversarial Questions

Generate counter-questions, then abstract into recurring failure categories across the codebase.

Deep Verification (Steps 3–3d)

3

Independent Verify

Trace actual code paths. Apply hidden behavior probing and the modular adversarial scaffold.

→
3d

Stress Test

Existence Question, Scale Shift, Time Travel, Requirement Inversion. Produce Survived: yes/no verdicts.

Reporting (Steps 4–5)

4

Evidence-Based Report

Findings with reasoning-aware confidence, anti-fabrication discipline, and Survived verdicts. Hypotheses reported separately.

→
5

Project Discovery

Propose updates to TODO.md, SPEC.md, PLAN.md based on findings. Waits for user confirmation.

6 Verification Domains

Every artifact type gets domain-specific verification with appropriate ground truth

{ }

Code

Source changes, logic flows, edge cases, concurrency issues

Ground truth: tests, type system, spec
▵

Architecture

Design decisions, spec coverage, dependency risks, breaking changes

Ground truth: requirements, constraints, patterns
▦

Data

Schemas, migrations, contracts, backward compatibility

Ground truth: production schema, validation rules
✎

Documentation

Technical docs, process docs, user-facing content, install guides

Ground truth: actual codebase, current API, git history
✓

Tests

Test suite integrity, mock honesty, coverage accuracy, flaky detection

Ground truth: production code, requirements, coverage reports
⚙

Analysis

Agent outputs, performance reports, research summaries

Ground truth: source material, cited references

Core Techniques

Seven research-backed methodologies working together

Chain-of-Verification

Decompose artifacts into verifiable claims, generate adversarial counter-questions, then verify each claim independently against ground truth.

Dhuliawala et al., 2023

Abstractive Red-Teaming

Identify recurring failure categories across the codebase: frequency assumptions, implicit ordering, stale state, missing completeness, silent fallthrough, assumed environment.

Anthropic, 2026

Hidden Behavior Probing

Four probing strategies: indirect (trace execution), scaffolded (chain findings), cross-reference (claims vs reality), and absence (what's NOT there).

AuditBench — Anthropic, 2026

Adversarial Scaffold

Five modules: suspicion modeling, attack selection (blast radius × probability), plan synthesis, execution, and subtlety detection.

Anthropic, 2025

Stress Techniques

Four techniques with forced variety: Existence Question (should this exist?), Scale Shift (10x? zero?), Time Travel (6 months later?), Requirement Inversion (exact opposite?). Produces Survived: yes/no verdicts.

Principles of Chaos Engineering

Tri-Modal Reasoning

Each claim classified as deductive (verify against source), inductive (generalize from 3+ instances), or abductive (best explanation from observations). Abductive findings reported as hypotheses with alternatives, never as facts.

v3.5

Anti-Fabrication

Confidence tied to reasoning mode: 80–100 requires cited source, 60–79 requires 3+ instances, 40–59 requires hypothesis with alternatives. Can't score >79 without citing where you looked. Absence claims must state what was checked.

v3.5

What It Catches

Real bug categories discovered across production codebases

Code Issues

  • Silent data corruption
  • Logic flaws & off-by-one errors
  • Initialization order bugs
  • Concurrent modification
  • State leaks across boundaries
  • Boundary condition failures
  • Resource exhaustion paths

Architecture Issues

  • Spec drift from implementation
  • Missing constraints
  • Over-engineering
  • Dependency risk
  • Breaking changes

Data Issues

  • Schema inconsistency
  • Data loss risk in migrations
  • Constraint gaps
  • Backward compatibility breaks

Documentation Issues

  • Stale instructions
  • API drift
  • Missing docs for new features
  • Broken examples
  • Misleading error messages
  • Version mismatch
  • Orphaned references
  • UI copy drift

Test Issues

  • Tautological tests
  • Mock leakage
  • Coverage lies
  • Missing negative tests
  • Fragile assertions
  • Test-code drift
  • Flaky indicators

Analysis Issues

  • Hallucinated facts
  • Stale references
  • Logical leaps
  • One-sided evidence

Agent Meta-Verification

  • Sycophantic deference
  • Hidden agenda detection
  • Anchoring bias
  • Confabulated confidence
  • Premature convergence
  • Evidence cherry-picking

Why Use It

Research-Backed

Built on 6 research references from Anthropic, academia, and industry. Not heuristics — proven methodology.

Multi-Domain

Verifies code, architecture, data, documentation, tests, and analysis — not just syntax or style.

Anti-Fabrication

Reasoning-aware confidence scoring. Can't claim absence without citing where you looked. Abductive findings are hypotheses, not facts.

Pattern Detection

Goes beyond individual bugs to find systemic failure categories that recur across your codebase.

Trust Integration

Built-in agent trust scoring: tracks confirmed bugs, false positives, and clean reviews for multi-agent workflows.

Battle-Tested

Extracted from real game development. Caught critical bugs including timer resets, patrol flip-flops, and missing collision bounds.

Reference Papers

The research foundation behind every verification step

CoV

Chain-of-Verification Reduces Hallucination in Large Language Models

Dhuliawala et al., 2023 — The core methodology: decompose, question, verify independently.

arxiv.org/abs/2309.11495 ↗ Auditing

Automated Auditing

Anthropic, 2025 — Systematic approaches to automated code and behavior auditing.

alignment.anthropic.com ↗ Red-Teaming

Abstractive Red-Teaming

Anthropic, 2026 — Finding failure categories, not just individual issues.

alignment.anthropic.com ↗ Probing

AuditBench

Anthropic, 2026 — Benchmarks for hidden behavior probing and agent meta-verification.

alignment.anthropic.com ↗ Scaffold

Strengthening Red Teams

Anthropic, 2025 — Modular adversarial scaffolding for systematic security testing.

alignment.anthropic.com ↗ Stress

Principles of Chaos Engineering

Hypothesis-driven resilience testing adapted for code review: formulate steady-state, vary conditions, observe, conclude.

principlesofchaos.org ↗

Installation

Option 1 — Clone & Copy

git clone https://github.com/fullo/claude-adversarial-skill.git
cp -r claude-adversarial-skill/skills/adversarial-verify ~/.claude/skills/

Option 2 — Claude Marketplace

claude marketplace add fullo-plugins https://github.com/fullo/claude-plugins-marketplace

claude plugin install adversarial-verify@fullo-plugins

Compatibility

Claude Code Cursor Windsurf Cline

Works with any tool that supports the Agent Skills format. Requires git for diff analysis.

Usage

Trigger the skill with natural language or the slash command

/adversarial-verify
run an adversarial review on my recent changes
CoV check the last commit
verify this code with total skepticism
verify the PLAN.md against the SPEC.md
adversarial check on this migration
verify this agent's analysis report
look for systemic failure patterns in the codebase
probe this function for hidden behaviors
verify the README matches the actual install process
stress test the auth module
what happens at 10x scale?
check if the planning agent's output is biased