Adversarial bug finding skill for Claude Code. Uses 3 isolated AI agents to find and verify real bugs with high fidelity.
Inspired by @systematicls's article on exploiting LLM sycophancy for better code review:
- Hunter - Scans your code and reports every possible bug (biased to over-report)
- Skeptic - Tries to disprove each bug (biased to dismiss false positives)
- Referee - Reads the code independently and makes final verdicts
Each agent runs in a completely isolated context — they can't see each other's reasoning, only structured findings. This prevents anchoring bias and produces high-fidelity results.
git clone https://github.com/danpeg/bug-hunt.git ~/.claude/skills/bug-huntClaude Code auto-discovers skills in ~/.claude/skills/.
/bug-hunt # Scan entire project
/bug-hunt src/ # Scan specific directory
/bug-hunt lib/auth.ts # Scan specific file
/bug-hunt -b feature-xyz # Scan files changed in feature-xyz vs main
/bug-hunt -b feature-xyz --base dev # Scan files changed in feature-xyz vs dev
Branch diff mode (-b) scans only files changed in a branch compared to a base branch (defaults to main). It reads the full file contents — not just the diff — so bug detection quality is preserved.
cd ~/.claude/skills/bug-hunt && git pullrm -rf ~/.claude/skills/bug-huntThe scoring incentives are load-bearing — they exploit each agent's desire to maximize its score:
- Hunter: +1/+5/+10 for low/medium/critical bugs. Motivates thoroughness.
- Skeptic: Earns points for disproving false positives, but pays 2x penalty for wrongly dismissing real bugs. Creates calibrated caution.
- Referee: Symmetric +1/-1 scoring with "ground truth" framing. Makes it precise rather than biased.
Based on the adversarial bug hunting technique described by @systematicls in "How To Be A World-Class Agentic Engineer."
Brand by Kitt at Curious Endeavor
Contributions welcome! See CONTRIBUTING.md for guidelines.
MIT