メニュー
Use when reviewing code for bugs with finder, challenger, and referee roles, especially when independent review is needed to reduce false positives.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Optimize an agent skill through bounded repeated trials, binary evaluations, targeted prompt mutations, and auditable keep/discard decisions. Use when asked to optimize, benchmark, evaluate, or systematically improve an existing skill. Produces an improved candidate skill, run artifacts, a results log, and a mutation changelog without overwriting the original.
Review uncommitted git changes for bugs/regressions via Codex MCP and present a structured report. Use when asked to review local changes or find bugs in current work. Requires Codex MCP to be configured.
Organize raw project documentation into a focused, cross-linked docs/ structure with an adaptive INDEX.md and a recommended AGENTS.md entry point. Preserve the source's natural document types; create stories only when they already exist and incremental slices only when the user wants them. Use when starting from raw documentation or restructuring scattered project docs.
Record one web application feature flow as a compact GIF using an existing Playwright installation. Use when the user asks for a visual demo, changelog artifact, documentation asset, or pull-request evidence. Creates local capture artifacts first; installation, publication, commits, uploads, and pull-request edits always require separate approval.
Coordinate specialized agent roles across complex tasks using dependency-aware planning, explicit ownership, structured handoffs, safe integration, and evidence-based quality gates. Use when work has multiple independent streams or genuinely benefits from isolated expertise; fall back transparently when the runtime cannot provide separate workers.
Draft and, after approval, capture an observed reusable lesson in an existing skill's gotchas.md. Use after a verified bug, architecture error, process failure, or misunderstanding when the lesson can be generalized without exposing project or user information.
| name | adversarial-bug-hunt |
| description | Use when reviewing code for bugs with finder, challenger, and referee roles, especially when independent review is needed to reduce false positives. |
| version | 1.1.0 |
| level | advanced |
| category | code-quality |
Three-role bug discovery pipeline. A bug-finder over-reports, an adversarial reviewer challenges, and a referee resolves. Prefer clean, isolated contexts; when the runtime cannot provide them, use the documented reduced-assurance fallback instead of inventing platform-specific tools.
gotchas.md before startingChoose the first available mode. Do not assume tool names or ask the user to install another agent runtime.
| Mode | How to run the roles | Assurance |
|---|---|---|
| Isolated workers | Launch each role with the runtime's native delegation mechanism in a separate clean context | Preferred |
| External isolated sessions | Run each role in a separate available process or session that can read the review scope | Preferred |
| Sequential fallback | Run the three prompt templates as separate passes in the current context, preserving each JSON handoff | Reduced |
For sequential fallback, tell the user before Phase 1:
This runtime cannot isolate the three reviewers. I can run the same finder → adversarial → referee pipeline sequentially, but the roles may influence one another, so the result has reduced assurance.
Record the selected mode in the final report. Never describe sequential fallback as independent or clean-context review.
If the user provided a scope (files, directory, "uncommitted changes"), use it. Otherwise ask. Default: uncommitted changes.
Read references/agent-prompts.md for the bug-finder prompt template.
Run the bug-finder prompt with the selected execution mode, passing the scope. In an isolated mode, start a fresh worker or session. In sequential fallback, run it as the first pass. The role is incentivized to over-report and produces the superset of possible bugs.
Receives: JSON array of findings using the schema in templates/finding-schema.json.
Read references/agent-prompts.md for the adversarial prompt template.
Run the adversarial prompt with the selected execution mode, passing the bug-finder's full JSON output. Use a fresh context in either isolated mode; otherwise run it as the second sequential pass. The role is incentivized to disprove and produces the subset of likely real bugs.
Receives: same findings array, each annotated with adversarial_verdict and adversarial_reasoning.
Read references/agent-prompts.md for the referee prompt template.
Run the referee prompt with the selected execution mode, passing both prior outputs. Use a fresh context in either isolated mode; otherwise run it as the third sequential pass. The referee resolves each dispute.
Receives: same findings array, each annotated with referee_verdict and referee_reasoning.
Read references/report-format.md for display format. See examples/ for concrete samples.
Show only CONFIRMED + UNCERTAIN findings. If all findings are DISPROVED: say "No confirmed issues found in the reviewed code." and stop.
Otherwise, ask the user which finding IDs to fix:
"Which issues would you like me to fix? You can list IDs (e.g., BUG-001, BUG-003) or say 'all'."
When the user replies, prepare a fix plan before editing. Use the runtime's native planning mode when one exists; otherwise present the plan directly in chat and wait for approval. The plan must reference the selected findings, evidence, and fix recommendations from the report.