メニュー
Run a skeptical Codex review from Claude Code that hunts for hidden regressions, unsafe assumptions, and release risk. Use before merges, deploys, migrations, or other risky changes.
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
ThumbGate provides pre-action gates for AI coding agents. It captures thumbs-up/down feedback on agent actions, auto-promotes repeated failures into prevention rules, and blocks known-bad tool calls via PreToolUse hooks. Trigger when the user wants to add safety guardrails to an AI agent workflow, capture structured feedback on agent output, generate prevention rules from failure patterns, gate high-risk actions before execution, or export DPO training pairs from production feedback. Works with any MCP-compatible agent including Cursor, Codex, Gemini CLI, Amp, and OpenCode.
Make any ThumbGate-facing copy — landing/pricing/compare/guide pages, README, postinstall banners, CLI receipts, Reddit/LinkedIn/Bluesky/Threads posts, docs, outreach drafts, blog/launch content — sound like ThumbGate: direct, technical, honest, anti-hype. Use BEFORE writing or editing customer-facing or marketing text, and when the user says "write a landing page", "draft a Reddit/LinkedIn post", "write the README", "make this on-brand", "punch up this copy", or "review this copy for voice". Do NOT use for code comments, commit messages, internal notes, or non-ThumbGate copy — and never to add fabricated traction or unsupported claims (honesty is the brand).
Capture thumbs up/down feedback into structured memories and prevention rules. Require one sentence of why before claiming memory promotion. Use when user gives explicit quality signals about agent work (e.g. "that worked", "that failed", "thumbs up/down"). Do NOT use for general questions, code generation, file operations, or any task that is not explicit feedback on prior agent output.
Give your agent persistent cross-session memory — recall what went wrong on similar tasks before starting, then capture whether the work succeeded or failed so repeated mistakes auto-promote into prevention rules. Runs locally via the ThumbGate MCP server (recall, capture_feedback, prevention_rules, feedback_stats) — no API key needed. Use when the user says "recall", "what went wrong last time", "remember this for next time", "capture feedback", "thumbs up/down on that", or at the start of a new task or session. Do NOT use when the user wants to force one specific mistake into an enforced block gate (use the thumbgate-guard skill), to list rules that already exist (thumbgate-rules), or to check ThumbGate wiring (thumbgate-doctor).
Course-correct from over-planning to execution — when the agent has produced a plan, research, or options instead of doing work the user already approved, drop the analysis and execute the smallest end-to-end action now (branch, run, deploy, verify). Use when the user says "stop planning, just do it", "why are you researching, execute", "just ship it", "enough analysis", or pushes back that you keep proposing instead of acting on an already-decided task. Do NOT use when the task is genuinely undecided and needs a plan first, when the user explicitly asked for research/options/a design, or for routine work where no over-planning has occurred — acting on an unscoped or destructive task is its own failure.
Show ThumbGate's enforcement record — how many risky actions were actually blocked versus warned, which gates fire most, the tokens/damage saved, and the full feedback to check to rejection pipeline. Reads live enforcement counters via the gate_stats and enforcement_matrix MCP tools (CLI fallback `npx thumbgate gate-stats`). Use when the user says "what has ThumbGate blocked", "show gate stats", "is enforcement working", "how many tokens did we save", "show the enforcement matrix", or "what got stopped". Do NOT use to list rule DEFINITIONS that exist (use the thumbgate-rules skill), to create a new rule (use the thumbgate-guard skill), or to check whether ThumbGate is installed and wired (use the thumbgate-doctor skill).
| name | adversarial-review |
| description | Run a skeptical Codex review from Claude Code that hunts for hidden regressions, unsafe assumptions, and release risk. Use before merges, deploys, migrations, or other risky changes. |
Default to the current uncommitted workspace unless the user explicitly provides base=<branch> or commit=<sha>.
Run one of:
node "${CLAUDE_PLUGIN_ROOT}/scripts/codex-bridge.js" adversarial-review --uncommitted
node "${CLAUDE_PLUGIN_ROOT}/scripts/codex-bridge.js" adversarial-review --base main
node "${CLAUDE_PLUGIN_ROOT}/scripts/codex-bridge.js" adversarial-review --commit <sha>
If the user asks for a specific threat model such as security, rollback safety, billing, or data loss, pass that in --prompt.
When reporting back: