Evil Twin

Your shadow self. Attacks your reasoning. Solves blind. Enters the forge.

Works standalone with just Claude. Connects into AI Buddies for full multi-AI competitive builds.

What's inside

Command	What it does	Speed
/evil-twin	Adversarial self-challenge — assumes you're wrong, finds the error	30-60s
/evil-twin --quick	Fast inline 3-point sanity check	5-10s
/doppelganger	Independent blind solve in git worktree — merges best of both	2-5 min
/evil-pipeline	Full chain: Twin → brainstorm → forge with Doppelganger	5-15 min

Install

claude plugin marketplace add cukas/evil-twin
claude plugin install evil-twin@cukas

If you have AI Buddies installed, Doppelganger auto-registers as a forge competitor on session start.

How it works

Evil Twin — "Am I wrong?"

Auto-triggers when Claude states confidence below 92%. A second Claude with an adversarial system prompt attacks your reasoning using three frameworks:

• Inversion — "You said X works because Y. What if Y is false?"
• Pre-mortem — "This shipped and broke production. What went wrong?"
• Second-order — "What does this break downstream?"

Up to 5 concrete failure scenarios. You rebuttal each one. Resolution table with updated confidence. If still below 85% → escalates to /tribunal or /brainstorm.

Doppelganger — "Is there a better way?"

A second Claude solves the same problem in a git worktree WITHOUT seeing your solution. Then both approaches are compared. Merges the best parts from both — doesn't just pick a winner.

Includes programmatic leakage detection to verify independence.

Evil Pipeline — "Go full send"

Chains everything together when stakes are high:

/evil-pipeline "implement auth middleware"

  1. Evil Twin challenges your approach          (30-60s)
     |
  2. Brainstorm/tribunal if still uncertain      (3-5 min)
     |
  3. Forge: Doppelganger + Codex + Gemini        (3-10 min)
     |    Doppelganger solves BLIND (no context from steps 1-2)
     |    Codex and Gemini solve independently
     |
  4. Convergence analysis + merge best parts

The key insight: in regular /forge, Claude has seen the entire conversation (unfair anchoring). Doppelganger replaces Claude's slot with a genuinely blind solve — like Codex and Gemini, it starts fresh.

Requires AI Buddies plugin.

When to use what

Situation	Tool
"Am I wrong about this?"	/evil-twin
Quick sanity check	/evil-twin --quick
"Is there a better way?"	/doppelganger
High-stakes, need full verification	/evil-pipeline
External perspectives needed	/tribunal (AI Buddies)
Who should take this task?	/brainstorm (AI Buddies)

Configuration

Config file: ~/.evil-twin/config.json

{
  "threshold": 0.92,
  "max_challenges": 5,
  "auto_trigger": true,
  "escalation_threshold": 0.85,
  "quick_default": false,
  "easter_eggs": false,
  "doppelganger_timeout": 300,
  "doppelganger_run_tests": true,
  "doppelganger_run_build": true,
  "pipeline_stop_on_flawed": true,
  "pipeline_timeout": 600
}

Architecture

evil-twin/
├── skills/
│   ├── evil-twin/SKILL.md        # Adversarial challenge protocol
│   ├── doppelganger/SKILL.md     # Blind verification protocol
│   └── evil-pipeline/SKILL.md    # Full pipeline orchestration
├── scripts/
│   ├── doppelganger-cli.sh       # CLI adapter for AI Buddies forge
│   ├── detect-project.sh         # Auto-detect language, build, test
│   └── leakage-check.sh          # Contamination detection
├── hooks/
│   ├── session-start.sh          # Confidence injection + buddy registration
│   └── lib.sh                    # Config helpers
└── commands/
    ├── evil-twin-help.md
    └── evil-twin-config.md

Standalone: Works with just Claude. No external AI keys needed. Connected: If AI Buddies is installed, Doppelganger auto-registers as a forge competitor. /evil-pipeline chains everything together.

License

MIT